Focus on Mapping the Brain

e are entering a new era in neuroscience in which techno- Contents logical development will allow us to obtain full anatomi- 483 From the connectome Wcal, high-resolution renderings of entire brain circuits and to brain function to map the activity of ever larger cellular populations as an animal C I Bargmann & E Marder performs specific behaviors. Assembling anatomical, molecular and 491 Making sense of brain functional maps has the potential to greatly advance our under- network data standing of how brains work. O Sporns In this Focus, experts outline the technologies needed to obtain 494 Why not these maps and discuss what will be needed beyond them to under- connectomics? J L Morgan & stand brain function. J W Lichtman Visualization of functional In a Historical Perspective, Cornelia Bargmann and Eve Marder connectivity in the human 501 Cellular-resolution discuss what has been learned from invertebrate circuits whose connectomics: cerebral cortex based on magnetic connectivity patterns are known and what will be needed beyond challenges of dense resonance imaging data. Brain neural circuit image by Joachim Böttger and anatomical maps to understand brain function in other organisms. reconstruction Daniel Margulies (Max Planck In a Commentary, Jeff Lichtman and Joshua Morgan express their M Helmstaedter Institute for Human Cognitive and views about why obtaining detailed, high-resolution structural Brain Sciences, Leipzig, Germany) maps should be an essential part of this endeavor. To understand 508 CLARITY for mapping with compositing help by Tobias S. the nervous system Hoffmann. Cover composition by the deluge of data these maps will engender once generated, Olaf K Chung & K Deisseroth Erin Dewalt. Sporns argues in another Commentary that data representation and 515 Mapping brain modeling will be critical. circuitry with a light Other experts discuss the newest technologies available for microscope obtaining brain maps. Moritz Helmstaedter presents the state of P Osten & T W Margrie the art and current challenges of electron microscopy–based cir- 524 Imaging human

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature cuit reconstruction. In three papers, the potential of using light connectomes at the macroscale to unveil the function and anatomy of brain circuits is presented. R C Craddock, S Jbabdi, Karl Deisseroth and Kwanghun Chung discuss their newly devel- C-G Yan, J T Vogelstein, npg oped method named CLARITY for rendering mammalian brains F X Castellanos, permeable to visible photons and molecules. Pavel Osten and Troy A Di Martino, C Kelly, Margrie review light-microscopy methods available for large-scale K Heberlein, S Colcombe & anatomical tracing and discuss ways to integrate molecular identity, M P Milham activity recording and anatomical information. In a Resource, Josh 540 Improved tools for the Brainbow toolbox Sanes and colleagues present improved tools for mapping the mouse Dawen Cai, K B Cohen, brain using the Brainbow technology. Finally, Michael Milham, Stan T Luo, J W Lichtman & Colcombe and their colleagues review methods for functional and J R Sanes anatomical analysis of human brains at the macroscale. We are pleased to acknowledge the financial support of Carl Zeiss Microscopy, Hamamatsu Corporation, LaVision BioTec, TissueVision, Inc. and Chroma Technology Corp. Nature Methods carries sole responsibility for all editorial content and peer review. Erika Pastrana

Editor, Nature Methods Daniel Evanko Copy Editor Odelia Ghodsizadeh Carol Evangelista, Ivy Robles Focus Editor Erika Pastrana Managing Production Editor Sponsorship David Bagshaw, Publisher Stephanie Diment Renee Lucas Yvette Smith, Reya Silao Senior Copy Editor Irene Kaganman Production Editors Brandy Cafarella, Marketing Nazly De La Rosa

nature methods | VOL.10 NO.6 | JUNE 2013 | 481 focus on mapping the brain historical perspective

From the connectome to brain function

Cornelia I Bargmann1 & Eve Marder2,3

In this Historical Perspective, we ask what behaviors5–7. In association with the recordings of information is needed beyond connectivity these individually recognizable, identified neurons, the diagrams to understand the function of nervous cells were filled with dye to visualize their structures systems. Informed by invertebrate circuits and projection patterns via light microscopy8–10. In whose connectivities are known, we highlight some cases, electron microscopy was used to observe the importance of neuronal dynamics and the anatomical synapses in these small circuits11–13. neuromodulation, and the existence of parallel But until the publication of the heroic electron micros­ circuits. The vertebrate retina has these features in copy reconstruction of the full nervous system of common with invertebrate circuits, suggesting that C. elegans14 in the mid-1980s, it was unimaginable they are general across animals. Comparisons across that the electron microscope could be used to deter- these systems suggest approaches to study the mine circuit connectivity rather than providing functional organization of large circuits based on ultrastructural detail to connectivity determined existing knowledge of small circuits. either with physiological or light microscopy–based An animal’s behavior arises from the coordinated activ- anatomical methods. ity of many interconnected neurons—“many” meaning Recent advances in electron microscopy and image 302 for Caenorhabditis elegans, 20,000 for a mollusc, analysis have made it possible to scale up this ultrastruc- several hundred thousand for an insect or billions for tural approach: to serially section and reconstruct pieces humans. Determining the connectivity of these neu- of both vertebrate and invertebrate nervous systems, rons, via combined anatomical and electrophysiologi- with the stated purpose of using detailed connectomes cal methods, has always been a part of neuroscience. As to reveal how these circuits work4,15–18. Such large- we were writing this, these ideas were being revisited scale projects will provide new anatomical data that will © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature from the perspective of massively parallel methods for offer invaluable insights into the functional organiza- dense reconstruction, or ‘connectomics’. One thread of tion of the structures studied. An unbiased approach this analysis involves the detailed, high-density map- to data acquisition always reveals surprises and new npg ping of point-to-point connections between neurons insights. Moreover, because of the scope and size of at synapses1–4. The specialized membrane structures these projects, such efforts will generate unprecedented and synaptic vesicles of synapses can be visualized amounts of data to be analyzed and understood. with an electron microscope, and consequently dense Here we ask what additional information is needed reconstructions of nervous-system connectomes rely beyond connectivity diagrams to understand circuit on electron microscopy of serial brain sections. In a function, informed by the invertebrate circuits whose complementary approach, detailed electrophysiologi- connectivity is known. For the prototypical case, the cal analysis shows how synapses and circuits function complete C. elegans nervous system, the anatomical at high resolution, and is increasingly being applied to connectome was largely established over 25 years large numbers of interconnected neurons. ago14. In a variety of other invertebrate preparations, The first approaches used to map complete circuits connectivity was established using combinations of came from studies of the smaller nervous systems electrophysiological recordings and neuronal tracing 30– of invertebrates. In the 1960s and 1970s, systematic 40 years ago, which enabled researchers to generate electrophysiological recordings from neurons in a wiring map that incorporates activity information. discrete ganglia enabled the identification of neu- Despite their different starting points from anatomy and ronal components of circuits that generate specific electrophysiology, these two approaches have uncovered

1Howard Hughes Medical Institute, The Rockefeller University, New York, New York, USA. 2Volen Center, Brandeis University, Waltham, Massachusetts, USA. 3Department of Biology, Brandeis University, Waltham, Massachusetts, USA. Correspondence should be addressed to C.I.B. ([email protected]). Received 27 February; accepted 5 April; published online 30 may 2013; doi:10.1038/nmeth.2451

nature methods | VOL.10 NO.6 | JUNE 2013 | 483 historical perspective FOCUS ON MAPPING THE BRAIN

Figure 1 | Connectivity of two a well-studied invertebrate circuits. (a) Connectivity diagram of the crab STG based on electrophysiological AB PD LPG recordings. Red and blue background Electrical synapse shading indicates neurons that are LP IC LG MG GM primarily part of the pyloric and Chemical inhibitory synapses gastric circuits, respectively. Purple PY VD Int1 DG AM shading indicates that some neurons switch between firing in pyloric and gastric time, and that there is no fixed boundary between the pyloric and gastric circuits. Yellow ASJL b PHAR AIMR highlights two neurons that are both ASJR ASIR AINL electrically coupled and reciprocally ALMR PVM PHAL ASIL AIML HSNL PHCR inhibitory. Green highlights one AWALAINRAWAR ASGL AVFL ADLR IL2R ASEL AWBL ADLL of many examples of neurons that IL2L AWBR PVQL ALML PLNR FLPR ADFL VC05AVM PVDR IL2VRIL2VL CEPVRCEPDR ASGR SDQL LUAR ASER ASHR PDEL PLMR are coupled both monosynaptically IL2DL AWCR PLNL ASHL PVPR IL2DR RIH AWCL ADFR ADEL BDULAVHLAVHR BDUR PHBR BAGL PVQR FLPLPVNL PVDLVC04 URBRCEPVLCEPDL ADER PQR PHBL URXL ASKL PDER PLMLPHCL and polysynaptically. (b) The ASKR AVG ALNL SDQR RIFR LUAL AFDLAFDR OLLL RIR RIFL HSNR ADAL connectome of C. elegans, URXR AUALBAGR PVWR AUARURYVRURBLURYVL ALNR PVT AIYL RMGLRMGRAIAL RIS ALAPVR PVNR URAVR URADR OLLR AIZL AIAR ADAR AVFRAVJRPVPL URAVLIL1VR AIZR AQR showing all 302 neurons and their IL1R RICL PVWL PVCR URADL AIYR DVA IL1L OLQVL URYDR SAAVL IL1VL RICR chemical synapses but not their OLQVR URYDL AIBL AVDR RIVR RMFL VA12 RIAR SMBVR DVC AVDL PDA IL1DL OLQDRRIAL RIVL AIBR SAAVR AVJL VD11 RIGR RIGL VB01 PVCL DVB VC02 gap junctions. Each neuron has a OLQDL SAADR IL1DR SMBVLSMBDR AVKR AVKL RMFR AVL RMEL RMHRRMHLRIBL AVERSAADLAVEL VB11 VB08 three-letter name, often followed RIBR AVBL AVARAVAL VC03 RIPL SMBDL RIPR SIADR RIMR DB01 RMER SIBDR RIML AVBR VB10 by a spatial designator. This top- RMEV SIADL SIBVL VB09 RMED RMDLSIAVR RID SIAVLSIBDL VB07 VB02 VA07 RMDDR SIBVR DB07 VA02 VB06 SMDVL DB02VC01 to-bottom arrangement (signal RMDDL SABD DA09AS11 DB03 RMDR PDB VA08VA09 DB04 VB04 AS01VA01 VD10AS02 VA04 RMDVRSMDVR VA11VD13 VA03 flow view) is arranged to reflect SMDDL AS09 AS06 SABVR DA01 DA03DA04 VA06 RMDVL SMDDR DB06DB05 AS03 DA05AS05 DA08 DA02 AS04 dominant information flow, SABVL VA05 VD01 VD09 which goes from sensory neurons AS07 DA06 AS10 VD12 VA10AS08DA07 (red) to interneurons (blue) to DD06 DD05 VD08 VD07 VD02DD01 VD05 motor neurons (green). Reprinted DD04DD02 from ref. 50. VD03 VD04VD06 DD03

similar principles and similar puzzles as to how circuit function STG neurons switch their activity between the two rhythms19, arises from the component neurons and their interactions. and the separation of the STG’s connectivity into two discrete circuits, although convenient for those who study the network, What do functional and anatomical maps reveal? does not really capture the highly interconnected reality of the We begin with the connectivity diagram of the stomatogas- ganglion’s architecture.

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature tric ganglion (STG) of the crab, Cancer borealis (Fig. 1a) and a Like all nervous systems, the circuit has many chemical synapses, graph of the connectome of C. elegans (Fig. 1b). In each case, the in which a presynaptic neuron releases a chemical neurotransmit- number of neurons is small, ~27 neurons or 302 neurons, respec- ter to activate receptors on the postsynaptic neuron. Chemical

npg tively, but the number of synaptic connections is much larger; synapses can be inhibitory or excitatory depending on the nature the neurons are extensively interconnected. The basic function of of the receptor and associated ion channels; the chemical synapses each circuit is known: to generate rhythmic stomach movements among STG neurons are inhibitory. Additional connections are for the crab STG and to control locomotion behavior in response created by the widespread electrical synapses, mediated by direct to sensory inputs for C. elegans. The intellectual strength of the cytoplasmic communication through gap junctions, through STG system is the ability to relate neuronal connectivity to which current flows depending on the voltages of the coupled neuronal activity patterns; the complementary strength of neurons. In the STG circuit, there are many instances of neurons C. elegans is the ability to relate neuronal connectivity to whole- that are connected by electrical synapses as well as by chemical animal behavior. inhibitory synapses (Fig. 1a). There are also many instances of The STG contains motor neurons and interneurons that gen- neurons connected by reciprocal inhibition. These wiring motifs erate two rhythmic motor patterns19. The pyloric rhythm is an contribute to circuit properties that are not easily predictable. oscillating, triphasic motor pattern that is continuously active and In addition, there are many ‘parallel pathways’ in which two neu- depends on a set of electrically coupled pacemaker neurons. The rons are connected via two or more synaptic routes, one direct gastric mill rhythm is episodically active and depends on descend- route and additional indirect routes (Fig. 1a). The complexity of ing modulatory inputs activated by sensory neurons for its gen- this connection map poses the essential question: are all synapses eration19,20. Although these rhythms are easily studied separately, important, or are some only important under certain conditions a close look at the STG connectivity diagram reveals that the neu- (as appears to be the case)21? How do we understand the impor- rons that conventionally are thought to be part of the pyloric tance of synaptic connectivity patterns that seem to oppose each circuit (neurons AB, PD, LP, PY, VD and IC) are highly intercon- other, such as the common motif of electrical coupling between nected with those conventionally thought part of the gastric mill neurons that also inhibit each other? circuit (neurons DG, GM, LPG, MG, LG and Int1) (AM is part The C. elegans wiring diagram was assembled in the near- of a third circuit that we will not discuss here). Indeed, many complete absence of prior functional information. It allowed an

484 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN historical perspective

Anterior touch Posterior touch the number of synapses and their importance in behavior is seen in the avoidance of head touch, where just two of 58 synapses ALM AVM PLM (again representing gap junctions) are the key link between the sensory neurons (ALM and AVM) and the essential interneuron 9 classes 9 classes AVD AVB PVC 6 classes (AVD). This general mismatch between the number of syn- of neurons of neurons of neurons apses and apparent functional importance has applied wherever C. elegans circuits have been defined. As a result, early guesses VA VB about how information might flow through the wiring diagram Gap junction DA DB Chemical synapse were largely incorrect. Sensory neuron Clearly, the wiring diagram could generate hypotheses to test, Interneuron Motor neuron Reversal Acceleration but solving a circuit by anatomical inspection alone was not suc- cessful. We believe that anatomical inspection fails because each Figure 2 | C. elegans neurons essential for avoidance of light touch. wiring diagram encodes many possible circuit outcomes. Inferred connections necessary for anterior and posterior touch avoidance are in purple and orange, respectively; other synapses are in black. The ‘essential’ synapses here shown in orange and purple comprise less than Parallel and antagonistic pathways complicate circuits 10% of the output synapses of the mechanosensory neurons. Image based Both of the wiring diagrams shown in Figure 1 are richly con- on ablation data from ref. 22. nected. In the STG, a large fraction of the synapses are electrical synapses. In some cases, the electrical synapses connect multiple immediate classification of neurons into large classes: sensory copies of the same neuron, such as the two PD neurons in the neurons (with distinctive sensory dendrites and cilia), motor STG. Notably, many electrical synapses connect neurons with dif- neurons (with neuromuscular junctions) and interneurons ferent functions. Almost invariably, the combination of electrical (a term that is used in C. elegans to describe any neuron that is and chemical synapses create ‘parallel pathways’, that is to say, not evidently sensory or motor, encompassing projection neurons multiple pathways by which neuron 1 can influence neuron 2 and local neurons)14. In each group, neurons were subdivided into (Fig. 1a). For example, in the STG, the PD neuron inhibits the IC unique types with similar morphologies and connections, collaps- neuron through chemical synapses but also can influence the IC ing the wiring diagram from 302 neurons to 119 neuronal types. neuron via the electrical synapse from LP to IC. Parallel pathways The flow of information through chemical synapses is predomi- such as those in the STG can be viewed as degenerate, as they nantly from sensory to interneuron to motor neuron, with many create multiple mechanisms by which the network output can be parallel pathways linking neurons both directly and indirectly switched between states23 (Fig. 3). A simulation study23 shows a (as in the STG), as well as gap junctions that may form electrical simplified five-cell network of oscillating neurons coupled with synapses (~10% of all synapses). Most neurons are separated from electrical synapses and chemical inhibitory synapses. The f1 and each other by no more than two or three synaptic connections. f2 neurons are connected reciprocally by chemical inhibitory The C. elegans map was immediately used to define neurons synapses, as are the s1 and s2 neurons. This type of wiring con- required for the touch-avoidance response, which is still the most figuration, called a half-center oscillator, often but not universally 22 24 © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature completely characterized of the animal’s behaviors . Light touch causes the neurons to be rhythmically active in alternation . In to the head elicits a reversal, and light touch to the tail elicits this example, two different oscillating rhythms are generated, one a forward acceleration. The neurons required for the touch- fast and one slow. The hub neuron at the center of the network

npg avoidance response were identified by killing cells with a laser can be switched between firing in time with the fast f1 and f2 microbeam and assessing the behavioral repertoire of the worms. neurons to firing in time with the slow s1 and s2 neurons by three Guided by the wiring diagram, this analysis revealed essential entirely different circuit mechanisms: changing the strength of the mechanosensory neurons in the head and tail, key interneurons electrical synapses, changing the strength of the synapses between required to propagate information, and motor neurons required f1 and s1 onto the hub neuron, and changing the strength of the for forward and backward movement (Fig. 2). The success of reciprocal inhibitory synapses linking f1 to f2 and s1 to s2 in the this approach inspired similar analyses of chemosensory behav- half-center oscillators. iors, foraging, egg-laying, feeding and more. At this point, over An example from the C. elegans connectome illustrates another 60% of C. elegans neuron types have defined functions in one or twist of circuit logic: divergent circuits that start at a common more behaviors. point but result in different outcomes. In this example, gap junc- This notable success, however, hides a surprising failure. For tions and chemical synapses from ADL sensory neurons generate C. elegans, although we know what most of the neurons do, we opposite behavioral responses to a C. elegans pheromone (Fig. 4a). do not know what most of the connections do, we do not know The chemical synapses drive avoidance of the pheromone, whereas which chemical connections are excitatory or inhibitory, and we the gap junctions stimulate a pheromone-regulated aggregation cannot easily predict which connections will be important from behavior25. Differing use of the chemical synapse subcircuit ver- the wiring diagram. The problem is illustrated most simply by sus the gap junction subcircuit allows ADL to switch between the classical touch-avoidance circuit22 (Fig. 2). The PLM sensory these two opposing behaviors in different contexts. ADL illus- neurons in the tail are solely responsible for tail touch avoid- trates the point that is not possible to ‘read’ a connectome if it is ance. PLM forms 31 synapses with 11 classes of neurons, but only intrinsically ambiguous, encoding two different behaviors. one of those targets is essential for the behavior—an interneuron Parallel and divergent systems of synapses are widespread fea- called PVC that is connected to PLM by just two gap junctions tures of invertebrate and vertebrate networks alike, and can be and two chemical synapses. An even greater mismatch between composed of sets of chemical synapses as well as sets of chemical

nature methods | VOL.10 NO.6 | JUNE 2013 | 485 historical perspective FOCUS ON MAPPING THE BRAIN

Figure 3 | Similar changes in a b c d circuit dynamics can arise from f1 f1 f1 g f1 three entirely different circuit synB gsynA mechanisms. (a) Circuit diagram gel (top) shows the ‘control’ condition f2 hn s2 f2 hn s2 f2 hn s2 f2 hn s2 gel in which f1 and f2 are firing g synA g in a fast rhythm (indicated by s1 s1 s1 s1 synB red shading) and the remaining f1 neurons are firing in a slow f2 rhythm (shaded in blue). hn is the hn hub neuron. In these diagrams, s2 electrical synapses are shown as s1

resistor symbols and chemical 100 mV 1 s inhibitory synapses with filled circuits. Traces (bottom) show the voltage waveforms of the five neurons. (b–d) Responses when the strength of the chemical synapses to the hub neuron (gsynA) was decreased (b), when the strength of the electrical synapses (gel) was decreased (c) and when the strength of the chemical synapses between f1 and f2 and between s1 and s2 (gsynB) was decreased (d). Image modified from ref. 23.

and electrical synapses. To understand information flow, there Many years of work on the effects of neuromodulators on the will be no substitute for recording activity. The methods for moni- STG have revealed that the functional connections that give rise toring neuronal activity have improved dramatically in recent to a specific circuit output are specified, or in fact ‘configured’, by years, with development of new multi-electrode recording tech- the neuromodulatory environment29. Every synapse and every niques and a suite of genetically encoded indicators that can be neuron in the STG is subject to modulation; the connectivity dia- used to measure calcium, voltage and synaptic release at cellular gram by itself only establishes potential circuit configurations, and subcellular levels. However, improved methods are needed whose availability and properties depend critically on which of to detect electrical synapses, which can also be difficult to see many neuromodulators are present at a given moment29. Under in electron micrographs. The regulation of electrical synapses some modulatory conditions, anatomically ‘present’ synaptic by voltage, neuromodulation, phosphorylation and small mole­ connections may be functionally silent, only to be strengthened cules is understudied26,27. A chemical method for measuring gap under other modulatory conditions. Likewise, modulators can junctions, local activation of molecular fluorescent probes, is a qualitatively alter the neurons’ intrinsic properties, transforming promising new direction that should spawn innovation28. neurons from tonic spiking to those generating plateau poten- tials or bursts29. These effects of neuromodulators can activate Neuromodulation reconfigures circuit properties or silence an entire circuit, change its frequency and/or the phase Superimposed on the fast chemical synapses and electrical syn- relationships of the motor patterns generated. apses in the wiring diagram are the neuromodulators—biogenic C. elegans has over 100 different neuropeptides as well as bio-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature amines (serotonin, dopamine, norepinephrine and histamine) genic amine neuromodulators. The integration of neuromodula- and neuropeptides (dozens to hundreds, depending on species)29. tion into its fast circuits appears to selectively enhance the use of These molecules are often released together with a fast chemical particular connections at the expense of others. For example, a

npg transmitter near a synapse, but they can diffuse over a greater dis- ‘hub-and-spoke’ circuit drives aggregation of C. elegans by cou- tance. Modulators also can be released from neuroendocrine cells pling multiple sensory inputs through gap junctions with a com- that do not make defined synaptic contacts or can be delivered as mon target neuron, RMG (Fig. 4b). Neuromodulation of RMG hormones through the circulation. As a result, the targets of neuro- by the neuropeptide receptor NPR-1 effectively silences this gap- modulation are invisible to the electron microscope. Signaling pri- junction circuit, while sparing other functions of the input sensory marily through G protein–regulated biochemical processes rather neurons that are mediated through chemical synapses30. than through ionotropic receptors, neuromodulators change neu- Neuromodulators are prominent in all nervous systems, and ronal functions over seconds to minutes, or even hours. act as key mediators of motivational and emotional states such as

Figure 4 | Two views of a multifunctional a C9 pheromone b C. elegans circuit. (a) Ambiguous circuitry of the ADL sensory neurons, which drive Pheromone Oxygen ADL attraction ASK URX avoidance avoidance of the ascaroside pheromone AVA NPR-1

C9 through chemical synapses onto multiple AIA Nociceptive Pheromone interneurons (right) but can also promote RMG AVD avoidance ASH RMG ADL avoidance aggregation (attraction toward pheromones) AIB through gap junctions with RMG (left). Image AWB IL2 modified from ref. 25. (b) Neuromodulation Pheromone separates overlapping circuits. Multiple sensory Aggregation avoidance Aggregation neurons form gap junctions with the RMG hub Gap junction Chemical synapse Sensory neuron Interneuron or motor neuron neurons and promote aggregation through this circuit, but each sensory neuron also has chemical synapses that can drive RMG-independent behaviors. The neuropeptide receptor NPR-1 inhibits RMG to suppress aggregation. Image modified from ref. 30.

486 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN historical perspective

sleep, arousal, stress, mood and pain. To understand their release 25 Synaptic 1 20 I and their effects on circuits, new methods are needed to monitor h neuromodulation in vivo. Electrophysiology remains the best tool 15 2 10

for characterizing the functional effects of neuromodulators but Period (s) 5 is low-throughput. Biochemical methods can be used to reveal 0 the presence of neuromodulators in tissue or in bulk extracellular 0 20 40 60 80 100 fluid but are less effective for detecting them near a particular Conductance (nS) Synaptic conductance synapse or release site. A new genetic method can be used to read 50 nS 60 nS 80 nS out the neuromodulatory state directly by monitoring receptor activation but with a timeframe of hours, whereas endogenous 2 22 22 31 modulation can change within minutes or seconds . Progress 1 11 11 is needed in all of these domains and beyond: there is a need to move from individual neurons and modulators to physiologically 10 mV lh conductance 2 s relevant modulatory states, which are likely to include multiple 80 nS 50 nS 30 nS neuromodulators acting at many sites. 2 22 22

Neuronal dynamics shape the activity of circuits 1 11 11 The existence of parallel circuits and neuromodulation means 10 mV that connectivity alone does not provide adequate information 2 s to predict the physiological output of circuits. Even without these Figure 5 | Changing either intrinsic neuronal properties or synaptic factors, the behavior of neurons over time is unpredictable from properties can alter network function. The dynamic clamp, a computer- anatomy because neuronal behavior is sensitive to intrinsic chan- neuron interface (top) was used to vary either the strength of synaptic nels and electrical properties that vary within and between cell connections between two neurons (synaptic) or the amount of an types. Channels, synapses and biochemical processes interact intrinsic hyperpolarization-activated inward current (Ih), in one neuron to generate explicitly time-delimited features, or dynamics, in as graphed (right). Traces (bottom) show the action potentials generated neurons and circuits. by the alternating, oscillating neuron pair as those properties were varied. Image modified from ref. 24. The importance of neuronal dynamics in circuit function can be seen most simply in a two-cell circuit (Fig. 5). Two isolated neurons from the STG that are not normally synaptically coupled on behavioral states35. Finally, synaptic plasticity can occur on were connected using the dynamic clamp, a computer-neuronal rapid timescales to strengthen and weaken synapses based on use, interface that allows a user to manipulate biological neurons adding complexity to circuit-level dynamics36. with conductances that imitate ion channels and synaptic con- Analyzing neuronal dynamics often requires the circuit to be nections32. The neurons are connected reciprocally by dynamic simultaneously monitored and manipulated, as shown in the clamp-created inhibitory synapses so that the neurons rhythmi- example of the dynamic clamp. Emerging techniques of optoge- 24 © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature cally alternate their activity . The dynamic clamp allows the netics and pharmacogenetics can be combined with recording investigator to change the strength of the synapses as well as the as well, but a limitation of all of these methods is that they act at amount of one of the membrane currents, hyperpolarization- the level of neurons or groups of neurons. To understand func-

npg activated inward current (Ih)—either of which dramatically tional connectivity, it will be useful to develop methods to silence alters the period of the circuit oscillation (Fig. 5). Thus, a given or activate specific channels and specific synaptic connections wiring diagram can produce widely different dynamics with between two specified neurons, without affecting all other func- different sets of circuit parameters, and conversely, different tions of the same cells. circuit mechanisms can give rise to similar oscillation dynamics. Without knowing the strength and time course of the synaptic Vertebrate retina also has complex circuit properties connections as well as the numbers and kinds of membrane What lessons will emerge as connectomes are scaled up from currents in each of the neurons, it would not be possible to small-scale to large-scale circuits? Many features will be com- simply go from the wiring diagram to the dynamics of even mon to small and large circuits. Vertebrate circuits, like inverte- two neurons. Synaptic connectivity alone does not sufficiently brate circuits, have multiple cell types with nonuniform intrinsic constrain a system. properties, extensive and massively parallel synaptic connectivity, Understanding neuron-specific and circuit-specific dynamics and neuromodulation. The balance of these components varies will be essential to understanding mammalian circuits as well as between animals and brain regions (the STG has more electri- invertebrate circuits. In some cases, unique dynamic properties cal synapses than most vertebrate brain regions; C. elegans uses are characteristics of particular cell types—for example, different mostly graded potentials instead of all-or-none action potentials), classes of inhibitory cortical interneurons are distinguished as but in reality, the diversity of circuits in the vertebrate brain is much by their dynamics as by their connectivity33. In other cases, at least as great as the difference between any one vertebrate neuronal dynamics are variable among similar cells or even within region and any invertebrate circuit. The essential distinction we one cell type. For example, pyramidal neurons in specific areas see in vertebrate brains is not a particular microcircuit property of the cortex exhibit persistent activity associated with working but their repeating structure (for example, the many cortical memory34, and neurons in brainstem modulatory systems switch columns) and their enormous scale compared to the worm brain their properties between tonic and phasic firing modes depending and the STG.

nature methods | VOL.10 NO.6 | JUNE 2013 | 487 historical perspective FOCUS ON MAPPING THE BRAIN

Examination of the vertebrate retina has begun to reveal the amacrine cells around dawn, under the control of acute light relationships between performance of a large circuit and prop- stimuli and circadian rhythm42, and acts on cells throughout the erties of a small circuit. The special power of studying the iso- retina to switch them from properties appropriate to night vision lated retina is the ability to experimentally control visual input to day vision. The photoreceptors themselves are modulated, and while simultaneously recording output—the spikes from retinal their coupling through gap junctions decreases to increase their ganglion cells that project to the brain. The retina represents an resolution but reduce their sensitivity. Downstream of the rod intermediate degree of complexity with features of both a small photoreceptors, which dominate night vision, dopamine closes circuit and a large circuit, and has been subject to the most com- the gap junctions between rod bipolar cells and AII amacrine plete anatomical and electrophysiological characterizations of cells, effectively diminishing rod input to the retinal ganglion cell any vertebrate brain region. Current connectomics studies of output of the retina. the retina, for example, include dense reconstruction of serial- What differences are there between small and large circuits? The section electron micrographs accompanied by analysis of the sheer size of the retina shows a sharp transition compared to the neurotransmitter phenotype and activity patterns of the recon- size of the STG and the worm brain, and the level of analysis moves structed neurons3,17. The combination of structure and function, from single cells to cell classes. Understanding a single pixel is and a rich history of elegant experiments, make this the ideal not sufficient to understand the retina, and here the properties of system for understanding neural computations in detail. simple and complex circuits diverge. For example, long-range com- The retina contains millions of neurons that fall into five major munication allows groups of retinal cells to perform computations neuronal classes (photoreceptors, horizontal cells, bipolar cells, that a single cell cannot. Wide-field cells such as starburst amacrine amacrine cells and retinal ganglion cells), which are subdivided cells can make judgments about motion that no single-pixel neuron into about 60 discrete cell types37. Each of the 60 cell types is could make but can then feed that information into narrow-field arrayed in a near-crystalline two-dimensional array, so that any single-pixel neurons to bias their properties. The scaling from fine pixel viewed by the retina is covered by at least one neuron of each resolution to broad resolution and back again emerges from the cell type. Ultimately, information leaves the retina through the diversity of spatial scales across the structure of the retina. 20 classes of retinal ganglion cells, each of which is considered to be a parallel but partially overlapping processing stream. Circuits interact to generate behavior The first views of the retinal connectome show all of the prop- The entire nervous system is connected, but reductionist neu- erties that we highlight in small circuits: cellular complexity, roscientists invariably focus on pieces of nervous systems. The extensive interconnectivity, parallel circuits with chemical and value of these simplified systems should not let us forget that electrical synapses, and neuromodulation. The heterogeneity of behavior emerges from the nervous system as a whole. At the the 60 retinal cell types is substantial, echoing the heterogeneity moment, obtaining the connectomes of even small parts of of individual neurons in the STG or C. elegans. Anatomically, the vertebrate nervous system is a heroic task. However, estab- some dendritic arbors cover only a tiny area of the visual field, but lishing the detailed pattern of connectivity for a small part of the others arborize much more broadly. Their intrinsic physiological nervous system may not be sufficient to understand how that properties are also extremely diverse, with some neurons that piece functions in its full context. By parceling out small regions,

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature spike (such as retinal ganglion cells), and many neurons that do one invariably loses information about the long-range connec- not spike (such as photoreceptors and bipolar cells)37. There are tions to and from that area. even amacrine neurons that perform independent computations The extent to which long-range connectivity clouds our under- 38

npg in different parts of their complex arbors . standing of connectomes will vary. For example, the vertebrate Synaptic connections in the retina are extensive and diverse, retina is anatomically isolated, functionally coherent and lacks and electron microscopy reconstructions have revealed many recurrent feedback synapses from other brain areas that are classes of synaptic connections that had not been observed in prominent in most other parts of the central nervous system. We physiological studies3,17. There is a great variety of excitatory might imagine the retina as a two-dimensional circuit, whereas and inhibitory chemical synapses, and there are many electri- most vertebrate circuits are three-dimensional; new principles cal synapses, that all vary in their strength and their modifica- will certainly arise from connectomes that include recurrent tion by experience. Both anatomical and physiological studies inputs. In the amygdala, for example, the intermixing of multi- demonstrate that the retina, like small circuits, consists of many ple cell types with different long-range inputs and outputs would partly parallel circuits with overlapping elements. In particu- preclude a meaningful understanding based on local anatomy lar, the retina operates over many orders of magnitude of light alone43. Choosing well among brain regions, and combining con- intensity, and the properties of its circuits change with its visual nectomes with molecular and functional information about the inputs. Within a few seconds in a new visual environment, retinal same cells, as is being done in the vertebrate retina3,17, will lead ganglion cells shift their properties to encode relevant features to the most informative results. of light intensity, contrast and motion, drawing on different features of the network39. Subsets of retinal ganglion cells change How can we ‘solve’ the brain? their weighting of center and surround inputs in a switch-like As we look to ways that other neural systems may be charac- fashion as light levels change40. A brief period of visual stimula- terized with similar power to the three described here, we can tion can even reverse the apparent direction-selectivity of retinal draw certain lessons. One is that precise circuit mapping and spe- ganglion cells41. cific neuron identification have had great importance for unify- Finally, neuromodulation has a role in retinal processing that ing structural and functional data from different laboratories. reshapes visual circuits. Dopamine is released from a subset of Extending this idea, other systems may not have individually

488 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN historical perspective

named cells, but all nervous systems have cell types distinguished Acknowledgments by anatomy, connectivity and molecular profile that can serve as We thank M. Meister for sharing his knowledge of the retina. C.I.B. is funded by the Howard Hughes Medical Institute. Research in the Marder laboratory relevant the basis of a common vocabulary. Improvements in molecular to this piece is funded by the US National Institutes of Health (NS17813, NS staining methods will only increase the power of a connectome 81013 and MH46742). anchored in cellular identity. We can also see that connectomic endeavors will need to be COMPETING FINANCIAL INTERESTS The authors declare no competing financial interests. supplemented by experiments that monitor, manipulate and model circuit activity. Monitoring and manipulation of circuit Reprints and permissions information is available online at http://www.nature. function have been considered above. To complement and inform com/reprints/index.html. these experimental approaches, the third step is to develop mod- 1. Bock, D.D. et al. Network anatomy and in vivo physiology of visual els that describe how a system’s output results from the interac- cortical neurons. Nature 471, 177–182 (2011). tions of its components. There is a tension between the desire to 2. Briggman, K.L. & Bock, D.D. Volume electron microscopy for neuronal study abstract models that are amenable to precise mathematical circuit reconstruction. Curr. Opin. Neurobiol. 22, 154–161 (2012). 3. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the analysis and the desire to study models with sufficient biologi- direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). cal realism to represent the system’s underlying structures and 4. Kleinfeld, D. et al. Large-scale automated histology in the pursuit of functions. In small circuits, it is now possible to construct connectomes. J. Neurosci. 31, 16125–16138 (2011). models and families of models that can be quite instructive44. 5. Burrows, M. Monosynaptic connexions between wing stretch receptors and flight motoneurones of the locust. J. Exp. Biol. 62, 189–219 (1975). In C. elegans, a few testable models emerged directly from analy- 6. Fentress, J.C. Simpler Networks and Behavior (Sinauer Associates, 1976). ses of anatomy. One was the concept of a motif, a set of connection 7. Getting, P.A., Lennard, P.R. & Hume, R.I. Central pattern generator patterns between three or four neurons that are over-represented mediating swimming in Tritonia. I. Identification and synaptic interactions. J. Neurophysiol. 44, 151–164 (1980). in the wiring diagram compared to the statistical expectation 8. Stretton, A.O. & Kravitz, E.A. Neuronal geometry: determination with a 45 based on individual connections . Perhaps these motifs per- technique of intracellular dye injection. Science 162, 132–134 (1968). form a canonical computation, or a few canonical computations, 9. Remler, M., Selverston, A. & Kennedy, D. Lateral giant fibers of cray fish: so that solving a few of them effectively solves a larger piece location of somata by dye injection. Science 162, 281–283 (1968). 10. Goodman, C.S. Isogenic grasshoppers: genetic variability in the morphology of the diagram. of identified neurons. J. Comp. Neurol. 182, 681–705 (1978). But how should we approach building models of large networks 11. Maynard, E.A. Electron microscopy of the stomatogastric ganglion in the without generating models that are as difficult to understand as lobster, Homarus americanus. Tissue Cell 3, 137–160 (1971). the biological systems that motivated and inspired them? The 12. King, D.G. Organization of crustacean neuropil. I. Patterns of synaptic connections in lobster stomatogastric ganglion. J. Neurocytol. 5, 207–237 beauty of the connectome is its precision and specificity, but it (1976). is hard to imagine useful network models that implement all of 13. King, D.G. Organization of crustacean neuropil. II. Distribution of synaptic the details of cell-to-cell connectivity obtained with the electron contacts on identified motor neurons in lobster stomatogastric ganglion. J. Neurocytol. 5, 239–266 (1976). microscope, when building such models would require enormous 14. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S. The structure of numbers of assumptions about other circuit parameters, and these the nervous system of the nematode Caenorhabditis elegans. Phil. Trans. parameters are likely to change in different modulatory states. So R. Soc. Lond. B 314, 1–340 (1986). 15. Seung, H.S. Reading the book of memory: sparse sampling versus dense © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature we face a conundrum: the new anatomical data will be instructive, mapping of connectomes. Neuron 62, 17–29 (2009). but it is not yet obvious what kinds of models will best reveal the 16. Meinertzhagen, I.A. & Lee, C.H. The genetic analysis of functional implications of these data for how circuits actually work. connectomics in Drosophila. Adv. Genet. 80, 99–151 (2012). 17. Anderson, J.R. et al. Exploring the retinal connectome. Mol. Vis. 17, npg We are in the midst of a fascinating international debate about 355–379 (2011). whether it is the right time to embark on a ‘big science’ project to 18. Lu, J., Tapia, J.C., White, O.L. & Lichtman, J.W. The interscutularis muscle monitor and model large brain regions. There are those who argue connectome. PLoS Biol. 7, e32 (2009). that we are now at the point at which investments in large-scale 19. Marder, E. & Bucher, D. Understanding circuit dynamics using the projects will considerably advance the field in ways not possible stomatogastric nervous system of lobsters and crabs. Annu. Rev. 46–48 Physiol. 69, 291–316 (2007). by a distributed small-lab approach . Big science works best 20. Nusbaum, M.P. & Beenhakker, M.P. A small-systems approach to motor when the goals of a project are well-defined and when the out- pattern generation. Nature 417, 343–350 (2002). comes can be easily recognized. Both were true about the human 21. Thirumalai, V., Prinz, A.A., Johnson, C.D. & Marder, E. Red pigment genome project, but neither is true, yet, about large-scale attempts concentrating hormone strongly enhances the strength of the feedback to the pyloric rhythm oscillator but has little effect on pyloric rhythm to understand the brain. Moreover, this is well-recognized, and all period. J. Neurophysiol. 95, 1762–1770 (2006). of the proponents of large-scale initiatives are acutely aware of the 22. Chalfie, M. et al. The neural circuit for touch sensitivity in Caenorhabditis necessity to develop new technology48 and of the extraordinary elegans. J. Neurosci. 5, 956–964 (1985). 49 23. Gutierrez, G.J., O’Leary, T. & Marder, E. Multiple mechanisms switch an complexity of biological systems . That said, the largest challenge electrically coupled, synaptically inhibited neuron between competing we face in future attempts to understand the dynamics of large rhythmic oscillators. Neuron 77, 845–858 (2013). circuits is not in collecting the data: what is most needed are new 24. Sharp, A.A., Skinner, F.K. & Marder, E. Mechanisms of oscillation in methods that allow our human brains to understand what we find. dynamic clamp constructed two-cell half-center circuits. J. Neurophysiol. 76, 867–883 (1996). Humans are notoriously bad at understanding multiple nonlinear 25. Jang, H. et al. Neuromodulatory state and sex specify alternative processes, although we excel at pattern recognition. Somehow, we behaviors through antagonistic synaptic pathways in C. elegans. have to turn the enormous data sets that are already starting to be Neuron 75, 585–592 (2012). 26. Pereda, A.E. et al. Gap junction-mediated electrical transmission: regulatory generated into a form we can analyze and think about. Otherwise, mechanisms and plasticity. Biochim. Biophys. Acta 1828, 134–146 (2013). we will be doomed to creating a machine that will understand the 27. Neyton, J. & Trautmann, A. Physiological modulation of gap junction human brain better than we can! permeability. J. Exp. Biol. 124, 93–114 (1986).

nature methods | VOL.10 NO.6 | JUNE 2013 | 489 historical perspective FOCUS ON MAPPING THE BRAIN

28. Guo, Y.M. et al. Imaging dynamic cell-cell junctional coupling in vivo 39. Hosoya, T., Baccus, S.A. & Meister, M. Dynamic predictive coding by the using Trojan-LAMP. Nat. Methods 5, 835–841 (2008). retina. Nature 436, 71–77 (2005). 29. Marder, E. Neuromodulation of neuronal circuits: back to the future. 40. Farrow, K. et al. Ambient illumination toggles a neuronal circuit Neuron 76, 1–11 (2012). switch in the retina and visual perception at cone threshold. Neuron 78, 30. Macosko, E.Z. et al. A hub-and-spoke circuit drives pheromone attraction 325–338 (2013). and social behaviour in C. elegans. Nature 458, 1171–1175 (2009). 41. Rivlin-Etzion, M., Wei, W. & Feller, M.B. Visual stimulation reverses 31. Inagaki, H.K. et al. Visualizing neuromodulation in vivo: TANGO-mapping the directional preference of direction-selective retinal ganglion cells. of dopamine signaling reveals appetite control of sugar sensing. Cell 148, Neuron 76, 518–525 (2012). 583–595 (2012). 42. Witkovsky, P. Dopamine and retinal function. Doc. Ophthalmol. 108, 17–40 32. Sharp, A.A., O’Neil, M.B., Abbott, L.F. & Marder, E. Dynamic clamp: computer- (2004). generated conductances in real neurons. J. Neurophysiol. 69, 992–995 (1993). 43. Ciocchi, S. et al. Encoding of conditioned fear in central amygdala 33. Kawaguchi, Y. & Kondo, S. Parvalbumin, somatostatin and cholecystokinin inhibitory circuits. Nature 468, 277–282 (2010). as chemical markers for specific GABAergic interneuron types in the rat 44. Marder, E. & Taylor, A.L. Multiple models to capture the variability in frontal cortex. J. Neurocytol. 31, 277–287 (2002). biological neurons and networks. Nat. Neurosci. 14, 133–138 (2011). 34. Egorov, A.V., Hamam, B.N., Fransen, E., Hasselmo, M.E. & Alonso, A.A. 45. Milo, R. et al. Network motifs: simple building blocks of complex Graded persistent activity in entorhinal cortex neurons. Nature 420, networks. Science 298, 824–827 (2002). 173–178 (2002). 46. Markram, H. The blue brain project. Nat. Rev. Neurosci. 7, 153–160 35. Aston-Jones, G. & Cohen, J.D. An integrative theory of locus coeruleus- (2006). norepinephrine function: adaptive gain and optimal performance. Annu. 47. Koch, C. & Reid, R.C. Neuroscience: Observatories of the mind. Rev. Neurosci. 28, 403–450 (2005). Nature 483, 397–398 (2012). 36. Zucker, R.S. & Regehr, W.G. Short-term synaptic plasticity. Annu. Rev. 48. Alivisatos, A.P. et al. The brain activity map project and the challenge Physiol. 64, 355–405 (2002). of functional connectomics. Neuron 74, 970–974 (2012). 37. Masland, R.H. The neuronal organization of the retina. Neuron 76, 49. Koch, C. Systems biology. Modular biological complexity. Science 337, 266–280 (2012). 531–532 (2012). 38. Grimes, W.N., Zhang, J., Graydon, C.W., Kachar, B. & Diamond, J.S. Retinal 50. Varshney, L.R., Chen, B.L., Paniagua, E., Hall, D.H. & Chklovskii, D.B. parallel processors: more than 100 independent microcircuits operate Structural properties of the Caenorhabditis elegans neuronal network. within a single interneuron. Neuron 65, 873–885 (2010). PLoS Comput. Biol. 7, e1001066 (2011). © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature npg

490 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain commentary

Making sense of brain network data Olaf Sporns

New methods for mapping synaptic connections and recording neural signals generate rich and complex data on the structure and dynamics of brain networks. Making sense of these data will require a concerted effort directed at data analysis and reduction as well as computational modeling.

New empirical methodologies deliver an precision, sensitivity and reproducibility ever increasing amount of data on patterns of network maps, partly because of various of synaptic connectivity. For any given ner- methodological biases. Consequently, much vous system, the complete map of its neural attention in the area of brain networks has components and their synaptic interconnec- been devoted to addressing potential prob- tions corresponds to the connectome1. The lems and pitfalls in data acquisition, regard- Nodes connectome delivers a description of struc- ing the reconstruction and inference of Hubs tural brain connectivity, in popular parlance connectional anatomy as well as the mea- often referred to as a ‘wiring diagram’. This surement of dynamic neural interactions. wiring diagram may describe brain connec- Clearly, progress critically depends on meth- tivity at different scales, capturing synaptic odological developments that increase the Modules Edges connections among single neurons or axonal completeness and accuracy of empirically projections among brain regions. An impor- measured network maps. Figure 1 | Simple network. Schematic illustrates tant distinction must be made between such But the scientific challenges with regard to nodes and edges of the network, and its connectome data and the highly variable and the study of brain networks do not end once community structure, with network modules and dynamic patterns of neural activity and inter- data are acquired. The sheer volume and hubs highlighted. actions in the connectome, also called ‘func- complexity of these data requires sophis-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature tional connectivity’. ticated new approaches to enable efficient biological functions—once thought to Dynamic patterns of functional connec- data integration and sharing via public neu- depend on the expression of single genes or tivity are distinct from connectome maps as roinformatics resources4,5. And another key on the action of single molecules—are now

npg they are not based on synaptic connections question looms on the horizon: how can we more commonly thought to depend on large but on statistical measures derived from neu- use brain network data to gain new insight ensembles of biological components inter- ronal recordings, ranging from simple cor- into fundamental neurobiological structures acting in complex signaling, metabolic and relation to sophisticated inferences of causal and processes? In this Commentary, I argue gene-regulatory networks. dependence. Despite these important dif- that the shift toward an explicitly network- The genomic revolution has given rise to ferences between structural and functional oriented approach to studying brain function a new discipline, systems biology, dedicated connectivity, both can be represented as net- requires the development of new strategies to unraveling the structure and dynam- works comprising a set of discrete nodes and for the analysis and modeling of brain net- ics of molecular and cellular networks that edges (Fig. 1), corresponding to neural ele- work data. underpin all aspects of biological function6. ments and their pairwise associations. Accomplishing this task requires tackling ‘big Creating accurate maps of brain networks Shift toward networks data’ with new computational tools and theo- presents many methodological challenges. The rise of genomics at the end of the 20th retical models. For example, the construction of connectome century ushered in a period of profound and With the arrival of new experimental data from observations of micrometer-scale still ongoing change in the biomedical sci- methods for the comprehensive mapping of anatomical patterns among individual neu- ences. New technological developments such neural structures and for large-scale record- rons2 or from imaging of millimeter-scale as the capability to sequence whole genomes ing of neural activity both in single cells and projections among neural populations or and to create comprehensive inventories of in populations, neuroscience appears poised regions3 poses challenges regarding the biomolecules and their chemical interac- to embark on its own ‘omics’ revolution7. tions have not only transformed the empiri- Connectomics has emerged as a major focus Olaf Sporns is in the Department of Psychological cal study of biological systems, but also radi- of a new ‘systems biology of the brain’, cen- and Brain Sciences, Indiana University, Bloomington, Indiana, USA. cally altered our conceptual understanding tering on the study of anatomical networks e-mail: [email protected] of complex biological processes. Complex across multiple scales as well as the complex

nature methods | VOL.10 NO.6 | JUNE 2013 | 491 focus on mapping the brain commentary

dynamics these networks generate when modules correspond to sets of network ele- occur in functional networks; instead func- becoming active. In parallel to recent devel- ments that are densely linked with each other tional networks are a manifestation of these opments driven by modern genomics and and less so with other elements in the net- processes as they unfold in the structural systems biology, the aim of connectomics is work (that is, other modules). connectome. Therefore, studies aiming to to disclose the architecture of structural brain Module detection is an appealing strategy identify important network nodes (so-called networks and to explain the mechanisms by for analyses of brain network data sets for a ‘hubs’) or to probe the robustness of a net- which these structural networks shape and host of reasons: (i) many studies have shown work after lesion (involving deletion of nodes constrain brain function. that brain network modules correspond to and/or edges) are more appropriately carried New methods for mapping the layout and prominent functional subdivisions, for exam- out on structural networks than on func- operation of brain networks deliver huge ple, so-called resting-state networks reliably tional networks. Whereas the identification volumes of ‘big data’ that are shared in public encountered in spontaneous and task-evoked of hubs in structural networks is relatively repositories and neuroinformatics databases. brain dynamics14; (ii) the concept of modu- straightforward, it remains somewhat ambig- These data pose fundamental challenges larity has validity and can be productively uous in the context of functional networks17. for analysis and modeling, and will require applied across both structural and functional Brain networks are dynamic and change extensive data integration across multiple networks, thus enabling comparisons across across time, on both long time scales (for domains of brain and behavior. Network the two domains; (iii) the topography and example, plasticity and remodeling during theory has become vitally important in this relations among modules provide important development and learning) as well as short endeavor. By relying on networks, con- means for characterizing individual differ- time scales (for example, fast remodeling of nectomics follows a path similar to the one ences or developmental patterns in network dendritic spines in structural connectivity or already taken in systems biology, harnessing organization; (iv) modular decomposition rapid edge dynamics in functional connec- some of the same opportunities as well as offers a low-dimensional description of tivity, the latter in part resulting from neuro- encountering some of the same challenges. complex network data, which is useful for modulation18). Hence, capturing characteris- Researchers in both fields continue to grapple purposes of data reduction and compression; tic features of how networks change through with methodological problems concerning (v) modules help to define the roles of nodes time will be as important as characterizing the accurate and complete mapping of system and edges in the global network topology, for the structure of nodes and edges at each interactions, the efficient representation and example, as connectors or bridges that cross- instant in time. In this area, there is currently sharing of big data sets, and the application of link different communities15; and (vi) modu- a dearth of data-driven approaches, in part large-scale computational models. larity is thought to promote robust network because network studies in other disciplines function16. Modularity may also become an (social networks, protein networks and oth- Analysis of brain network data important tool for joining brain network data ers) are only just beginning to explicitly take Brain networks are built from empirical across multiple scales. For example, modu- network dynamics into account. The prob- data that are extraordinarily rich and com- larity should allow the identification of struc- lem is particularly pressing in the analysis of plex. There is ongoing debate concerning turally and functionally coherent populations functional networks where the topology and the proper definition of network edges and of neurons in cellular-scale data, which may strengths of dynamic interactions are contin-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature nodes. Structural brain network data usually then be aggregated into larger-scale network ually and rapidly modulated by endogenous comprise multiple measures of connectional communities at regional and systems levels. changes in state and exogenous changes in anatomy, the neurobiological interpretation Establishing links between the connec- input and task. Studies of the human brain

npg of which is not straightforward, as is the tome on the one side and the dynamic pat- with electrophysiological or imaging meth- case for inferences of connection weight or terns of neural activity and the interactions ods consistently show that functional brain strength from synaptic morphology8, or from on the other requires relating data across networks rapidly reconfigure with changes axonal microstructure or myelination status9. the two different domains of structural and in the momentary demands of the environ- Functional brain network data derived from functional networks. To reiterate the dif- ment19 and exhibit nonstationarity during time-series recordings of neuronal activity ference between structural and functional ‘rest’20. Development of methods to charac- can be expressed in many ways, from sim- network data, edges in structural networks terize these network dynamics is becoming ply recording linear statistical dependen- refer to aspects of the physical infrastructure critical in the area of data representation and cies (for example, Pearson correlations) to of brain connectivity, that is, synaptic con- analysis. sophisticated inferences on dynamic causal nections jointly comprising the ‘wiring dia- Recently, far-ranging proposals for interactions10. Which of these structural gram’, whereas edges in functional networks cataloging all functional activity maps a and dynamic network measures best repre- reflect aspects of statistical dependencies given brain can generate (referred to as sents the underlying neurobiological reality among neuronal time series, corresponding the brain’s ‘functional connectome’) have remains unclear11,12. Progress in this area is to simple correlation or covariance or more been made21, culminating in the Brain important because the initial definition of sophisticated measures of nonlinear coupling Research through Advancing Innovative nodes and edges can have profound conse- or causal dependence. Neurotechnologies (BRAIN) initiative quences for how brain networks are config- These differences in edge definition entail announced in April 2013 by President ured and interpreted. clear differences in the way structural and Barack Obama. Brain networks are cen- Of particular importance for revealing the functional network data should be analyzed tral to achieving the goals of this ambitious organization of complex networks are meth- and interpreted. For example, communi- plan. Dynamic recordings of functional ods for detection of network communities cation processes governing the exchange activity maps can naturally and efficiently or modules13 (Fig. 1). Generally speaking, of information along connections do not be represented in the form of functional

492 | VOL.10 NO.6 | JUNE 2013 | nature methods commentary focus on mapping the brain

networks. Such networks, especially if the analysis of empirically measured brain plex networked system can deliver funda- recorded at single-cell and millisecond res- dynamics. Connectome data on structural mentally new insights and knowledge. This olution, will be highly dynamic and vari- connectivity can supply the coupling matrix promise can only be fulfilled with the help of able, thus posing difficult computational for such computational models, and initial new approaches to data analysis and reduc- and analytic challenges. results suggest that these models can partly tion, driven by network theory and compu- To address these challenges, the BRAIN reproduce the spatial pattern and temporal tational modeling. As increasing amounts of initiative must include intense efforts direct- dynamics of empirically observed functional brain network data become available, there ed not only at data analysis and modeling but networks26. First applications of such com- will be a growing need for neuroscientists to also toward extending fundamental network putational network models have already work more closely with colleagues in other science and theory to the brain. Such efforts provided considerable theoretical insight into disciplines who are tackling complex tech- have proven productive in other areas of neu- the network basis of functional connectivity nological, social and biological systems. roscience. Catalogs of functional activation in the resting brain27. The adoption of a network perspective, the maps recorded with noninvasive neuroimag- As such computational models become development of new ways to analyze and ing methods in the human brain are available more realistic and sophisticated, they may describe brain connectivity, and the deploy- as public repositories that can be data-mined also become important tools for explaining ment of sophisticated integrative models are and analyzed online22,23. These databases empirical observations and predicting new all important steps toward making sense of have enabled important insights regarding findings. A highly ambitious project to build brain network data. the link between functional activations, net- a working model of a human brain incor- ACKNOWLEDGMENTS works of coactivation patterns and functional porating cellular and subcellular detail has Supported by the J.S. McDonnell Foundation. connectivity. And these functional network commenced28, and structural connectome maps have been linked to various domains data coming from both micro and macro COMPETING FINANCIAL INTERESTS The author declares no competing financial interests. of behavior and cognition. scales will likely be an important ingredient. A parallel effort to construct a ‘virtual brain’ 1. Sporns, O., Tononi, G. & Kötter, R. PLoS Comput. Growing role of computational models builds on whole-brain connectome data sets Biol. 1, 245–251 (2005). 2. Denk, W., Briggman, K.L. & Helmstadter, M. Nat. One of the key rationales behind studies of to generate complex neural dynamics whose Rev. Neurosci. 13, 351–358 (2012). protein networks is that patterns of protein- spatiotemporal signatures can be compared 3. Van Essen, D.C. & Ugurbil, K. Neuroimage 62, protein interactions can be predictive of a against empirical brain recordings29, with the 1299–1310 (2012). 4. Akil, H., Martone, M.E. & van Essen, D.C. Science protein’s functional roles as well as its involve- explicit aim to create a computational tool 331, 708–712 (2011). 24 ment in disease states . The notion that that can be of clinical benefit. Simulations of 5. Marcus, D.S. et al. Front. Neuroinform. 5, 4 (2011). structure can predict function also informs brain activity in human patients, informed by 6. Hood, L. et al. Science 306, 640–643 (2004). 7. Sporns, O. Discovering the Human Connectome. one of the chief goals of connectomics, which mapping their individual connectome, may (MIT Press, 2012). is to furnish network models that can bridge be just over the horizon. 8. Bourne, J.N. & Harris, K.M. Curr. Opin. Neurobiol. brain structure and function1,7. A large body The confluence of connectome mapping 22, 372–382 (2012). of evidence suggests that neuronal struc- and computational modeling once again 9. Jbabdi, S. & Johansen-Berg, H. Brain Connect. 1, 169–183 (2011).

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature ture, and especially connectional anatomy, parallels ongoing developments in genomics 10. Friston, K.J. Brain Connect. 1, 13–36 (2011). is crucial for endowing neural elements with and systems biology, where computational 11. Smith, S.M. Neuroimage 62, 1257–1266 (2012). specific functional properties. This structure- models have become an integral compo- 12. Wig, G.S., Schlaggar, B.L. & Petersen, S.E. Ann. NY Acad. Sci. 1224, 126–146 (2011). npg function relationship holds across scales and nent of many research projects. The idea 13. Fortunato, S. Phys. Rep. 486, 75–174 (2010). species, and it can even predict individual that computer models can help understand 14. Power, J.D. et al. Neuron 72, 665–678 (2011). variations in brain responses and behavior25. the workings of a cell has a long history, and 15. Guimerà, R. & Amaral, L.A.N. Nature 433, 895– 900 (2005). The structure-function relationship opens the availability of comprehensive data on cell 16. Kashtan, N. & Alon, U. Proc. Natl. Acad. Sci. USA a crucial role for computational models that structure, genes and gene products has now 102, 13773–13778 (2005). build on structural data to predict brain put the construction of a working model of a 17. Zuo, X.N. et al. Cereb. Cortex 22, 1862–1875 30 (2012). dynamics and associated functional attri- cell within reach . 18. Bargmann, C.I. Bioessays 34, 458–465 (2012). butes. Typically, such models consist of neu- As is the case with the virtual cell, building 19. Bassett, D.S. et al. Proc. Natl. Acad. Sci. USA 108, ronal units whose dynamics are described by a virtual brain will open a wealth of new pos- 7641–7646 (2011). differential equations based on biophysics sibilities for testing and manipulating brain 20. Jones, D.T. et al. PLoS ONE 7, e39731 (2012). 21. Alivisatos, A.P. et al. Neuron 74, 970–974 (2012). (for example, ionic membrane conductance). networks at all levels and probing for archi- 22. Fox, P.T. et al. Hum. Brain Mapp. 25, 185–198 These neuronal units are then interconnected tectural features that are necessary for effi- (2005). via a structural network of synaptic links that cient and flexible network function. As the 23. Yarkoni, T. et al. Nat. Methods 8, 665–670 (2011). 24. Vazquez, A. et al. Nat. Biotechnol. 21, 697–700 allows units to communicate and influence emerging field of connectomics continues to (2003). each other’s dynamic state. Neural com- mature, computational models will become 25. Saygin, Z.M. et al. Nat. Neurosci. 15, 321–327 munication may involve noisy signal trans- indispensable tools for guiding empirical (2012). 26. Adachi, Y. et al. Cereb. Cortex 22, 1586–1592 mission and conduction delays, and neural studies of brain networks. (2012). activity is typically tracked by measuring 27. Deco, G., Jirsa, V.K. & McIntosh, A.R. Nat. Rev. activation level, firing rate or membrane Conclusion Neurosci. 12, 43–56 (2011). 28. Markram, H. Sci. Am. 306, 50–55 (2012). potential across time. Hence, such models The promise of concerted efforts to map the 29. Jirsa, V.K. et al. Arch. Ital. Biol. 148, 189–205 deliver time-series data that can be analyzed brain’s structural and functional connections (2010). using methods very similar to those used for is that a renewed focus on the brain as a com- 30. Karr, J.R. et al. Cell 150, 389–401 (2012).

nature methods | VOL.10 NO.6 | JUNE 2013 | 493 commentary FOCUS ON MAPPING THE BRAIN

Why not connectomics? Joshua L Morgan & Jeff W Lichtman

Opinions diverge on whether mapping the synaptic connectivity of the brain is a good idea. Here we argue that albeit their limitations, such maps will reveal essential characteristics of neural circuits that would otherwise be inaccessible.

Neuroscience is in its heyday: large initia- not surprisingly, see the problem in struc- ficient justification for such studies. But tives in Europe1 and the United States2 are tural terms. The nervous system is a physi- not everyone agrees. First, connectomics getting under way. The annual neuroscience cal tissue that is quite unlike other organ can require an industrialized effort that is meetings, with their tens of thousands of systems4. For one thing, it contains far more akin to initiatives that allowed genomics attendees, feel like small cities. The num- types of cells. The retina alone has more to flourish. In light of present severe bud- ber of neuroscience research papers pub- than 50 different kinds of cells5, whereas the get limitations, connectomics might be lished more than doubled over the last two liver has closer to five. Second, neurons, by ill-advised. Moreover, some have argued decades, overtaking those from fields such virtue of their complicated geometry, con- that pursuing connectomics would be a as biochemistry, molecular biology and cell nect (via synapses) to many more cellular waste of money, even if it were free. They biology3. Given the number of momentous partners than the limited associations of have argued that anatomical maps fun- advances we hear about, surely must we not immediately adjacent cells in other organs. damentally do not reveal how the brain be on the threshold of knowing how the Third, the neuronal contacts give rise to works. Below we address some of these healthy brain works and how to fix it when directional circuits that have no analogs in arguments. it does not? other tissues. Fourth, the fine structure of Alas, the brain remains a tough nut to these neural circuits is quite diverse and dif- Top ten arguments against connectomics crack. No other organ system is associ- fers from the marked structural redundancy Number ten: circuit structure is different ated with as long a list of incurable diseases. in other organs where the same multicellu- from circuit function. One argument put

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature Worse still, for many common nervous sys- lar motif (for example, the renal nephron) forward is that the nervous system’s mac- tem illnesses there is not only no cure but no is iterated many times. Fifth, perhaps the roscopic functions (that is, behaviors) are clear idea of what is wrong. Few psychiatric most intriguing difference between neural derived directly from the functional (that

npg illnesses, learning disorders or even severe tissue and other organs is that the cellular is, electrical) properties of the neurons pain syndromes such as migraine have structure of neural tissue is a product of rather than the anatomical connections pathognomonic signs: no blood tests, radio- both genetic instruction and experience. between neurons. Hence it is the relation- graphic or electrophysiological findings, or Thus, the structure of each of our nervous ship between the firing patterns of action even brain biopsies enable diagnoses. This systems is personalized by our own set of potentials of neurons and function that is predicament is unlike that for other organ experiences. the key to bridge the gap between the cells systems, where disease is nearly always The special structural features of the of the nervous system and behavior. The associated with tissue pathological and/or nervous system are likely the reason why it recent proposal of an initiative to get a ‘brain biochemical signs. A patient may come for is more difficult to understand than other activity map’ reflects this view, as do many help because of pain in their belly, but the organs are. We contend, however, that a more initiatives for Ca2+ imaging from ever larger physician discovers the cellular or molecu- complete rendering of neural circuit synaptic numbers of nerve cells. lar cause of the pain before a treatment regi- connectivity (that is, connectomics6) would Although no one doubts that neural men begins. Why is this so different for the go a long way to solving this problem. With connections underlie signaling between brain? this understanding, diseases that manifest as nerve cells, the focus on neuronal function Those of us with an interest in a deeper abnormalities of behavior, thought or learn- (action potentials) is based in part on the understanding of the structure of the brain, ing, or as pain might become as clearly linked belief that the structural wiring details per to underlying pathological structure, as is se are insufficient to derive the firing pat- Joshua L. Morgan and Jeff W. Lichtman are in the the case for diseases in other organ systems. terns. Part of the problem is that the spik- Department of Molecular and Cellular Biology and Knowing what is wrong is a good first step to ing properties of a neuron come not only The Center for Brain Science, Harvard University, Cambridge, Massachusetts, USA. finding a solution. from the electrical signals they receive from e-mail: [email protected] or For some proponents of connectomics, their presynaptic partners but also from [email protected] the potential clinical payoffs provide suf- the distribution and characteristics of their

494 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN commentary

intrinsic voltage-gated channels. Although pinnings of a behavior of Caenorhabditis in a nervous system. The connectome cer- in principle connectomics can reveal all the elegans may actually be more difficult to tainly would not be a map of the behavioral presynaptic partners, distributions of ion decode than the circuit underlying a learned state at the moment the brain is preserved. channels would not be easily revealed by a behavior in a mammal. We think the connectome would be much connectional map. Furthermore, even if one more than that, as it could provide the knew the molecular identities and precise Number nine: signals without synapses entire behavioral capability of the brain. But location of all the voltage-gated channels in and synapses without signals. There is extracting such information may require every neuron (determined by very sophisti- abundant evidence that neurotransmit- detailing an unprecedented amount of ana- cated immunological labeling, for example), ter released at a synapse is not always tomical data. This anatomically stringent this information would still not be sufficient restricted to the ‘intended’ postsynaptic view of connectomics contrasts with inter- to provide the activity patterns underly- target. Spillover of glutamate from excit- esting recent ideas about determining the ing behavior. The reason is that behavior atory synapses has been shown to affect connections between brain cells without is often driven by the particular trains of nearby postsynaptic sites7. It is also well information about anatomy or physiology11. action potentials set up by sensory inputs. known that neuronal activity can affect A connectional matrix minus the anatomy This sensory experience is extrinsic to the the behavior of nearby glial cells, which information, of course, can reveal neither nervous system and hence inaccessible from can then convey this activity to other glial the sites of spillover nor the proximities nec- just looking at a connectional map. cells through junctions between glia. In essary for the spread of peptides, whereas all brains, there are also many signals that an anatomical connectome could reveal Response. Brains can encode experiences pass between brain cells and other organ less direct paths for neurotransmitter and and learned skills in a form that persists for systems via hormones. Steroids originat- neuropeptide signaling. However, the hor- decades or longer. The physical instantia- ing in the adrenal cortex have effects on monal milieu via blood flow or cerebrospi- tion of such stable traces of activity is not brain function as do those from sex organs. nal fluid would still remain invisible unless known, but it seems likely to us that they Growth hormone, thyroid hormone, insu- the effects of diffusible factors on behavior are embodied in the same way intrinsic lin, leptin and many others also affect brain had a structural correlate12. behaviors (such as reflexes) are: that is, in function. Furthermore, some forms of neu- There are also anatomical approaches to the specific pattern of connections between rotransmitter release do not rely on classi- deal with the problem of synapses without nerve cells. In this view, experience alters cal synapses8. Lastly, neurons often release signals. Some of the earliest evidence for connections between nerve cells to record peptides that act over large areas, using vol- a class of silent synapses came from serial a memory for later recall. Both the sensory ume transmission as opposed to restricted electron microscopy reconstructions13, and experience that lays down a memory and its transmission at adjacent synapses9. All of we imagine that such information might later recall are indeed trains of action poten- the aforementioned examples show that become available for glutaminergic synaps- tials, but in-between, and persisting for long brain function is profoundly affected by es as well. If silent synapses are structurally periods, is a stable physical structural entity chemical cues in ways that would be diffi- different from transmitting ones by virtue that holds that memory. In this sense, a map cult or impossible to infer from anatomical of the neurotransmitter receptor, then in

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature of all the things the brain has put to memory wiring diagrams. principle that difference can be revealed. is found in the structure—the connectional Conversely, it is generally accepted that a Immunolabeling for AMPA-type glutamate map. An ‘activity map’ of the brain that only substantial fraction of excitatory synapses receptors should solve this problem.

npg shows trains of action potentials would cer- can be structurally present but functionally tainly be an incomplete map, as most behav- silent10. These synapses can be switched on Number eight: ‘junk’ synapses. Given the iors and memories will not be visible in any via a Hebbian learning step. Obviously in trillions of synapses in a human brain, most finite recording session. Decoding the way the absence of knowing which synapses are individual synapses are functionally negli- experience via electrical activity becomes silent, a connectome provides a distorted gible. The brain is probably organized such stably embedded in physical neuronal net- picture of the functionally useful connec- that the precise number and identity of syn- works is the unmet challenge that connec- tions in the brain. apses are unimportant, and what really mat- tomics attempts to solve. ters is the general likelihood of connectivity. In our view, more challenging perhaps Response. It is true that that a map of syn- Thus, some fraction of the synaptic connec- will be ferreting out the relationships aptic connectivity is not identical to a map tions could rightly be called ‘junk’: they have between neural circuits and behaviors that of the signaling pathways of the brain. In no functional role but have so little cost that are intrinsic and unlearned. In most ani- terms of signals without synapses, there is such scraps persist. If junk synapses are com- mals, the behavioral repertoire is mostly no denying the importance of the brain’s mon, then going to the trouble of itemizing genetically encoded. Such inherited circuits chemical milieu on behavior. The ability of all the synaptic connections is both a waste may have arbitrary features that came about pharmacological agents to rapidly induce of time and a distraction from more impor- accidentally at some instant in an animal’s sleep, tranquility, excitement, hallucinations tant functional questions. evolution. We therefore think that geneti- and so on means that the behavioral state cally encoded behaviors could be instanti- can be dramatically altered probably with- Response. Our view is that it is premature ated in quite diverse ways compared to a out any modification to the connectome. to assign an influence to individual synaps- more limited number of rules governing Therefore, we should be careful not to es in wiring diagrams when so few complete experience-dependent formation of circuits. confuse the goals of connectomics with wiring diagrams have been described. The From this perspective, the structural under- the aims of maps of the activity patterns lesson from genomics is that many noncod-

nature methods | VOL.10 NO.6 | JUNE 2013 | 495 commentary FOCUS ON MAPPING THE BRAIN

ing sequences that initially were thought to a have been revealed from fortuitous cir- be unimportant and sometimes called junk cuits that lend themselves to easy analysis. turned out to have functional importance. For example, triad synapses in the lateral Moreover, we are not convinced that the geniculate nucleus allow the connectivity nervous system tolerates large numbers of of cell A, B and C to all be seen in a single synapses that are not serving a useful func- electron microscopy section as all the syn- tion. In fact, the synapses observed in the apses are adjacent17. If more complicated adult may reflect the small percentage of patterns of conditional connectivity exist 18,19 synapses that survive an extended period of 33% connected (as well they might ), then there is no synapse elimination during development14. way to find these without reconstructing b the entire circuit and including all the cel- Number seven: same structure, many lular elements (Fig. 1). functions. Connectomics might be tracta- ble if each behavior was partitioned to a dif- Number five: statistical synapses should ferent circuit element (that is, one circuit, suffice. Statistical mechanics was a great one behavior). However, good evidence advance in physics: when many similar shows that multiple sensations or behaviors events occur at the same time (such as the use the same neurons in different ways15. collisions of molecules in a gas), the behav- Neurons can rapidly switch their functional ior of the system can be predicted with roles in response to chemical signals such accuracy without ascertaining the behavior as peptides, other neuromodulators or Figure 1 | Potential results from two approaches of each and every particle’s trajectory. Given activity levels. Because of this flexibility, it to studying circuit connectivity. (a) Probing the trillions of synapses in a brain, is there is really not possible to assign a neuron to pairs of neurons in multiple subjects determines really any alternative except to take a statis- a function without knowing the behavioral the probability that neighboring neurons are tical approach to synaptic connectivity? connected. (b) Connectome of the same tissue state—something that is invisible in con- reveals network motifs. Members of the European Union’s initia- nectomics. tive on Future and Emerging Technologies have decided, for example, to fund the large Response. Without an understanding of lar function, it will be impossible to relate ‘Human Brain Project’. In this case, the the physiological responses of neurons to structure to function. connectional associations will be deter- various chemical messengers such as pep- mined by characterizing “...the morpholo- tides, we agree it will be difficult to define Response. In mammals, at least, variabil- gies of different types of neuron present in the roles of neurons that switch their func- ity seems to be the rule. Even the pattern different regions of the human brain. tion depending on the chemical milieu. The of nerve-muscle connections in the same Combined with modeling, the results extent to which this kind of switching is a muscle is quite different from one instan- would enable the project to predict a large 16 © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature general feature of nervous systems is not yet tiation to the next . However, variability in proportion of the short-range connectiv- known. connections does not necessarily mean that ity between neurons, without measuring But whatever the case, it is important to the result is incomprehensible. In muscle, the connectivity experimentally”1. That is,

npg emphasize one limitation of connectomics: for example, there is a skewed distribution they will create the wiring diagram without it is not a replacement for insights gained of the size of motor units that is the same having to go to the trouble of obtaining an by physiological or pharmacological stud- in each muscle even though the exact loca- anatomic connectome. ies. Rather, connectomics may associate tion and branching pattern of each axon is specific physiological phenomena with unique. Thus, each instantiation appears to Response. Because of the heterogeneous specific neural-circuit motifs so that the be a variation on the same common theme, nature of brain tissue, statistical approach- next time that motif is observed in the just as every chess match is different, but es to studying neural-circuit wiring are, of same tissue, it will signify a physiological they all obey the same rules. One of the rea- course, much more difficult than those for process without the requirement of repeat- sons we believe connectomics is necessary, a homogeneous gas. For a statistical model ing the physiological analysis each time. is that it is the only way to derive network to capture how brains actually generate Because in many nervous systems the same principles despite variability in the connec- behavior, ideally it would need to include neuron types are duplicated many thou- tivity of individual neurons. all the neuronal connectivity motifs. sands or millions of times, it seems likely Indeed, deriving network principles Unfortunately, the number of ways the that the same motifs will be repeated mul- without connectomics may be nearly dozens of different cell types are intercon- tiple times. Thus, a little physiology may impossible in some cases. Without connec- nected, and especially how these connec- go a long way. tomics, neural circuit diagrams can only be tions are contingent on the full cohort of assembled by identifying connected pairs synaptic partners of each cell, is presently Number six: same function, many struc- of cells in many different subjects. This unknown. But the advocates of statistical tures What if there is great variability in the approach, although accurate as far as it approaches to studying connectivity are circuits that give rise to a single behavioral goes, cannot reveal wiring rules as simple not unduly concerned because they believe output? Without a stereotyped pattern of as ‘cell A only connects to cell C when cell that models that combine a limited amount neural circuits that underlie a particu- B connects to cell C’. Such conditional rules of connectivity data with the right set of

496 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN commentary

learning rules will produce fully functional Number three: merely descriptive neuro- ron the list of upstream and downstream neural circuits. anatomy, just more expensive. The only neurons. This is no doubt beneficial, hav- We, however, think the only way to know hope for curing diseases and uncover- ing been the starting point for understand- whether the results of this strategy are actu- ing the ways brains work is to get insights ing the neural pathways for touch-induced ally consistent with a biological brain, is to into underlying mechanisms. Mechanistic movements22, for example. But the worm compare the predicted wiring to an actual insights come from experiments that test connectome may not be as useful as initial- connectome. This is why we submit that hypotheses by manipulating variables. ly imagined because of its surprising com- ‘measuring the connectivity experimental- There is of course a time and a place for plexity. The high level of interconnectivity ly’ is a good idea: it provides all the circuit descriptive studies in a field, and neurosci- among the 300 neurons—a revelation in motifs; with analysis, it may also provide the ence owes Cajal a great debt for his land- itself—does not lend itself to easy analysis. physical instantiation of the learning rules mark description of the cellular underpin- The interconnectedness of C. elegans neu- (see below) and is a direct path to building a nings of neural tissue, but the time for mere rons potentially provides for a much more working model of the brain. description is long past. Indeed, connec- adaptable and diverse behavioral repertoire tomics sounds modern but is really just a than one with a simpler wiring diagram but Number four: the mind is no match for throwback: it is neuroanatomy gussied up at the expense of easy comprehension by the complexity of the brain. Might it be with a fancy new word and ultra-expensive humans. the case that ‘connectomisists’ have bit off machines. Connectomics harkens to a time A particular technical limitation of the more than they can chew? Might it be that when description was all we could do. We worm’s connectome is that its inhibitory the brain’s structural complexity far out- can do more now. and excitatory connections cannot be dif- strips the complexity of the thoughts that ferentiated in electron microscopy images emanate from even the best and the bright- Response. The Hubble telescope, archae- (unlike the situation for mammals), reduc- est brains? Although some may hope that ology, the human genome and much of ing the overall power of the wiring diagram there are organizing principles and regu- structural biology are also merely descrip- to reveal new concepts. In addition, there larities that will permit substantial com- tive, but few would argue that they have are particular challenges in interpreting a pression of the connectional information not provided fundamental insights. The map of connectivity in a highly differenti- into a simpler framework, is there really hope of descriptive approaches is that they ated nervous system where each neuron has any biological reason such regularities provide specific data that lead us to general many subcellular compartments, where all are inevitable? Could the most succinct hypotheses (inductive reasoning). They can the signaling is due to local potentials, and description that embodies all that a brain reveal associations and frameworks that evolution has specialized nearly every cell does be the brain itself? If so, then there were not readily apparent before. In a vari- into a unique type. This specification con- is little point in mapping the connections ety of fields, big-data initiatives are being trasts to the millions of cells of a single class in great detail because in the best case, used to challenge the sacrosanct ‘hypothesis found in many parts of the (perhaps less one would end up with a description that first’ world view of scientific investigation. evolved) vertebrate nervous system. Thus, would be as complicated and intractable as the worm’s nervous system may use a fun-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature the brain itself. Number two: not much was learned from damentally different strategy than the one the connectomes we have. The 10-year predominating in much larger nervous sys- Response. The proximate goal of connec- effort to generate the connectome of tems in which cell types reflect populations 20 npg tomics is to generate detailed renderings of C. elegans is widely cited, but it has had of neurons whose connectivity is organized the connections between neurons in large less utility than had been imagined origi- by experience. volumes of the brain. The purposes that nally. No one can claim that the relationship As to the Human Connectome project, such connectional maps could be put to are between structure and function has been this is an important and ambitious multi- numerous. Comparisons between healthy settled in this very small nervous system. In institution study to gain information about and diseased brains might point to the phys- some respects, connectomes distract from the organization of the brains of many indi- ical underpinnings of psychiatric diseases. more mechanistic analyses because they viduals including many twins. One of its Comparisons between young and old brains reveal many more synaptic interconnections goals is to map the pathways that project might provide an understanding of what between neurons than would seem neces- between various brain regions (‘projec- kinds of network changes are associated sary. Given the struggle to make sense of the tomics’). However, none of its many goals with development and aging. Comparisons mind of a worm with about 300 neurons, it relate to describing the synaptic connectiv- between human and other primate brains is very unlikely that we will get anywhere ity of the brain. may provide insight into what underlies trying to fathom a mammalian nervous intelligence. So even if ‘understanding’ the system that could be roughly a billion times Number one. “If you want to Google on brain is not likely to be an early triumph of bigger. connectome and look [it] up, you can see connectomics, there are potentially many Speaking of big brains, there is already a some gorgeous pictures that are being made fundamental things that this kind of data big Human Connectome project21, so why of the wiring diagram of the human brain, will reveal. To put it perhaps a bit too blunt- bother with maps of mice and flies? showing you how the wires move from front ly, we think that understanding big data, to back and up and down, side to side. But may be overrated as a goal. Connectomes Response. At the very least the worm con- again, it’s a static picture. It’d be like, you can immediately provide insights even if nectome has been helpful in constraining know, taking your laptop and prying the understanding lags. circuit analysis by providing for each neu- top off and staring at the parts inside, you’d

nature methods | VOL.10 NO.6 | JUNE 2013 | 497 commentary FOCUS ON MAPPING THE BRAIN

be able to say, yeah, this is connected to ab that, but you wouldn’t know how it worked” (Francis Collins, NPR Science Friday; 5 April 2013).

Response. Maybe picking an argument in this case is not so wise, but prying the lid off a computer to see how everything is connected does not seem like such a bad idea, at least as a first step. Without know- ing the parts list and what is connected to what, how could one ever really know how Figure 2 | Two types of wiring diagrams. (a) Schematic of mammalian retina drawn by it works? Santiago Ramón y Cajal in 1901. Image courtesy of J. De Carlos. (b) Dense connectivity of the In our view, the ideal brain-imaging Twitter network of David Rodrigues as rendered by D. Rodrigues (http://www.flickr.com/photos/ technology would provide both a complete dr/2048034334/). map of synaptic connectivity as well as a complete map of the activity of all neurons seeking the physical wiring abnormalities dents with a sense of which cell types con- and synapses in real time during normal that underlie brain disorders, even if this nect with one another. These diagrams also behaviors. Even better, would be to do this requires high resolution connectomics. serve as a model of how information flows in a human being who can report on their In both situations, the structural data is through a circuit of cells because many of thoughts while behaving. Unfortunately, we overwhelming but nonetheless holds fun- these diagrams include arrows that harken are a long way from such technologies; so damental truths. back to the arrows found in the original what do we do in the meantime? To quote Francis Collins referring to Cajal diagrams (Fig. 2a). We think the computer analogy gives us the human genome project: “When you The sense one might take from these a lead. The computer can be turned off and have for the first time in front of you this kinds of diagrams is that this kind of struc- on without losing much data because the 3.1-billion-letter instruction book that tural class-wise connectivity information instructions that make it work are embed- conveys all kinds of information and all explains the way a neural circuit works. ded in its ‘static’ physicality. A deep under- kinds of mystery about humankind, you One may quibble over the extent to which standing of how information is stably stored can’t survey that going through page after the details of these connectional maps have in the structure of hard discs, the input and page without a sense of awe. I can’t help been worked out, but the implication is that output wires of each chip, the physical struc- but look at those pages and have a vague knowing this kind of neuronal-class con- tures that explain the working of those chips sense that this is giving me a glimpse of nectivity is sufficient to get some under- and so on would be enormously helpful in God’s mind”23. We likewise feel that the standing about how neurons process infor- making sense of a computer. Might the same static maps of the brain may engender awe mation.

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature be said of nervous systems? and it might not be so unrealistic to hope In certain cases, this view is probably We think that the static connectivity of that in staring into such a map we might correct. In an animal in which the major- the brain has embedded within it much get a glimpse of the human mind. ity of neurons are each genetically unique

npg of the ‘instruction book’ that guides the and each has a stereotyped connectivity, billions of impulses through networks of Why we still want to do connectomics connections of individual neurons are the neurons to ultimately generate outward The large (but we are sure still incomplete) same thing as a class-wise wiring diagram behavior. We see a connection between list of arguments above might seem discour- (because to a first approximation each cell the relationship between the connectome aging enough to put a damper on all but the is its own class). Thus, one can imagine a and the brain’s functional properties to not most irrationally enthusiastic proponents of detailed map describing exactly which only computers but also to the way the static connectomics. In addition to the reasons we cells are connected, and that map would be human genome encodes much of how an provided in our rebuttals, we remain com- exactly the same as what actually exists in organism works. mitted to pursuing connectomics because the animal. Structural defects in the genome that of the absence of one fundamental kind of But in other cases these diagrams are give rise to a range of diseases are perhaps information about nervous systems that we quite different from the actual connec- a good analogy for defects in wiring that do not think is attainable in any other way. tional array. They lack both quantitative give rise to diseases of brain function. information and ignore the different wir- In each case, tracing the causal linkages What Cajal did not do for neuroscience. ing patterns of the cells of the same class. between the static structure (be it genotype Thanks to both his genius and the extraor- Thus, a retinal schematic diagram shows or connectotype) and the ultimate pheno- dinary power of the sparse labeling of the what classes of cells are interconnected type (be it cancer or schizophrenia) is dif- Golgi stain, Cajal developed a view about but does not show critical details such as ficult. But this difficulty has not deterred neurons that has had a dominating effect on how many different amacrine cells con- cancer biologists from seeking the ulti- the way in which we imagine neural circuits verge on each retinal ganglion cell or which mate causes (the physical genes) that affect to work. Schematic textbook diagrams of particular amacrine cells are making those the likelihood of cancer, and we think retinal circuits, cortical circuits, cerebellar connections. A cortical schematic diagram should not deter neuroscientists from circuits, among many others, provide stu- does not show whether two interconnected

498 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN commentary

pyramidal cells have a common third input. Regardless of the mechanism, we current bottleneck for obtaining connec- A cerebellar schematic diagram does not wonder whether eventually a subfield of tomics data is not image acquisition but show whether the cohort of parallel fibers neuroscience will exist that is devoted to image segmentation29. Many connectomics connected to a given Purkinje cell predicts understanding the encoding and decod- data sets rely on a human painting voxels on which parallel fibers will connect with adja- ing of experience-based changes in neu- a computer screen. cent Purkinje cells. ral networks. Engramics would certainly One potential speedup is crowdsourcing This kind of information is, however, require a hefty amount of connectomics (http://eyewire.org/). Alternatively learn- being addressed by new generation of along with sophisticated analysis and ulti- ing algorithms can use human segmenta- connectomic circuit maps that reveal con- mately simulations to test the idea that a tions to guide computer efforts30. At the nectivity between dozens or hundreds particular neural network encodes a par- moment the results still require substantial of cells in a single piece of tissue17,24–26. ticular memory. human editing. All of these directions are These new efforts and several earlier being pursued, with competitions between attempts20,27, are beginning to shift the What is ahead for connectomics? groups stimulating new approaches to this paradigm from class-wise connectiv- Assuming that connectomics has a future, daunting problem (http://brainiac.mit.edu/ ity questions to questions about the par- several mostly technical obstacles will need SNEMI3D/). Ironically, much of this effort ticular pattern of connections in a cell attention. in image analysis is to make machines do class. Variability in the connections of what human brains do easily—something cells of the same class, is the essential Statistical and analytic tools. In many that we might better understand once con- feature that is lacking in classical wiring ways, maps of complete connectomes might nectomics analysis of the visual system is diagrams but that exists in actual neural look less like the canonical wiring diagrams complete. circuits. Because the vertebrate brain is of Cajal and more like modern renderings of mostly composed of many copies of each social networks (Fig. 2b). How to compare Proteome meets the connectome. The of many cell types (unique neurons such and analyze these connectional graphs of structural mapping of connections will as the Mauthner cell in teleosts are quite the brain is a new challenge that will require only be useful if the cell type of each of exceptional), this aspect of neural circuit new mathematical techniques for analyzing the neurons involved in a circuit is known. organization cannot be ignored. In sum, graphs, new statistical tests to compare one Serial reconstruction of a neuron’s shape Cajal’s use of sparse labeling did not pro- circuit with another and finally a cadre of and location will in some cases be suffi- vide a means for coming to grips with the neuroscientists who will find mining big cient for the type of neuron to be identi- way redundant populations are used in cir- data appealing. Ultimately, however, these fied. But it is likely that there is hetero- cuits. Connectomics, however, does. tools will only be useful if connectomes can geneity among neurons that look alike. be produced easily. Fortunately the past few decades have seen Engramics? In 1904, Richard Semon, a an explosion in knowledge of the molecu- German evolutionary biologist, coined Faster data collection. There are two rea- lar classes of nerve cells31. What remains the term ‘engram’ as the physical mem- sons to collect connectomics data faster. to be done is finding ways to insert that

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature ory trace that is somehow embedded in First, the production of connectomes must kind of information into mapping stud- an organism after an experience. He was eventually be easy enough so that multiple ies. Correlative fluorescence immunos- motivated to think about the physicality experimental conditions can be compared taining with electron microscopy is one 32

npg of memory in an attempt to formulate a and findings can be replicated. Second, approach . Other techniques are certainly means of inheritance of acquired char- the variety of questions that can be asked on the horizon. acteristics28. Although this Lamarckian increases as volumes enlarge and multiple notion is refuted for gene-based inheri- samples can be processed. If a complete Conclusion tance, there is little doubt that humans do map of every synapse in a human brain is Despite the many arguments against acquire information (by learning) during wanted, a serious obstacle is the fact that undertaking a connectomics analysis, we their lives that affects their behaviors and at current speeds it would take 10 million think it must be done for three reasons. then pass this information on to their chil- years to complete. First, the conditional patterns of synaptic dren (among others) to alter their behav- connectivity generated during develop- iors in nongenetic ways. A longstanding How fast can we go? The best technol- ment and by experience are inaccessible goal of neuroscience is to determine the ogy currently available for producing con- to techniques that sample from only a physical basis of an engram. nectomics data sets is high-throughput few cells at a time. Second, neuroscien- An interesting feature of the highly electron microscopy. Large-scale electron tists cannot claim to understand brains as redundant populations of neurons in microscopy data sets are now being pro- long as the network level of brain organi- mammalian nervous systems is that they duced that are about 300 cubic microm- zation is uncharted; without this detailed undergo dramatic changes in their connec- eters. For many circuits, this size is large information, neuronal physiology is tivity in early postnatal life. Because these enough to encompass multiple com- connected to systems physiology by a alterations are activity-dependent and plete axonal and dendritic arbors. There black box. Third, there is a high likeli- transform diffusely connected networks are a number of ways that this technol- hood that such exploration will reveal into many distinct subnetworks, this pro- ogy could be modified to produce a unexpected properties by shining a cess could be the physical underpinnings 10–50 fold increase in data-acquisition light (or electrons) on this most mysteri- of memory14. rates in the next few years. However, the ous tissue.

nature methods | VOL.10 NO.6 | JUNE 2013 | 499 commentary FOCUS ON MAPPING THE BRAIN

ACKNOWLEDGMENTS 6. Lichtman, J.W. & Sanes, J.R. Curr. Opin. 20. White, J.G., Southgate, E., Thomson, J.N. & Our work was supported by a Conte Center grant (US Neurobiol. 18, 346–353 (2008). Brenner, S. Phil. Trans. R. Soc. Lond. B Biol. Sci. National Institute of Mental Health), the US National 7. Diamond, J.S. Nat. Neurosci. 5, 291–292 314, 1–340 (1986). Institutes of Health, the Gatsby Charitable Trust (2002). 21. Sporns, O., Tononi, G. & Kötter, R. PLoS Comput. and Center for Brain Science Harvard University. We 8. Oláh, S. et al. Nature 461, 1278–1281 (2009). Biol. 1, e42 (2005). thank D. Rodrigues for use of his rendering of his 9. Bargmann, C.I. Bioessays 34, 458–465 (2012). 22. Chalfie, M., Sulston, J.E., Thomson, J.N. & Twitter network and J. De Carlos for Cajal’s drawing 10. Kerchner, G. & Nicoll, R. Nat. Rev. Neurosci. 9, White, G. J. Neurosci. 5, 956–964 (1985). of the retina. 813–825 (2008). 23. Swinford, S. The Sunday Times of London 11. Zador, A.M. et al. PLoS Biol. 10, e1001411 (11 June 2006). COMPETING FINANCIAL INTERESTS (2012). 24. Anderson, J.R. et al. Mol. Vis. 17, 355–379 The authors declare no competing financial interests. 12. Butcher, A.J. et al. J. Biol. Chem. 286, 11506– (2011). 1. Markram, H. et al. A Report to the European 11518 (2011). 25. Bock, D.D. et al. Nature 471, 177–182 (2011). Commission. (April 2012). 599–613 (1974). Nature 471, 183–188 (2011). 2. Collins, F. & Prabhakar, A. The White 14. Lichtman, J.W. & Colman, H. Neuron 25, 269– 27. Freed, M.A. & Sterling, P. J. Neurosci. 8, 2303– House Blog 902 (1994). & Unwin Ltd., 1921). (2 April 2013). 16. Lu, J., Tapia, J.C., White, O.L. & Lichtman, J.W. 29. Helmstaedter, M. Nat. Methods 10, 501–507 3. Pautasso, M. Sustainability 4, 3234–3247 PLoS Biol. 7, 13 (2009). (2013). (2012). 17. Sherman, S.M. & Guillery, R.W. J. Neurophysiol. 30. Kaynig, V. & Vazquez-Reina, A. IEEE Trans. Med. 4. Lichtman, J.W. & Denk, W. Science 334, 618– 76, 1367–1395 (1996). Imaging 1, 1–7 (2012). 623 (2011). 18. Seung, H.S. Neuron 62, 17–29 (2009). 31. Lein, E.S. et al. Nature 445, 168–176 (2007). 5. Masland, R.H. Curr. Opin. Neurobiol. 11, 431– 19. Song, S., Sjöström, P.J., Reigl, M., Nelson, S. & 32. Micheva, K.D., Busse, B., Weiler, N.C., O’Rourke, N. 436 (2001). Chklovskii, D.B. PLoS Biol. 3, e68 (2005). & Smith, S.J. Neuron 68, 639–653 (2010). © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature npg

500 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain perspective

Cellular-resolution connectomics: challenges of dense neural circuit reconstruction

Moritz Helmstaedter

Neuronal networks are high-dimensional graphs that single-cell and single-neurite resolution. Neuroscience are packed into three-dimensional nervous tissue at is very data-poor in this respect, contrary to the extremely high density. Comprehensively mapping assumptions made by contemporary simulation initia- these networks is therefore a major challenge. tives (The Human Brain Project7; http://www.human Although recent developments in volume electron brainproject.eu/), and it is not known whether there are microscopy imaging have made data acquisition cases in which the measurement of cellular-resolution feasible for circuits comprising a few hundreds to a connectivity graphs can in fact be replaced by simpli- few thousands of neurons, data analysis is massively fied low-order statistical assumptions about neuronal lagging behind. The aim of this perspective is to wiring. To date, only a few neuronal circuits have been summarize and quantify the challenges for data analyzed comprehensively—with the connectivity map analysis in cellular-resolution connectomics and of the entire nervous system of Caenorhabditis elegans, describe current solutions involving online crowd- comprising 302 neurons8,9, being the most notable and sourcing and machine-learning approaches. largest for decades. Brains are unique not in the number of cells they Neuronal-circuit reconstruction is so difficult comprise (about 85 billion neurons in the case of the because of the small size of neuronal processes and human brain1,2) but in the extent of direct and spe- the extremely high packing density of the neuropil cific communication between their cells via synaptic (Fig. 1a; here the term neuropil is used to describe connections (each neuron has on the order of 1,000 nervous tissue containing densely packed axons and synaptically coupled partner neurons). Mapping the dendrites, even if interspersed with cell bodies, glia © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature resulting complex connectivity graph is the goal of cells and blood vessels). For more than a century, the connectomics. At the coarse level, inter-areal projec- analysis of neuronal circuitry was therefore focused tions are tracked either noninvasively using variants on using very sparse labeling techniques that stained npg of diffusion imaging in humans3 (see Review4 in this only every 10,000th to 100,000th neuron in a given issue) or invasively using tracer injections combined volume (Fig. 1a). with high-throughput imaging in mice (Mouse Brain Only recent advances in high-throughput electron Architecture Project5 (http://brainarchitecture.org/) microscopy have made possible the imaging of sub- and Allen Brain Connectivity Atlas (http://connectivity. stantial volumes of neuronal tissue at high resolution, brain-map.org/) among other initiatives; see Review6 allowing the reconstruction of larger circuits at a much in this issue). However, the mapping coverage achieved faster pace. The main challenge, however, is the analy- using these approaches is still not sufficient to resolve sis of imaging data, which is currently the limiting step the complete set of neuronal networks contained in by several orders of magnitude. Overcoming this gap the tissue. One voxel of magnetic resonance imaging in data analysis is therefore the main methodological data with typical one-cubic-millimeter resolution, for focus of cellular-resolution connectomics. example, contains millions of neuronal cell bodies and several kilometers of neuronal wires. Resolution requirements At the other end of the resolution spectrum, the The mapping of dense circuits requires the recon- aim of connectomics is to image and analyze neuronal struction of a large fraction of the neuronal wires in a circuits densely in sufficiently large volumes but at given volume of nerve tissue (Fig. 1a). In most cases,

Structure of Neocortical Circuits Group, Max Planck Institute of Neurobiology, Munich-Martinsried, Germany. Correspondence should be addressed to M.H. ([email protected]). Received 11 February; accepted 15 April; published online 30 may 2013; doi:10.1038/nmeth.2476

nature methods | VOL.10 NO.6 | june 2013 | 501 perspective focus on mapping the brain

Figure 1 | Density of neuronal circuits and minimal resolution a A requirements. (a) Fraction of labeled neurons using light microscopy (left) B and electron microscopy (right). In light microcopy images single neurons C can be detected only because of extremely sparse staining. Images show (left to right): two synaptically coupled neurons from layers 4 and 2/3 D in rat somatosensory cortex (modified from ref. 48), sketch of cortical neurons (from ref. 49, image from Wikimedia), population of fluorescently E

labeled neurons in neocortex that were presynaptic to one postsynaptic F pyramidal neuron (modified from ref. 50), reconstruction of mouse retina inner plexiform layer (modified from ref. 36) and depiction of C. elegans 10–5 10–4 10–3 10–2 10–1 1 neurons (from wormbase.org and openworm.org, courtesy of C. Grove and Fraction of labeled neurons S. Larson). (b) Minimal resolution requirements for cellular connectomics. The thinnest structures in various neuronal circuits are indicated, from left b Mushroom body Dendritic spine Amacrine cell Pyramidal to right: mushroom body dendrite (modified from ref. 12) dendrite of layer dendrites neck dendrite cell axon 4 neuron in mouse neocortex (image by B. Cowgill), starburst amacrine cell dendrite, modified from ref. 11, GABAergic axon in mouse neocortex (image by K. Boergens and N. Marahori). Not drawn to scale. Red arrows indicate examples of the respective neurites.

30 40 50 60 70 80 neuropil is largely isotropic: neuronal processes can locally turn Smallest neurite diameter (nm) in any direction, even in cases where there is a preferred direction such as the light axis in the retina or the radial developmental axis Volume requirements: minimal circuit dimension in the neocortex. This means that the lowest-resolution dimension When mapping neuronal circuits (Fig. 2a), it is important to of the imaging techniques used should account for the smallest detect synaptic contacts between neurons, but it is in many cases neurite diameter in the chosen tissue volume. Examples of the even more important to be able to exclude synaptic connectivity smallest neurites in several parts of the nervous system (Fig. 1b) between neurons to determine the structure of a wiring diagram are dendritic spine necks (40–50 nanometer (nm) diameter, for (measuring the ‘zeros’ of the connectivity matrix). To exclude the example, in mouse hippocampus10 and neocortex), the thinnest existence of a synapse between two neurons, one has to image at parts of mouse cortical axons (~50 nm; K.M. Boergens and M.H.; least one of these two neurons in entirety. More precisely, in cases unpublished data), amacrine cell dendrites in the mouse retina where neurons have an input part (for example, pyramidal cell (~50 nm (ref. 11; K.L. Briggman, personal communication) and dendrites, which are assumed to be only postsynaptic for chemi- dendrites in the fly mushroom body (~30 nm (ref. 12)). The mini- cal synapses) and an output part (for example, the axon of a local mal required imaging resolution is then roughly half the smallest interneuron), at least one of the two neuronal arbors have to be neurite diameter, if one requires each neurite diameter to be repre- fully contained in the imaged volume. sented by at least two voxels in the three-dimensional (3D) image This notion implies that for each circuit to be mapped, one

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature data set. It may be possible in some cases to relieve this resolution can define a minimal circuit volume that fulfills the criterion of requirement slightly based on prior knowledge about the shape of sufficient completeness for a sufficient number of the relevant neurites and the continuity of intracellular organelles. Thus, the neurons (Fig. 2b–e). Then, the imaging technique should be capa-

npg minimal required imaging resolution is ~20–30 nm for most cir- ble of imaging volumes such that the smallest imaged dimension cuits but can be as small as 10–15 nm in certain model systems. is at least as large as the minimal required circuit dimension.

Figure 2 | Minimal circuit dimensions. a b Mouse retina c Mouse S1 cortex e Olfactory bulb Connectome inner plexiform layer layer 4 glomerulus Postsynaptic (a) A ‘connectome’, the connectivity matrix ‘1’ ~40 µm between a set of presynaptic and postsynatptic neurons, with 15% positive entries (‘1’, blue dots) representing the existence of a synaptic ~80 µ m connection between two neurons and 85% ~300 µ m

negative entries (‘0’, white) representing the Presynaptic absence of a synaptic connection. Detection ~1 mm of a synaptic connection is a local decision ‘0’ d Mouse cortex layer 2/3 (on the order of a few micrometers; top right). To exclude synaptic connectivity between two neurons, at least one of them has to be ~300 µ m

imaged in its entirety, yielding the notion ~500 µ m of a minimal circuit volume (dashed box) and the minimal circuit dimension (the smallest >1–2 mm of the volumes’ dimensions). (b–e) Approximate minimal circuit dimensions for several example circuits, based on the requirement to measure the existence and the absence of synaptic connections and the spatial extent of the relevant neurons in the circuit. Not drawn to scale. Dashed boxes indicate the respective minimal circuit volumes.

502 | VOL.10 NO.6 | june 2013 | nature methods focus on mapping the brain perspective

a Manual e ultrathin serial sectioning TEM(CA) Human cortex

1 cm

b Automated Mouse cortex serial sectioning, M.ret. S1 layer 2/3 1 mm tape collection w.f. SEM Z.f. larv. w.b. * ssTEM(CA) M.o.b. glom. SB C. elegans w.b. 100 µm D.m. EM m.b. M. retina

c d Minimal circuit dimension s.f. 10 µm D.m. medulla ATUM- FIB-SEM ssSEM

0 20 40 60 80 Diamond knife, FIB, serial block-face serial block-face Minimal required resolution (nm) SEM SEM

Figure 3 | Volume electron microscopy techniques for cellular connectomics and their spatial resolution and scope. (a–d) Sketches of the four most widely used methods for dense-circuit reconstruction: conventional manual ultrathin sectioning of neuropil (a) followed by TEM or TEMCA imaging16 (a), ATUM- SEM17 (b), SBEM18 (c) and FIB-SEM19 (d). In a,b, tissue is first sectioned and then (potentially later) transferred into the electron microscope for imaging. In c,d, the tissue block is abraded while imaging inside the electron microscope. (e) Approximate minimal resolution and smallest spatial dimension typically attainable with the imaging techniques in a–d, based on published results (gray shading); dashed lines indicate likely future extensions. Values also depend on the quality of staining and neurons of interest in a circuit. Approximate minimal resolution and minimal circuit dimension required to image indicated circuits. C. elegans w.b., C. elegans whole-brain reconstruction8,9,35; solid line indicates longest series from one worm and dashed line, the combined series length from three worms9. D.m. m.b., Drosophila melanogaster mushroom body; minimal required resolution based on estimate of smallest dendrites (30 nm diameter; Fig. 1b); D.m. medulla, D. melanogaster medulla, 1 cartridge (diameter of ~6 µm) with smallest processes less than 15 nm diameter. Human cortex, minimal circuit volume containing entire L5 pyramidal neuron dendrites and their local axons. Mouse cortex S1 layer 2/3, minimal circuit volume (Fig. 2d). *, mouse cortex S1 layer 4 minimal circuit volume (Fig. 2c). M.o.b.glom., mouse olfactory bulb, 1 glomerulus, only intraglomerular circuitry (Fig. 2e). M. retina s.f., mouse retina, small field (Fig. 2b). M.ret. w.f., mouse retina, wide field (including the largest amacrine and ganglion cells). Z.f. larv. w.b.: zebrafish larva whole brain.

For example, the minimal circuit volume for bipolar-to- tissue combined with slicing, abrading or evaporating the tissue. ganglion cell connectivity in the mouse retina is dictated by the Then images are combined into a 3D image volume. Only in this size of the axon of bipolar cells, which is contained in a volume sense are today’s imaging methods volume imaging methods. of ~20 µm × 20 µm × 40 µm for most bipolar cells (20 µm in the The key techniques are serial-section transmission electron 9,13–15 © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature plane of the retina, and 40 µm along the light axis; Fig. 2b). If microscopy (ssTEM ) in some cases combined with fast one wants to map the circuits of roughly a dozen bipolar cells per camera arrays (TEMCA16, Fig. 3a), automated serial-section bipolar cell type, the resulting minimal circuit volume is ~80 µm × tape-collection scanning electron microscopy (ATUM-SEM17; 11,18

npg 80 µm × 40 µm (Fig. 2b); thus the minimal circuit dimension is Fig. 3b), serial block-face SEM (SBEM , Fig. 3c) and focused 40 µm. Other examples are the mouse olfactory bulb, where the ion beam SEM (FIB-SEM19, Fig. 3d); these microscopy tech- entire circuit within one glomerulus, excluding mitral cells, is niques are reviewed in ref. 20. Briefly, these methods differ in contained in a volume of ~300 µm × 300 µm × 300 µm (Fig. 2e). the sequence and quality of cutting and imaging, and provide dif- If one aims to include mitral cells into the circuit analysis, the ferent minimal imaging resolutions and dimensions. Essentially, minimal volume increases to ~1 mm × 500 µm × 500 µm (Fig. 2e; the resolution is currently lowest, and the acquirable dimension A. Schaefer; personal communication). In mouse neocortex, the is smallest, along the cutting axis. minimal circuit volume is thought to be smallest in layer 4 of In ssTEM (Fig. 3a), tissue is first cut into ultrathin sections, primary somatosensory cortex (~300 µm × 300 µm × 300 µm, which are then imaged using TEM or speed-enhanced TEMCA. Fig. 2c), it is ~1 mm × 1 mm × 500 µm in layer 2/3, ~3 mm × As sections are cut manually and are physically maintained for 3 mm × 1 mm in layer 5 and at least as large in other cortical areas the imaging step, these techniques have a minimal resolution of where neurons have widespread axonal projections (Fig. 2d). ~40 nm. They provide large imaging areas in the plane of imag- Together, the required minimal imaging resolution, as dictated by ing but yield typically small dimensions along the cutting axis the smallest neurite diameters, and the minimal circuit dimension, because manually cutting and maintaining more than a few thou- as dictated by the types and spatial extent of the relevant neurons, sand sections is hardly possible. ATUM-SEM (Fig. 3b) overcomes differ widely between different circuit model systems. To map these the cutting-thickness limitation of ssTEM by automated slicing circuits, it is therefore crucial to choose the appropriate volume and automated tape collection of the sections, yielding a minimal imaging method (Fig. 3). resolution of up to 30 nm and providing long series of collected sections. These are imaged using SEM as the collection tape is Volume electron microscopy techniques electron-dense. The two block-face techniques (Fig. 3c,d) interlace All current imaging techniques used for cellular-resolution con- and invert the sequence of imaging and cutting: the entire tissue nectomics rely on sequential two-dimensional (2D) imaging of block is transferred into the electron microscope chamber, and

nature methods | VOL.10 NO.6 | june 2013 | 503 perspective focus on mapping the brain

a b d Contouring Difficult configuration 1

Synapse

2 1 .... 6 c Skeleton tracing e Missed branchpoint 3 Axon Dendrite

4 5 6

Figure 4 | Manual and automated reconstruction challenges in electron microscopy–based connectomics. (a) Volume reconstruction of a synaptic contact between an excitatory axon and dendritic spine in mouse somatosensory cortex imaged using SBEM (top right; K.M. Boergens and M.H., unpublished data). Images show electron microcopy slices at various depths; distance between slices 1 and 6 is 500 nm. (b,c) Manual data annotation is performed mostly either by contouring of neurite cross-sections (b) or center-line reconstruction (skeleton tracing, c). Contouring is ~50 times faster but provides no direct volume reconstruction. Modified from ref. 51. (d) Sketch of three possible neurite configurations for a configuration that automated algorithms typically cannot resolve but human annotators can (panels show three possible configurations). Dashed lines, imaging plane in which analysis is most difficult. Note that one neurite runs parallel to imaging plane (blue and green). Such a configuration is only resolvable if minimal resolution is greater than the smallest neurite diameter. Analysis may be facilitated when including priors about neurite diameters (pink neurite). (e) Branchpoint that can be missed by human annotators (unforced attention-related error25) corresponding to the dashed line in scheme on the left (from ref. 25; blue and white, skeletons traced independently by two expert annotators).

images are taken from the surface of the block using SEM. Then, the foreseeable future. To reconstruct neuronal circuits from the the top of the tissue block is abraded either using a diamond- imaging data, synapses have to be identified, and the pre- and knife microtome installed into the electron microscope chamber postsynaptic neurites contributing to each synapse have to be fol- (SBEM, Fig. 3c) or using a focused ion beam attached to the elec- lowed back to their respective cell bodies to attribute the synapse tron microscope (FIB-SEM, Fig. 3d). Diamond knife–based cut- to the correct neuron. As soon as the imaging resolution and ting currently limits the abrasion thickness to 25 nm but allows staining quality are sufficient for detecting postsynaptic densities for long imaging sequences (smallest dimension is currently and synaptic vesicles (which are ~40 nm in diameter and thus at

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature 300 µm (ref. 11), K.M. Boergens and M.H.; unpublished the same scale as most minimal neurite diameters, see above), the data), whereas FIB-SEM is unique in providing a minimal detection of synapses becomes feasible. Synapse detection is more resolution of 4 nm, but so far limited in its smallest dimension difficult in systems where one presynaptic site can have multiple

npg (~40 µm (ref. 19)). postsynaptic partners (such as in the mammalian retina and the The relationship between the available electron microscopy fly optical system), presynaptic vesicle pools are small and/or imaging techniques and the resolution and size requirements of postsynaptic specializations are difficult to identify (for example, several example circuits is summarized in Figure 3e. In a nutshell, symmetrical synapses of inhibitory axons in the neocortex). ssTEM is most amenable for circuits that can tolerate a minimal In all cases, volume imaging substantially improves the required resolution of more than 40 nm and a minimal circuit detection of synapses compared to the identification of syn- dimension of less than a few hundred micrometers, whereas FIB- apses in two-dimensional images. For example, the 3D imag- SEM is uniquely suited for circuits that require very high imaging ing of a synapse in mouse neocortex (Fig. 4a, at 25 nm slice resolution, such as most circuits in the fly nervous system, but thickness using SBEM, K.M. Boergens and M.H., unpublished can tolerate smaller volumes. ATUM-SEM and SBEM are cur- data) substantially improves the reliability of synapse detec- rently best suited for dense-circuit reconstruction in systems with tion: judging the synapse from just one of the 2D images may ~25–30 nm minimal resolution and several hundred microm- be questionable, but the sequence of several 2D images of eters minimal circuit dimension. The aims of ongoing method the same synaptic location relieves uncertainty of detection. developments are to increase the spatial scope along the cutting Synapse detection is additionally improved by high-resolution dimension, increase imaging speed in the plane of imaging and 3D imaging as obtained using FIB-SEM22 and can in these cases potentially decrease the cutting or abrasion thickness (Fig. 3e). be automated22,23. Automated synapse-detection methods for SBEM and ssTEM data are likely to become available soon. Circuit reconstruction The main challenge for cellular connectomics, however, is not All high-resolution imaging techniques generate large volumes of the detection of synapses but the reconstruction of neuronal wires images, which totaled several hundred gigabytes per data set for (Fig. 4b–e). This immense difficulty has two main origins: first, published projects11,16,21, are currently around a dozen terabytes neurites vary greatly in diameter and in local entanglement, gen- per data set and will likely be several petabytes per data set in erating a substantial frequency of locations at which the path of

504 | VOL.10 NO.6 | june 2013 | nature methods focus on mapping the brain perspective

Figure 5 | Imaging and analysis times illustrating the analysis a Data acquisition Reconstruction Gaming gap in cellular connectomics. (a) Data acquisition times and 109 respective reconstruction durations for completed projects and estimates for ongoing or planned projects. Blue and red shading indicates experiment and reconstruction, respectively. 106 Data-acquisition projections assume ~5 megahertz imaging Analysis Completed speed and SBEM experiment (*); human cortex minimum Completed gap

circuit volume–acquisition projection assumes an additional Time (hours) 103 ~200-fold increase in imaging speed as potentially attainable

with multibeam electron microscopes (**). A recent C. elegans ** ** * * *** SBEM SBEM 26 ssTEM

data analysis had a tenfold increase in the speed of annotation ssTEMCA (dashed line, ***). Gaming estimates are provided for successful science games in total annotation hours and an online casual volume S1 cortex L4 game (Angry Birds) in annotation hours per year. FoldIt estimate S1 cortex L2/3 is based on the number of players; Galaxy Zoo estimate is based C. elegansMouse (refs. retina 6,7) (ref.Mouse 9) S1 cortex L4 Retina sparse (ref. 9) Human cortex dense Mouse S1 cortex L2/3 Fish larva whole brainGalaxy ZooAngry astronomy Birds per year on the reported number of classifications and an estimate of Mouse V1 cortex (ref. 14) V1 cortex sparseOlfactory (ref. bulb14) glomerulus Zebrafish larva whole brain Mouse retina dense (ref. 36) Mouse retina dense (ref. 36) Foldit protein folding (ref. 43) on average 30 seconds per classification. ( , ) Screen shots Human cortex minimum circuit b c Mouse olfactory bulb glomerulus of ongoing online gaming initiatives in cellular connectomics: C. elegans dense (refs. 6,7,24,33) Eyewire (b; image courtesy of H.S. Seung) and Brainflight (c). b c

neurites can only be revealed after intense inspection (Fig. 4d). Second, and most notably, errors of neurite continuity have a highly correlated effect such that any neurite break along the path from a synapse to its soma creates an error in the assignment of the synapse to a neuron. As these entangled paths can be as long as several millimeters in many circuits (note that this is the neu- ronal path length between synapse and soma, not the Euclidean and make attention-based errors, whereas machines are efficient distance), neurite reconstruction has to be extremely reliable to at solving easy locations but fail when neurites become small and provide tolerable error rates in the connectivity matrix. their packing is dense. In spite of these difficulties, human annotators can reconstruct The key approach to resolving the reconstruction problem in electron microscopy data by either contouring the neurites in electron microscopy–based connectomics has therefore been sequential image planes (Fig. 4b (refs. 21,24)) or by reconstruct- to combine massive human data-annotation efforts with auto- ing only the neurite center line in three dimensions (Fig. 4c mated image analysis. In essence, human annotators provide the (refs. 25,26). Most notably, human annotators can resolve difficult long-range connectivity information and solve the most difficult locations with high accuracy (Fig. 4d). However, such annota- annotation problems, and automated algorithms provide the local tion is very slow, and annotators make many unforced attention- volume reconstruction at less difficult locations. This analysis 25 © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature related mistakes (Fig. 4e). approach has been implemented in two variants: either automated In an attempt to speed up reconstruction, a main focus in con- algorithms are first used to presegment the imaging data, fol- nectomic data analysis was to devise automated reconstruction lowed by manual inspection and correction (iterated proofread- 15,22,27–33 15,21,35 npg algorithms . Unfortunately, however, these algorithms so ing , usually involving well-trained full-time annotators), far do not provide sufficient reconstruction accuracy to replace or manual annotation is parallelized, followed by automated human annotators. Currently, automated reconstructions gener- consensus computation, and by combination with local volume ate erroneous neurite breaks or erroneous neurite mergers every segmentation (forward-only annotation, usually using lightly few tens of micrometers of neuronal path length, which results trained part-time annotators such as undergraduate students25). in the loss or misassignment of most of the synapses of a given neuron. Automated algorithms can, however, substantially sup- The analysis gap in cellular connectomics port data annotation by human annotators, as discussed below. What are the resulting experiment and analysis times for cellular- Why do automated algorithms fail at resolving what human resolution connectomics today? In Figure 5a, durations for the annotators can resolve? Figure 4d is an illustration of a local neur- accomplished circuit imaging and reconstruction experiments are ite configuration that could be interpreted in multiple ways by only reported as well as estimates for ongoing and envisioned projects. subtle local changes of the membranes separating the respective Data acquisition took around several hundred to a thousand neurites. Presently available automated algorithms are notoriously hours for the reconstruction of the C. elegans connectome9 bad at resolving such locations, especially when the concerned (ssTEM), the direction-selectivity circuit analysis in the mouse neurites are very thin. It can be suspected that such neurite con- retina11 (SBEM), a local circuit analysis experiment in V1 mouse figurations can only be resolved when explicitly comparing the cortex16 (ssTEMCA) and the dense local reconstruction of the different possible segmentation solutions (Fig. 4d), which is what inner plexiform layer in mouse retina36 (SBEM). This time was the human visual system may compute when inspecting such a spread over several months for the ssTEM methods and com- location in the data. Recent developments in algorithms aim to pressed into 6–8 weeks in the SBEM experiments. account for such segmentation-based data analysis31,34. Imaging and acquisition speed has since substantially increased, Analysis of electron microscopy data is thus meeting a substan- bringing much larger volumes within the reach of a several-month tial challenge: humans can solve difficult locations but are slow continuous experiment: for example, one glomerulus in mouse

nature methods | VOL.10 NO.6 | june 2013 | 505 perspective focus on mapping the brain

olfactory bulb (~300 µm × 300 µm × 300 µm), the minimal circuit the estimates of circuit reconstruction reported in Figure 5 are volume in layer 4 of mouse S1 cortex (~300 µm × 300 µm × based on manual-algorithmic integration that already yielded a 300 µm) or the entire brain of the larva-stage zebra fish 10–50-fold increase in reconstruction efficiency. (~400 µm × 500 µm × 500 µm). Highly parallel SEM may promise Thus, on one hand, the key prospect for the improvement of auto- an additional speed-up of more than two orders of magnitude, mated analysis remains in lowering error rates (that is, generating which would even enable the imaging of a relevant volume of longer correctly reconstructed neurite segments), but on the other human cortex (~2 cm × 2 cm × 1 cm). Improvements in imag- hand, the focus is on better integration with human annotation. ing and acquisition rates appear to have moved many circuits of In an interesting twist, all automated methods rely heavily on train- interest into the realm of experimental feasibility, which is ~1,000– ing data. It is typically very expensive, however, to generate train- 2,000 hours of imaging time (Fig. 5a; continuous experiments of ing data, and this in many cases limits algorithm improvements. up to a year (~8,000 hours) may become feasible with improve- Therefore, the attempts to recruit annotators to perform large ments in automation). amounts of human annotation, as described next, also may provide However, if one considers the required reconstruction time, a critical boost to improve automated reconstruction methods. cellular-resolution connectomics imposes daunting challenges. The Online crowd-sourcing has been successful in a few scientific two recently published mouse connectomics studies made use of very fields: most notably the protein-folding game FoldIt45 and the galaxy- sparse circuit reconstructions: only around 30 neurons were recon- classification initiative Galaxy Zoo46. Although these approaches structed for the direction-selectivity circuit analysis11 and the V1 proved very fruitful in their respective settings, the time requirements cortex study16, already amounting to ~1,000 hours of analysis time. in cellular connectomics are still large by comparison. Galaxy Zoo has For dense-circuit reconstruction, the C. elegans reconstruction (over several years) probably recruited the work time required for took around 12 years of part-time work by one scientist9,26, also the reconstruction of one dense cortical circuit (Fig. 5a). The com- amounting to around a few thousand hours of work. Including parison to casual gaming, however, seems promising: even a single the additional studies8,35, the C. elegans connectome may have successful game is reported to be played millions of hours every day, consumed around 10,000–20,000 hours spread across three amounting to as many hours a year as would be needed to reconstruct decades. The most recent mammalian dense-circuit reconstruc- a relevant piece of human cortex (~1 billion hours; Fig. 5a). tion in mouse retina36 took ~30,000 hours, by hundreds of part- This setting of successful citizen-science games and the order of time undergraduate annotators. magnitude of hours spent online playing games has fueled hopes Thus, all cellular-resolution connectomics studies to date have of recruiting the required analysis via the internet. As a first pre- involved thousands to tens-of-thousands of hours of recon- requisite, data annotation had to move online, overcoming the struction. Although imaging speed has substantially increased, challenges of streaming the required data to clients’ browsers. reconstruction speed is massively lagging behind. For any of The Open Connectome project using the collaborative annota- the proposed dense-circuit reconstructions (mouse neocortex, tion toolkit for massive amounts of image data (CATMAID) soft- ­olfactory bulb, fish brain and human neocortex; Fig. 5a), analysis- ware47 focused on ssTEM and ssSEM data analysis, the Eyewire time estimates are at least one if not several orders of magnitude project for SBEM data (http://www.eyewire.org/; Fig. 5b) and the larger than what has been accomplished to date (Fig. 5a): require- Brainflight project (http://www.brainflight.org/; Fig. 5c) are nota-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature ments for these envisioned projects are around several hundred ble examples of online annotation, with Eyewire and Brainflight thousand hours of manual labor per project. These enormous using game-like features to enhance engagement. numbers constitute what can be called the analysis gap in cellular It is an exciting endeavor today to explore whether citizen sci-

npg connectomics: although imaging larger circuits is becoming more ence can help to overcome the massive analysis gap we are facing in and more feasible, reconstructing them is not. cellular-resolution connectomics. Only if hundreds of thousands of hours can be recruited will dense-circuit reconstruction be feasible. Addressing the analysis problem It is also possible that improved automation will provide the crucial How is the field approaching this massive methodological chal- gain in the efficiency of analysis, but most likely it is going to be a lenge? There are two lines of solutions. First, improvements in combination of the two that will make dense-circuit reconstruc- automated algorithms promise to increase the efficiency of recon- tion more and more realistic. Notably, it can be expected that once struction by human annotators. Second, massive online crowd- such large data sets are being reconstructed, the large amounts of sourcing initiatives are being built to ‘recruit’ the required work training data needed to create more powerful analysis algorithms hours over the internet. will be available. Then, finally, can connectomics analysis become The current accuracy of automated classifiers is at least three a routine method with applications in many neuroscience laborato- orders of magnitude less than what is needed for reconstruction ries, and cellular-resolution connectomics may further contribute of entire neurons. It is therefore unlikely to expect that automated to unraveling the computations neuronal circuits perform. algorithms will soon take over image analysis for connectomics completely. Improvements in algorithms have saturated at the Acknowledgments 27,28,32,37–39 I am grateful to the members of my laboratory for many fruitful discussions, level of voxel classification (‘voxel classifiers’ ), and specifically to K.M. Boergens, Y. Buckley, F. Isensee, N. Marahori, A. Mohn and the current focus is on segmentation-based classifiers (‘super- H. Wissler for help with generating figures, and E. Dow for discussions concerning voxel classifiers’31,34,40). Rather than waiting for such algorithms game development. I thank M. Berning, K.M. Boergens, E. Dow and A. Schaefer to reach the required accuracy, automated methods have been for helpful comments on the manuscript. made useful in reducing the analysis challenge when they were COMPETING FINANCIAL INTERESTS well integrated with human annotation15,25,41–44. For example, The author declares no competing financial interests.

506 | VOL.10 NO.6 | june 2013 | nature methods focus on mapping the brain perspective

Reprints and permissions information is available online at http://www.nature. 27. Jain, V. et al. Supervised learning of image restoration with convolutional com/reprints/index.html. networks. IEEE 11th International Conference on Computer Vision 1–8 (2007). 1. Herculano-Houzel, S. The human brain in numbers: a linearly scaled-up 28. Jain, V., Seung, H.S. & Turaga, S.C. Machines that learn to segment primate brain. Front. Human Neurosci. 3, 31 (2009). images: a crucial technology for connectomics. Curr. Opin. Neurobiol. 20, 2. Azevedo, F.A. et al. Equal numbers of neuronal and nonneuronal cells 653–666 (2010). make the human brain an isometrically scaled-up primate brain. J. Comp. 29. Turaga, S.C. et al. Convolutional networks can learn to generate Neurol. 513, 532–541 (2009). affinity graphs for image segmentation. Neural Comput. 22, 511–538 3. Wedeen, V.J. et al. The geometric structure of the brain fiber pathways. (2010). Science 335, 1628–1634 (2012). 30. Chklovskii, D.B., Vitaladevuni, S. & Scheffer, L.K. Semi-automated 4. Craddock, R.C. Imaging human connectomes at the macroscale. reconstruction of neural circuits using electron microscopy. Curr. Opin. Nat. Methods 10, 524–539 (2013). Neurobiol. 20, 667–675 (2010). 5. Bohland, J.W. et al. A proposal for a coordinated effort for the 31. Andres, B. et al. 3D segmentation of SBFSEM images of neuropil by determination of brainwide neuroanatomical connectivity in model a graphical model over supervoxel boundaries. Med. Image Anal. 16, organisms at a mesoscopic scale. PLOS Comput. Biol. 5, e1000334 796–805 (2012). (2009). 32. Andres, B., Köthe, U., Helmstaedter, M., Denk, W. & Hamprecht, F. 6. Osten, P. & Margrie, T.W. Mapping brain circuitry with a light microscope. Segmentation of SBFSEM volume data of neural tissue by hierarchical Nat. Methods 10, 515–523 (2013). classification. in Pattern Recognition, Proc. DAGM 2008, Lecture Notes in 7. Markram, H. The blue brain project. Nat. Rev. Neurosci. 7, 153–160 (2006). Computer Science (ed., G. Rigoll) 5096, 142–152 (Springer, 2008). 8. Varshney, L.R., Chen, B.L., Paniagua, E., Hall, D.H. & Chklovskii, D.B. 33. Funke, J., Andres, B., Hamprecht, F.A., Cardona, A. & Cook, M. Efficient Structural properties of the Caenorhabditis elegans neuronal network. automatic 3D-reconstruction of branching neurons from EM data. Proc. PLOS Comput. Biol. 7, e1001066 (2011). IEEE Conference on Computer Vision and Pattern Recognition 1004–1011 9. White, J.G., Southgate, E., Thomson, J.N. & Brenner, S. The structure of (2012). the nervous system of the nematode Caenorhabditis elegans. Philos. Trans. 34. Jain, V. et al. Learning to agglomerate superpixel hierarchies. in Advances R. Soc. Lond. B Biol. Sci. 314, 1–340 (1986). in Neural Information Processing Systems (eds., J. Shawe-Taylor, 10. Chicurel, M.E. & Harris, K.M. Three-dimensional analysis of the structure R.S. Zemel, P. Bartlett, F.C.N. Pereira and K.Q. Weinberger) 24, 648–656 and composition of CA3 branched dendritic spines and their synaptic (2011). relationships with mossy fiber boutons in the rat hippocampus. J. Comp. 35. Jarrell, T.A. et al. The connectome of a decision-making neural network. Neurol. 325, 169–182 (1992). Science 337, 437–444 (2012). 11. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the 36. Helmstaedter, M., Briggman, K.L., Turaga, S., Jain, V., Seung, H.S. & direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). Denk, W. Connectomic reconstruction of the inner plexiform layer in the 12. Butcher, N.J., Friedrich, A.B., Lu, Z., Tanimoto, H. & Meinertzhagen, I.A. mouse retina. Nature (in the press). Different classes of input and output neurons reveal new features in 37. Turaga, S.C. et al. Convolutional networks can learn to generate affinity microglomeruli of the adult Drosophila mushroom body calyx. J. Comp. graphs for image segmentation. Neural Comput. 22, 511–538 (2010). Neurol. 520, 2185–2201 (2012). 38. Mishchenko, Y. Automation of 3D reconstruction of neural tissue from 13. Fahrenbach, W.H. Anatomical circuitry of lateral inhibition in the eye large volume of conventional serial section transmission electron of the horseshoe crab, Limulus polyphemus. Proc. R. Soc. Lond. B micrographs. J. Neurosci. Methods 176, 276–289 (2009). Biol. Sci. 225, 219–249 (1985). 39. Seyedhosseini, M. et al. Detection of neuron membranes in electron 14. Harris, K.M. et al. Uniform serial sectioning for transmission electron microscopy images using multi-scale context and radon-like features. microscopy. J. Neurosci. 26, 12101–12103 (2006). Med. Image Comput. Comput. Assist. Interv. 14, 670–677 (2011). 15. Mishchenko, Y. et al. Ultrastructural analysis of hippocampal neuropil 40. Vazquez-Reina, A. et al. Segmentation fusion for connectomics. IEEE from the connectomics perspective. Neuron 67, 1009–1020 (2010). International Conference on Computer Vision 177–184 (2011). 16. Bock, D.D. et al. Network anatomy and in vivo physiology of visual 41. Sommer, C., Straehle, C., Koethe, U. & Hamprecht, F.A. ilastik: interactive cortical neurons. Nature 471, 177–182 (2011). learning and segmentation toolkit. Proc. IEEE International Symposium on © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature 17. Hayworth, K.J., Kasthuri, N., Schalek, R. & Lichtman, J.W. Automating the Biomedical Imaging 8, 230–233 (2011). collection of ultrathin serial sections for large volume TEM reconstructions. 42. Straehle, C.N., Kothe, U., Knott, G. & Hamprecht, F.A. Carving: scalable Microsc. Microanal. 12 (suppl. 2), 86–87 (2006). interactive segmentation of neural volume electron microscopy images. 18. Denk, W. & Horstmann, H. Serial block-face scanning electron microscopy Med. Image Comput. Comput. Assist. Interv. 14, 653–660 (2011). npg to reconstruct three-dimensional tissue nanostructure. PLoS Biol. 2, e329 43. Jeong, W. et al. Ssecrett and NeuroTrace: interactive visualization and (2004). analysis tools for large-scale neuroscience data sets. IEEE Comput. 19. Knott, G., Marchman, H., Wall, D. & Lich, B. Serial section scanning Graph. Appl. 30, 58–70 (2010). electron microscopy of adult brain tissue using focused ion beam milling. 44. Roberts, M. et al. Neural process reconstruction from sparse user scribbles. J. Neurosci. 28, 2959–2964 (2008). Med. Image Comput. Comput. Assist. Interv. 14, 621–628 (2011). 20. Briggman, K.L. & Bock, D.D. Volume electron microscopy for neuronal 45. Cooper, S. et al. Predicting protein structures with a multiplayer online circuit reconstruction. Curr. Opin. Neurobiol. 22, 154–161 (2012). game. Nature 466, 756–760 (2010). 21. Cardona, A. et al. An integrated micro- and macroarchitectural analysis Example of a computer game used for scientific discovery. of the Drosophila brain by computer-assisted serial section electron 46. Land, K. et al. Galaxy Zoo: the large-scale spin statistics of spiral galaxies microscopy. PLoS Biol. 8, e1000502 (2010). in the Sloan Digital Sky Survey. Mon. Not. R. Astron. Soc. 388, 1686–1692 22. Kreshuk, A. et al. Automated detection and segmentation of synaptic (2008). contacts in nearly isotropic serial electron microscopy images. PLoS 47. Saalfeld, S., Cardona, A., Hartenstein, V. & Tomancak, P. CATMAID: ONE 6, e24899 (2011). collaborative annotation toolkit for massive amounts of image data. 23. Becker, C., Ali, K., Knott, G. & Fua, P. Learning context cues for synapse Bioinformatics 25, 1984–1986 (2009). segmentation in EM volumes. Med. Image Comput. Comput. Assist. Interv. Online annotation platform for electron microscopy image data 15, 585–592 (2012). presented in successive 2D image planes. 24. Fiala, J.C. Reconstruct: a free editor for serial section microscopy. 48. Feldmeyer, D., Lubke, J., Silver, R.A. & Sakmann, B. Synaptic connections J. Microsc. 218, 52–61 (2005). between layer 4 spiny neurone-layer 2/3 pyramidal cell pairs in juvenile 25. Helmstaedter, M., Briggman, K.L. & Denk, W. High-accuracy neurite rat barrel cortex: physiology and anatomy of interlaminar signalling within reconstruction for high-throughput neuroanatomy. Nat. Neurosci. 14, a cortical column. J. Physiol. (Lond.) 538, 803–822 (2002). 1081–1088 (2011). 49. Ramón y Cajal, S. Textura del sistema nervioso del hombre y de los Algorithm-based apparatus for combining the reconstructions of vertebrados (Imprenta N. Moya, 1904). many slightly trained annotators into a consensus reconstruction, 50. Rancz, E.A. et al. Transfection via whole-cell recording in vivo: bridging a prerequisite for large-scale crowd-sourcing. single-cell physiology, genetics and connectomics. Nat. Neurosci. 14, 26. Xu, M. et al. Computer assisted assembly of connectomes from electron 527–532 (2011). micrographs: application to Caenorhabditis elegans. PLoS ONE 8, e54050 51. Helmstaedter, M. & Mitra, P.P. Computational methods and challenges for (2013). large-scale circuit mapping. Curr. Opin. Neurobiol. 22, 162–169 (2012).

nature methods | VOL.10 NO.6 | june 2013 | 507 perspective focus on mapping the brain

CLARITY for mapping the nervous system

Kwanghun Chung1,2 & Karl Deisseroth1–4

With potential relevance for brain-mapping work, of irregularly arranged lipid interfaces that character- hydrogel-based structures can now be built from izes this tissue likewise creates an effective scattering within biological tissue to allow subsequent barrier to photon penetration for optical interroga- removal of lipids without mechanical disassembly tion of mammalian brains7, unlike the Caenorhabditis of the tissue. This process creates a tissue-hydrogel elegans (worm) or larval Danio rerio (zebrafish) hybrid that is physically stable, that preserves fine nervous systems, which are more accessible owing structure, proteins and nucleic acids, and that is in part to smaller size and less myelination. Single- permeable to both visible-spectrum photons and photon microscopy can provide optical transmis- exogenous macromolecules. Here we highlight sion of information from only about 50 micrometers relevant challenges and opportunities of this below the mammalian brain surface, and even well- approach, especially with regard to integration optimized two-photon microscopy cannot be used to with complementary methodologies for brain- image deeper than about 800 micrometers, far short mapping studies. of enabling visualization of full projection patterns Mammalian brains are staggeringly complex in terms and global arrangement of cell populations in the of both scale and diversity; many billions of neurons intact brain8. are present, among them likely at least hundreds of Over the past few decades, a great deal of tech- genetically distinct cell types, with each type of cell nological innovation has been stimulated by these represented by many distinct projection patterns. challenges8–21. First, newer automated methods for CLARITY1 is a newly developed technology that can mechanical sectioning of tissue have overcome some be used to transform intact biological tissue into a of the drawbacks of traditional sectioning methods hybrid form in which tissue components are removed that were laborious, expensive and damaged the © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature and replaced with exogenous elements for increased tissue. Serial block-face mechanical9–13 or optical- accessibility and functionality. CLARITY has the ­ablative14 methods, in combination with imaging rea- potential to facilitate rapid extraction of anatomical- douts such as two-photon tomography14,15, electron npg projection information important for many fields of microscopy16 or array tomography17, have been used neuroscience research2,3; such information can be to map macroscopic to nanoscopic brain structures collected along with molecular-phenotype informa- (see Review18 in this issue). In some cases, molecu- tion at the resolution of single cells. Alone or in combi- lar labeling is built into these processes, and ongoing nation with other methods4,5, such an approach could work includes approaches for addressing additional contribute to the study of function and dysfunction in challenges such as generation of contrast in tissue this complex system. before sectioning19,20. Tools for automated analysis, In general, obtaining system-wide detailed informa- efficient reconstruction and error-free alignment also tion from neural tissue is a formidable challenge (to continue to be developed21 as well as for registration say nothing of subsequent data curation and analysis). with activity information, and indeed detailed wiring In the mammalian central nervous system, seamlessly information linked to activity has been obtained from intertwined neural processes leave little extracellular well-defined volumes22,23. space, creating barriers to macromolecule diffusion Second, optical clearing methods have been for in situ hybridization, antibody staining or other ­developed that involve immersion of the specimen in forms of molecular phenotyping deeper than the first medium that matches the refractive index of the tissue, few cellular layers of intact tissue6. The high density thereby reducing light scattering and extending

1Department of Bioengineering, Stanford University, Stanford, California, USA. 2CNC Program, Stanford University, Stanford, California, USA. 3Department of Psychiatry and Behavioral Sciences, Stanford University, Stanford, California, USA. 4Howard Hughes Medical Institute, Stanford University, Stanford, California, USA. Correspondence should be addressed to K.C. ([email protected]) and K.D. ([email protected]). Received 18 March; accepted 22 April; published online 30 may 2013; doi:10.1038/nmeth.2481

508 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain perspective

Figure 1 | Imaging of nervous system projections in the intact mouse brain with CLARITY. Adult Thy1-EYFP H line mice (4 months old) were perfused transcardially followed by hybridization of biomolecule-bound monomers into a hydrogel mesh and lipid removal as described1. The clarified mouse brain was imaged from the dorsal region to the ventral region (3.4 mm to the midpoint) using a 10× water-immersion objective. Image adapted from ref. 1.

the range of optical imaging24–26. For example, benzyl alcohol−benzyl ben- zoate is an organic solvent that effectively renders biological specimens transpar- ent but reduces the stability of fluores- cent protein signals24,25; in Scale, another clearing method, an aqueous solution is used to preserve fluorescence signals, but the rate and the extent of clearing remain 26 limiting . Notably all current tissue- 1 mm clearing methods leave the densely packed lipid bilayers intact and therefore still face challenges with regard to penetration by visible-spectrum infrastructure (in a process conceptually akin to petrification or light and molecules, making these methods largely incompatible fossilization, except that not only structure but also native bio- with whole-tissue molecular phenotyping. molecules such as proteins and nucleic acids are preserved). This The CLARITY approach1 helps address ongoing challenges outcome is achieved by first infusing small organic hydrogel- by enabling molecular and optical interrogation of large assem- monomer molecules into the intact brain along with cross- bled biological systems, such as the entire adult mouse brain. linkers and thermally triggered polymerization initiators; Light-microscopy (Fig. 1) and biochemical-phenotyping (Fig. 2) subsequent temperature elevation triggers formation, from within techniques can be used to rapidly access the entire intact clari- the brain, of a hydrogel meshwork covalently linked to native pro- fied mouse brain with fine structural resolution and molecular teins, small molecules and nucleic acids but not to lipids, which detail (to the level of spines, synapses, proteins, single-amino-acid lack the necessary reactive groups (Fig. 3a). Subsequent whole- neurotransmitters and nucleic acids) while in the same prepara- brain electrophoresis in the presence of ionic detergents actively

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature tion maintaining global structural information including brain- removes the lipids (Fig. 3b), resulting in a transparent brain- wide macroscopic connectivity. To clarify tissue (Fig. 3), lipid hydrogel hybrid that both preserves, and makes accessible, struc- bilayers are replaced with a more rigid and porous hydrogel-based tural and molecular information for visualization and analysis. npg

a 3D rendering b DR Figure 2 | Intact mouse brain molecular DR phenotyping and imaging with CLARITY. RR (a) Three-dimensional (3D) visualization of immunohistology data, showing tyrosine hydroxylase (TH)-positive neurons and fibers in the mouse brain. The intact clarified brain was 1 z = 2,460 µm stained for 6 weeks as described , with primary antibody for 2 weeks, followed by a 1-week wash, SNR SNR c VTA then stained with secondary antibody for 2 weeks followed by a 1-week wash and imaged 2,500 µm VTA from ventral side using the 10× water-immersion objective. D, V, A and P indicate dorsal, ventral, anterior and posterior, respectively. (b–d) Optical PO sections at different depths, corresponding z = 1,005 µm respectively to the upper, middle and lower

PO dashed box regions in a. Note that TH-positive d neurons are well-labeled and clearly visible even

CPu CPu at a depth of 2,460 µm in the intact brain. P CPu, caudate putamen; PO, preoptic nucleus; V VTA, ventral tegmental area; SNR, substantia D nigra; RR, retrorubral nucleus; DR, dorsal raphe. A Scale bars, 700 µm (a) and 100 µm (b–d). z = 420 µm Image adapted from ref. 1.

nature methods | VOL.10 NO.6 | JUNE 2013 | 509 perspective focus on mapping the brain

Figure 3 | CLARITY technology and instrumentation. (a) Tissue is cross- a Step 1: hydrogel monomer infusion (days 1–3) linked with formaldehyde in the presence of infused hydrogel monomers. Proteins DNA + + Thermally triggered polymerization then results in a hydrogel-tissue hybrid ER

which physically supports tissue structure and chemically incorporates 4 °C Formaldehyde Hydrogel Vesicle native biomolecules into the hydrogel mesh. (b) In electrophoretic tissue monomer clearing (ETC), an electric field is applied across the hybrid immersed in an ionic detergent solution to actively transport ionic micelles into the Plasma hybrid and extract membrane lipids out of the tissue, leaving structures membrane

and cross-linked biomolecules in place and available for imaging and Step 2: hydrogel-tissue hybridization (day 3) molecular phenotyping1. The ETC setup consists of the custom ETC chamber, a temperature-controlled buffer circulator (RE415, Lauda), 37 °C a buffer filter (McMaster) and a power supply (Bio-Rad). The sample is electrophoresed by applying 20–60 volts to the electrodes. Buffer solution is circulated through the chamber to maintain temperature and the composition of the buffer solution constant throughout the Hydrogel clearing process. The cut-through view (bottom right) shows placement of the hydrogel-embedded tissue in the sample holder (Cell Strainer, b Step 3: ETC (days 5–8) Extracted lipids BD Biosciences) located in the middle of the chamber between the two in SDS micelle electrodes. The end of each electrode exposed outside the chamber is SDS micelle connected to a power supply. Image adapted from ref. 1.

Here we address opportunities and challenges for CLARITY methods in brain mapping, in combination with other emerging Temperature-controlled genetic and imaging technologies. buffer circulator Top view: cut through Imaging methods for CLARITY Electrode CLARITY is compatible with most fluorescence microscopy tech- ETC niques. Conventional laser-scanning approaches, such as confocal Buffer Target tissue chamber and multiphoton microscopies, are well-suited to image samples filter Inlet port prepared using CLARITY because the excitation and emission Electrode Electrophoresis Electrode wavelengths of light involved can penetrate deep into the trans- power supply connector parent tissue. Objective working distance can be limiting in this setting, although already single-photon imaging of the intact clarified adult mouse brain with a 3.6-millimeter working dis- the sample is illuminated with a thin sheet of light from the side tance objective has been achieved without noticeable degrada- and fluorescence emission is collected along an axis perpendicu- tion in resolution (Fig. 1)1. Even greater depth of imaging, for lar to the plane of illumination by wide-field acquisition26–31.

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature larger brains, tissues or organisms, is possible with longer work- This unique illumination and imaging modality substantially ing distance and high-numerical-aperture objectives (such as the reduces both photobleaching and imaging time (for imaging 5–8-millimeter working distance versions available from several whole mouse brain, closer to hours at single-cell resolution, and

npg manufacturers), but it will be important for optics to be refined days at single-neurite resolution). CLARITY will naturally work and for objectives to be developed that are matched to the refrac- well with SPIM; for example, the high and uniform transparency tive index of clarified tissue to minimize aberration and maintain of clarified tissue will minimize light-plane broadening that causes both high resolution and long working distance. degradation in resolution toward the center of the tissue in SPIM Single-photon laser-scanning confocal imaging is in many and therefore enable uniform high-resolution imaging of large respects appropriate for CLARITY, given the broad range of samples. Other rapidly developing imaging technologies, such suitable fluorescent labels available for multicolor confocal as structured illumination and Bessel-beam illumination31, will interrogation of clarified tissue. But confocal microscopy does enhance resolution and may potentially allow imaging in large expose the entire depth of tissue to excitation light and therefore and intact tissue specimens even beyond the diffraction limit. The induces substantial photobleaching of fluorescent molecules, a CLARITY-SPIM combination of high resolution, independence particularly acute problem in the setting of slow, high-resolution from mechanical sectioning and reconstruction, and fast data- whole-brain imaging. Nonlinear imaging techniques such as acquisition rate may become particularly useful for distinguishing two-photon microscopy address this out-of-focus photobleach- and tracing neural projections with high throughput and accuracy ing issue, but both single and two-photon imaging methods suffer in the course of generating maps of brain connectivity. from low image acquisition rate. With these scanning methods, When indicated, after global maps of intact systems are gener- imaging a mouse brain may take days at single-cell resolution ated, individual portions of tissue such as the downstream target and months at single-neurite resolution—a key practical issue of an imaged projection may be extracted for analyses such as when dealing with the larger samples that are characteristic of electron microscopy that require sectioning. CLARITY is com- CLARITY approaches. patible with subsequent electron microscopy and preserves some For imaging large samples at high resolution and at high data- ultrastructural features such as postsynaptic densities1, although acquisition rates, we suggest that selective-plane illumination because of the absence of lipid, conventional electron microscopy microscopy (SPIM) may be a method of choice. In this technique, staining does not currently provide enough contrast to identify all

510 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain perspective

relevant ultrastructures and boundaries, pointing to the need for can be visualized35,36. This approach was validated with elec- optimization of the electron microscopy preparation and staining tron microscopy and was used to map synaptic connectivity in process. Alternatively, when combined with array tomography the nematode35 and the fruit fly37; most recently, applicability and/or super-resolution imaging in neural applications10,32, to mammalian systems was demonstrated: mammalian GRASP clarified tissue may provide for nanoscopic molecular charac- (mGRASP) revealed labeling of both excitatory and inhibitory terization of synaptic connections defined by paired presynap- synapses38. Second, in recent years many mouse lines and viral tic and postsynaptic proteins, so that the synaptic wiring logic targeting techniques have been developed for intersectional of incoming projections to a brain region can be established in labeling of spatially, genetically, synaptically and functionally detail. Of course, no single imaging modality that may be used defined neuronal circuit elements. For instance, an engineered with CLARITY provides all potentially important information on rabies virus encoding EGFP can be injected into a particular functional connectivity. For example, conventional light micros- brain region to label neurons that are presynaptic to the injection copy may not readily provide definitive identification of synaptic site39–41. More recently, activity-dependent expression of chan- contacts, whereas electron microscopy analysis does not typically nelrhodopsin (ChR2) conjugated with EYFP was used to label and reveal rich molecular information on detected synapses relevant drive a subset of synaptically recruited neurons associated with to functional properties, nor would an absence of synaptic junc- fear memory42, and a technique has been developed and validated tions between cells identified by any method demonstrate absence that allows permanent marking (by filling a cell with fluorescent of direct cellular communication, which can occur via nonlo- protein labels) of neurons across the mouse brain that were active cal and volume-transmission influences, as in the case of neuro­ during a relatively restricted time window (on a time scale of modulators or extrasynaptic neurotransmitter action. Consistent hours)43. These versatile targeting approaches, in combination with possible integrative strategies, clarified tissue appears com- with CLARITY, could provide a high-throughput approach for patible with diverse imaging readout modalities that complement globally mapping synaptically connected and synaptically acti- each other; moreover, the multiround molecular phenotyping vated populations across the brain that could complement the use capacity of CLARITY (see below) may be suitable for provid- of electron microscopy in some settings. ing enriched detail on the composition and relationships of sub­ Although new genetic strategies such as these are quite power­ cellular structures such as synapses33. ful, for studying and mapping the human brain, more conven- tional antibody-based (nongenetic) molecular phenotyping Integrating circuit maps with molecular information alone is particularly important. Lipophilic dyes (for example, A unique feature of CLARITY is its potential for intact-tissue 1,1′-dioctadecyl-3,3,3′3′-tetramethylindocarbocyanine perchlor­ molecular phenotyping studies. The hydrogel-tissue hybridi- ate) can be injected into postmortem human brain to effectively zation preserves endogenous biomolecules ranging from label local projections through passive diffusion44; however, ­neurotransmitters to proteins and nucleic acids1; soluble pro- in fixed tissue without active axonal transport, diffusion rates teins and cell-membrane proteins alike are secured by chemical (and therefore tracing distance) are substantially limited and tethering to the hydrogel mesh. Moreover, removal of the lipid may not be ideally suited to mapping long-range connectivity. membrane makes this retained molecular information accessible With CLARITY, molecular markers have already been used to

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature via passive diffusion of macromolecular probes into the tissue, identify individual structures and projections (cell bodies and and the enhanced structural integrity of clarified tissue allows fibers) in human tissue on the millimeter length scale1; although multiple rounds of antibody staining, elution or destaining and this approach has not yet been tested for tracking long-range

npg restaining that are not typically feasible with conventionally projections in the human brain, neurofilament protein staining fixed tissue1,34. Immunostaining of 1-millimeter-thick clarified reliably highlights a major subset of axonal fibers, and because tissue blocks takes days, and immunostaining of entire intact continuity of labeled projections is preserved in the intact tissue, adult mouse brains is possible on practical timescales of several fibers can be traced with diminished risk of alignment and/or weeks (Fig. 2). This process could be accelerated by increasing reconstruction error. Cell type–specific markers (for example, probe diffusion rate with electrophoresis or other methods. The parvalbumin and tyrosine hydroxylase) can also be used to label fact that CLARITY supports multiple (at least three) rounds of both cell bodies and projections in human brain tissue1. CLARITY molecular phenotyping may be of value to define cell types and may in this way help unlock a rich source of clinically relevant to link form and function in brain-mapping studies, allowing information, providing a means to interrogate and make accessi- the integration of rich local and global morphological details ble brain-bank samples that are unique or rare, for the purpose of (for example, type of synapse, shape of cell body, and informa- understanding disease mechanisms as well as the native structure, tion about dendritic arborization and axonal projections) with organization and complexity of the human brain. genetic fingerprints. In addition, the combination of CLARITY with genetic meth- Limitations, challenges and opportunities ods for identification of synapses or for labeling specific circuits CLARITY is currently in the early stages of development; innova- is also of interest. First, GFP reconstitution across synaptic tion will be needed over the coming years. First, we have noted partners (GRASP) has emerged as a light microscopy–based that tissues can expand after electrophoretic tissue clearing and ­synapse-detection technique35. In this method, two nonfluores- return to the original size after refractive-index matching1; cent split-GFP fragments can be virally expressed in the synaptic although macroscopic observation indicates that the changes in membrane of two separate neuronal populations; these two frag- volume are largely isotropic and reversible, quantitative monitor- ments reconstitute fluorescent GFP only across a synaptic cleft ing at microscopic and nanoscopic resolution will be required so that the location of synapses between the two populations to confirm that loss in structural connectivity or occurrence of

nature methods | VOL.10 NO.6 | JUNE 2013 | 511 perspective focus on mapping the brain

other tissue artifacts is minimal. Second, although adherence to ­molecular phenotyping capability of CLARITY becomes partic- the described protocol1 will minimize tissue damage, the precise ularly important. As in conventional histology, broadly staining extent to which CLARITY under best current practice secures markers (which are dense yet distinct) can be useful as landmarks specific molecular information represented in proteins, nucleic to navigate and identify brain regions of interest45. In this regard, acids and small molecules must be further explored. Only ~8% of DAPI staining is compatible with CLARITY1 and appears par- total protein content is lost in CLARITY1, substantially lower than ticularly useful as a Nissl-like registration method. with other methods over the same timescale (for example, ~24% Registration of CLARITY data with other data is both a loss with conventional fixative (paraformaldehyde) and deter- ­challenge and an opportunity. For example, in an animal- gent (0.1% Triton X-100) histological solution compositions1). subject brain in which a particular set of neurons becomes known However, the nature of the remaining loss should be investigated, (via imaging of activity 46–49 and/or via optogenetic control4) to and we note that the current chemistry of CLARITY (by design) be involved in a specific behavioral function or dysfunction, to does not preserve lipids or other molecules lacking functional then clarify the same preparation and obtain brain-wide wiring groups required for chemical tethering to the hydrogel mesh, such and molecular information on those same cells (and their con- as phosphatidylinositol 4,5-bisphosphate, or PIP2. Additional nection partners) would be of substantial value. This approach chemical-treatment strategies may need to be explored for could be used to address fundamental questions in both basic specialized experimental questions. and clinical or preclinical neuroscience but will require the It also remains to be determined how accessible different classes development of efficient workflows for the registration of the of biological information from clarified tissue may be, in clinical different types of experimental data. CLARITY may also help or animal settings, for high-content quantitative data-extraction to conceptually link future high-resolution activity maps50 with pipelines. Mapping this landscape will be of value for developing structural or functional macroscale maps (for example, http:// strategies to maximize information extraction after global (brain- www.neuroscienceblueprint.nih.gov/connectome/), providing wide) projection mapping has been completed, and may be carried an anatomical foundation that could help researchers decipher out in the context of normal function, disease states and treatment the meaning of brain-activity patterns linked to health and regimens (in mapping effects of brain-stimulation treatments or disease4,50. Beyond neuroscience, CLARITY is currently being in drug screening, for example). Another open question is how explored for the evaluation, diagnosis and prognosis of patho- long clarified tissue may be maintained or stored for such analy- logical states including cancer, infection, autoimmune disease sis, assessment, imaging or remodeling in subsequent rounds of and other clinical conditions as well as for the study of normal CLARITY processing. tissue, organ and organism function, including development CLARITY can be seen as a prototype for a general approach for and relationships of cells and tissues. Resources that may help building new structures and installing new functions from within enable the general user to establish the methodology are avail- biological systems—an approach that may find other instantia- able online (http://clarityresourcecenter.org/). tions as the technology is developed for introducing components Acknowledgments or monomers into biological tissue that are then triggered sub- We acknowledge all members of the Deisseroth laboratory for discussions and sequently to form a polymer, gel, mesh, network or assembled support. This work was funded by a US National Institutes of Health Director’s Transformative Research Award (TR01) to K.D. from the National Institute © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature structure with desired physical parameters (for example, stiff- ness, transparency, pore size, conductivity and permeability) or of Mental Health, as well as by the National Science Foundation, the Simons Foundation, the President and Provost of Stanford University, and the Howard active properties (for interfacing, catalysis or functionalization). Hughes Medical Institute. K.D. is also funded by the National Institute on Drug Abuse and the Defense Advanced Research Projects Agency Reorganization npg Relevant exogenous components may include proteins, oligonu- cleotides, stains, chemicals or even small mechanical, electronic and Plasticity to Accelerate Injury Recovery program, and the Wiegers, Snyder, Reeves, Gatsby, and Yu Foundations. K.C. is supported by the Burroughs Wellcome or optical components. These introduced components could be Fund Career Award at the Scientific Interface. designed for either constitutive or inducible functionality—in the latter case, such that they can be activated by external elements COMPETING FINANCIAL INTERESTS including heat, mechanical force, redox changes, electromagnetic The authors declare competing financial interests: details are available in the online version of the paper. triggers such as light, and other accelerators, thereby enabling temporally precise initiation of the structural and functional tis- Reprints and permissions information is available online at http://www.nature. sue transformation. The capacity to trigger functionality could com/reprints/index.html. help enable versions of this general approach (not involving lipid removal) that may be compatible with ongoing vital functions 1. Chung, K. et al. Structural and molecular interrogation of intact biological of the tissue. systems. Nature advance online publication, doi:10.1038/nature12107 (10 April 2013). On the brain-mapping front, as pioneering technologies such as 2. Petersen, C.C.H. The functional organization of the barrel cortex. Neuron array tomography and serial block face scanning electron micros- 56, 339–355 (2007). copy have already proven, turning large data sets into useful and 3. Mombaerts, P. et al. Visualizing an olfactory sensory map. Cell 87, tractable deliverables still poses an immense challenge. Relevant 675–686 (1996). 4. Deisseroth, K. Optogenetics and psychiatry: applications, challenges, and computational approaches for registration in three dimensions opportunities. Biol. Psychiatry 71, 1030–1032 (2012). will need to be developed, and particularly in the setting of 5. Kasthuri, N. & Lichtman, J.W. The rise of the ‘projectome’. Nat. Methods 4, variable, quenching and/or bleaching fluorescent signals, auto- 307–308 (2007). 6. Nicholson, C. Diffusion in brain extracellular space. Brain 6, 1277–1340 mated image acquisition, segmentation and tracing algorithms (2008). will be helpful. We note that molecular landmarks are useful for 7. Cheong, W., Prahl, S. & Welch, A. A review of the optical properties of navigating three-dimensional data, and here the intact-tissue biological tissues. IEEE J. Quantum Electronics 26, 2166–2185 (1990).

512 | VOL.10 NO.6 | JUNE 2013 | nature methods focus on mapping the brain perspective

8. Helmchen, F. & Denk, W. Deep tissue two-photon microscopy. 29. Tomer, R., Khairy, K., Amat, F. & Keller, P.J. Quantitative high-speed Nat. Methods 2, 932–940 (2005). imaging of entire developing embryos with simultaneous multiview 9. Rauschning, W. Surface cryoplanning: a technique for clinical anatomical light-sheet microscopy. Nat. Methods 9, 755–763 (2012). correlations. Uppsala J. Med. Sci. 91, 251–255 (1986). 30. Krzic, U., Gunther, S., Saunders, T.E., Streichan, S.J. & Hufnagel, L. 10. Toga, A.W., Ambach, K., Quinn, B., Hutchin, M. & Burton, J.S. Multiview light-sheet microscope for rapid in toto imaging. Nat. Methods Postmortem anatomy from cryosectioned whole human brain. J. Neurosci. 9, 730–733 (2012). Methods 54, 239–252 (1994). 31. Keller, P.J., Schmidt, A.D., Wittbrodt, J. & Stelzer, E.H.K. Reconstruction 11. Ewald, A.J., McBride, H., Reddington, M., Fraser, S.E. & Kerschmann, R. of zebrafish early embryonic development by scanned light sheet Surface imaging microscopy, an automated method for visualizing whole microscopy. Science 322, 1065–1069 (2008). embryo samples in three dimensions at high resolution. Dev. Dyn. 225, 32. Planchon, T.A. et al. Rapid three-dimensional isotropic imaging of living 369–375 (2002). cells using Bessel beam plane illumination. Nat. Methods 8, 417–423 12. McCormick, B.H. et al. Construction of anatomically correct models of (2011). mouse brain networks. Neurocomputing 58–60, 379–386 (2004). 33. Dani, A., Huang, B., Bergan, J., Dulac, C. & Zhuang, X. Superresolution 13. Li, A. et al. Micro-optical sectioning tomography to obtain a high- imaging of chemical synapses in the brain. Neuron 68, 843–856 (2010). resolution atlas of the mouse brain. Science 330, 1404–1408 (2010). 34. Micheva, K.D., Busse, B., Weiler, N.C., O’Rourke, N. & Smith, S.J. 14. Tsai, P. et al. All-optical histology using ultrashort laser pulses. Neuron Single-synapse analysis of a diverse synapse population: proteomic 39, 27–41 (2003). imaging methods and markers. Neuron 68, 639–653 (2010). 15. Ragan, T. et al. Serial two-photon tomography for automated ex vivo 35. Wählby, C. & Erlandsson, F. Sequential immunofluorescence staining and mouse brain imaging. Nat. Methods 9, 255–258 (2012). image analysis for detection of large numbers of antigens in individual 16. Denk, W. & Horstmann, H. Serial block-face scanning electron microscopy cell nuclei. Cytometry 47, 32–41 (2002). to reconstruct three-dimensional tissue nanostructure. PLoS Biol. 2, e329 36. Feinberg, E.H. et al. GFP reconstitution across synaptic partners (GRASP) (2004). defines cell contacts and synapses in living nervous systems. Neuron 57, 17. Micheva, K.D. & Smith, S.J. Array tomography: a new tool for imaging the 353–363 (2008). molecular architecture and ultrastructure of neural circuits. Neuron 55, 37. Wickersham, I.R. & Feinberg, E.H. New technologies for imaging synaptic 25–36 (2007). partners. Curr. Opin. Neurobiol. 22, 121–127 (2012). 18. Osten, P. & Margrie, T.W. Mapping brain circuitry with a light microscope. 38. Gordon, M.D. & Scott, K. Motor control in a Drosophila taste circuit. Nat. Methods 10, 515–523 (2013). Neuron 61, 373–384 (2009). 19. Livet, J. et al. Transgenic strategies for combinatorial expression 39. Kim, J. et al. mGRASP enables mapping mammalian synaptic connectivity of fluorescent proteins in the nervous system. Nature 450, 56–62 with light microscopy. Nat. Methods 9, 96–102 (2012). (2007). 40. Wickersham, I. et al. Retrograde neuronal tracing with a deletion-mutant 20. Tsai, P. et al. Correlations of neuronal and microvascular densities in rabies virus. Nat. Methods 4, 47–49 (2007). murine cortex revealed by direct counting and colocalization of cell nuclei 41. Miyamichi, K. et al. Cortical representations of olfactory input by and microvessels. J. Neurosci. 18, 14553–14570 (2009). trans-synaptic tracing. Nature 472, 191–196 (2011). 21. Kleinfeld, D. et al. Large-scale automated histology in the pursuit of 42. Liu, X. et al. Optogenetic stimulation of a hippocampal engram activates connectomes. J. Neurosci. 31, 16125–16138 (2011). fear memory recall. Nature 484, 381–385 (2012). 22. Bock, D.D. et al. Network anatomy and in vivo physiology of visual 43. Guenthner, C.J., Miyamichi, K., Yang, H., Heller, H.C. & Luo, L. Permanent cortical neurons. Nature 471, 177–182 (2011). genetic access to transiently active neurons using targeted recombination 23. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the in active populations (TRAP). Neuron (in the press). direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). 44. Godement, P., Vanselow, J., Thanos, S. & Bonhoeffer, F. A study in 24. Dodt, H., Leischner, U. & Schierloh, A. Ultramicroscopy: three-dimensional developing visual systems with a new method of staining neurones and visualization of neuronal networks in the whole mouse brain. Nat. Methods their processes in fixed tissue. Development 101, 697–713 (1987). 4, 331–336 (2007). 45. Helmstaedter, M. Cellular-resolution connectomics: challenges of dense 25. Ertürk, A., Mauch, C., Hellal, F. & Förstner, F. Three-dimensional imaging neural circuit reconstruction. Nat. Methods 10, 501–507 (2013). of the unsectioned adult spinal cord to assess axon regeneration and glial 46. Dombeck, D.A., Graziano, M.S. & Tank, D.W. Functional clustering of © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature responses after injury. Nat. Med. 18, 166–171 (2012). neurons in motor cortex determined by cellular resolution imaging in 26. Hama, H. et al. Scale: a chemical approach for fluorescence imaging and awake behaving mice. J. Neurosci. 29, 13751–13760 (2009). reconstruction of transparent mouse brain. Nat. Neurosci. 14, 1481–1488 47. Ahrens, M.B. et al. Brain-wide neuronal dynamics during motor adaptation (2011). in zebrafish. Nature 485, 471–477 (2012). npg 27. Keller, P.J. et al. Fast, high-contrast imaging of animal development with 48. Ziv, Y. et al. Long-term dynamics of CA1 hippocampal place codes. scanned light sheet-based structured-illumination microscopy. Nat. Nat. Neurosci. 16, 264–266 (2013). Methods 7, 637–642 (2010). 49. Ahrens, M.B., Orger, M.B., Robson, D.N., Li, J.M. & Keller, P.J. Whole-brain 28. Truong, T.V., Supatto, W., Koos, D.S., Choi, J.M. & Fraser, S.E. Deep and functional imaging at cellular resolution using light-sheet microscopy. fast live imaging with two-photon scanned light-sheet microscopy. Nat. Methods 10, 413–420 (2013). Nat. Methods 8, 757–760 (2011). 50. Alivisatos, P. et al. The brain activity map. Science 339, 1284–1285 (2013).

nature methods | VOL.10 NO.6 | JUNE 2013 | 513 FOCUS ON MAPPING THE BRAIN review

Mapping brain circuitry with a light microscope Pavel Osten1 & Troy W Margrie2,3

The beginning of the 21st century has seen a renaissance in light microscopy and anatomical tract tracing that together are rapidly advancing our understanding of the form and function of neuronal circuits. The introduction of instruments for automated imaging of whole mouse brains, new cell type–specific and trans-synaptic tracers, and computational methods for handling the whole-brain data sets has opened the door to neuroanatomical studies at an unprecedented scale. We present an overview of the present state and future opportunities in charting long-range and local connectivity in the entire mouse brain and in linking brain circuits to function.

Since the pioneering work of Camillo Golgi and small volumes of brain tissue by electron microscopy Santiago Ramón y Cajal at the turn of the last (see Review5 and Perspective6 in this Focus). century1,2, advances in light microscopy (LM) and Advancements in LM methods, the focus of this neurotracing methods have been central to the progress Review, are being applied to the mapping of point-to- in our understanding of anatomical organization in point connectivity between all anatomical regions in the mammalian brain. The Golgi silver-impregnation the mouse brain by means of sparse reconstructions of method allowed the visualization of neuron morpho­ anterograde and retrograde tracers7. Taking advantage logy, providing the first evidence for cell type–based of the automation of LM instruments, powerful data- and connectivity-based organization in the brain. The processing pipelines, and combinations of traditional

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature introduction of efficient neuroanatomical tracers in and modern viral vector-based tracers, teams of scien- the second half of the last century greatly increased tists at Cold Spring Harbor Laboratory (CSHL), Allen the throughput and versatility of neuronal projec- Institute for Brain Science (AIBS) and University of

npg tion mapping, which led to the identification of many California Los Angeles (UCLA) are racing to com- anatomical pathways and circuits, and revealed the basic plete a connectivity map of the mouse brain—dubbed principles of hierarchical and laminar connectivity the ‘mesoscopic connectome’—which will provide the in sensory, motor and other brain systems3,4. scientific community with online atlases for viewing The beginning of this century has seen a methods- entire anatomical data sets7. These efforts demonstrate driven renaissance in neuroanatomy, one that is dis- the transformative nature of today’s LM-based neuro- tinguished by a focus on large-scale projects that yield anatomy studies and the astonishing speed with which unprecedented amounts of anatomical data. Instead of large amounts of data can be disseminated online, and the traditional ‘cottage-industry’ approach to studying have an immediate impact on research in neuroscience one anatomical pathway at a time, the new projects aim laboratories around the world. to generate complete data sets—so-called projectomes As the mouse mesoscopic connectomes are being and connectomes—that can be used by the scientific completed, it is clear that LM methods will continue community as resources for answering specific experi- to impact the evolution of biological research and spe- mental questions. These efforts range in scale and res- cifically neuroscience: new trans-synaptic viral tracers olution from the macroscopic to the microscopic: from are being engineered to circumvent the need to resolve studies of the human brain by magnetic resonance synapses, which has constrained the interpretation of imaging to reconstructions of dense neural circuits in cell-to-cell connectivity in LM studies, and new assays

1Cold Spring Harbor Laboratory, Cold Spring Harbor, New York, USA. 2Department of Neuroscience, Physiology and Pharmacology, University College London, London, UK. 3Division of Neurophysiology, The Medical Research Council National Institute for Medical Research, London, UK. Correspondence should be addressed to P.O. ([email protected]) or T.W.M. ([email protected]). Received 2 march; accepted 15 April; published online 30 may 2013; doi:10.1038/nmeth.2477

nature methods | VOL.10 NO.6 | JUNE 2013 | 515 review FOCUS ON MAPPING THE BRAIN

a b c PMT PMT Camera

Ti:sapphire PMT Camera 515-nm laser

Piezo element Detection objective

Microtome Illumination objective Knife y y Light sheet x y x z x z Motorized x-y-z stage z Motorized x-y-z stage Motorized x-y-z stage J. Kuhl

Figure 1 | Whole-brain LM methods. (a) In STP tomography, a two-photon microscope is used to image the mouse brain in a coronal plane in a mosaic grid pattern, and a microtome sections off the imaged tissue. Piezo objective scanner can be used for z-stack imaging (image adapted from ref. 15). (b) In fMOST, confocal line-scan is used to image the brain as 1-micrometer thin section cut by a diamond knife (image adapted from ref. 16). (c) In LSFM, the cleared brain is illuminated from the side with a light sheet (blue) through an illumination objective (or cylinder lens19) and imaged in a mosaic grid pattern from the top (image adapted from ref. 20). In all instruments, the brain is moved under the objective on a motorized x-y-z stage; PMT, photomultiplier tube.

combining anatomical and functional measurements are being Three instruments have been designed that combine two- applied to bridge the traditional structure-function divide in the photon microscopy17 with subsequent tissue sectioning by study of the mammalian brain. In this Review, we aim to provide ultrashort laser pulses in all-optical histology10, by a milling an overview of today’s state of the art in LM instrumentation and machine in two-photon tissue cytometry12 or by a vibrating blade to highlight the opportunities for progress as well as the challenges microtome in serial two-photon (STP) tomography15 (Fig. 1a). that need to be overcome to transform neuronal-tracing studies Whereas in both all-optical histology and two-photon tissue into a truly quantitative science that yields comprehensive descrip- cytometry the sectioning obliterates the imaged tissue, the integra- tions of long-range and local projections and connectivity in whole tion of vibratome-based sectioning in STP tomography allows the mouse brains. We also discuss present strategies for the integration collection of the cut tissue for subsequent analysis by, for example,

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature of anatomy and function in the study of mouse brain circuits. immunohistochemistry (see below). In addition, the tissue prepa- ration by simple formaldehyde fixation and agar embedding in Automated light microscopes for whole-brain imaging STP tomography has minimal detrimental effects on fluorescence

npg The field of neuroanatomy has traditionally been associated with and brain morphology. This makes STP tomography applicable labor-intensive procedures that greatly limit the throughput of to a broad range of neuroanatomical projects that use genetically data collection. Recent efforts to automate LM instrumentation encoded fluorescent protein–based tracers, which are sensitive to have standardized and dramatically increased the throughput conditions used for fixation, dehydration and tissue clearing. This of anatomical studies. The main challenge for these methods is method is also versatile in terms of the mode and resolution of to maintain the rigorous quality of traditional neuroanatomical data collection. For example, imaging the mouse brain as a data studies, which results from detailed visual analysis, careful data set of 280 serial coronal sections, evenly spaced at 50 micrometers collection and expert data interpretation. and at x-y resolution of 1 micrometer, takes about ~21 hours and There are currently two approaches to the automation of LM generates a brain atlas–like data set of ~70 gigabytes. A complete for imaging three-dimensional (3D) whole-brain data sets: one visualization can be achieved by switching to 3D scanning of based on the integration of block-face microscopy and tissue sec- z-volume stacks between the mechanical sectioning steps, which tioning and the other based on light-sheet fluorescence micros­ allows the entire mouse brain to be imaged, for instance, at copy (LSFM) of chemically cleared tissue. The first approach has 1-micrometer x-y resolution and 2.5-micrometer z resolution in been developed for wide-field imaging, line-scan imaging, con- ~8 days, generating ~1.5 terabytes of data15. The instrument is focal microscopy and two-photon microscopy8–16. Common to commercially available from TissueVision Inc. The Allen Brain all these instruments is the motorized movement of the sample Institute is using this methodology for its Mouse Connectivity under the microscope objective for top-view mosaic imaging, fol- project (see below). lowed by mechanical removal of the imaged tissue before the next Two instruments have been designed to combine bright-field line- cycle of interleaved imaging and sectioning steps (Fig. 1a,b). As scan imaging and ultramicrotome sectioning of resin-embedded the objective is always near the tissue surface, it is possible to use tissue in methods named knife-edge scanning microscopy (known high-numerical-aperture lenses to achieve submicrometer resolu- as KESM)13 and micro-optical sectioning tomography (MOST)14 tion close to the diffraction limits of LM. (Fig. 1b). The latter was used to image Golgi-stained mouse brain

516 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN review

at 0.33 × 0.33 × 1.0 micrometer x-y-z resolution, generating >8 The pioneering effort in the field of anatomical projects applied terabytes of data in ~10 days13,14. The MOST instrument design at the scale of whole animal brains was the Allen Mouse Brain was also recently built for fluorescence imaging (fMOST) by Atlas of Gene Expression, which cataloged in situ hybridization confocal laser scanning microscopy, with the throughput of one maps for more than 20,000 genes in an online 3D digital mouse mouse brain at 1.0-micrometer voxel resolution in ~19 days16. brain atlas29,31,32. The proposal by a consortium of scientists led Knife-edge scanning microscopy imaging is now also available by Partha Mitra to generate similar LM-based atlases of ‘brain- as a commercial service from 3Scan. wide neuroanatomical connectivity’ in several animal models7 has The second approach for automated whole-brain imaging in short time spurred three independent projects, each promis- is based on LSFM (also known as selective-plane illumination ing to trace all efferent and afferent anatomical pathways in the microscopy18 and ultramicroscopy19; Fig. 1c). This approach mouse brain. The aim of the Mouse Brain Architecture Project allows fast imaging of chemically cleared ‘transparent’ mouse (http://brainarchitecture.org/) from the Mitra team at CSHL brains without the need for mechanical sectioning19,20 but, is to image >1,000 brains; the Allen Mouse Brain Connectivity at least until now, with some trade-offs for anatomical trac- Atlas project (http://connectivity.brain-map.org/) led by Hongkui ing applications. The chemical clearing procedures reduced Zeng at AIBS has a goal of imaging >2,000 brains; and the Mouse the signal of fluorescent proteins, but this problem appears to Connectome Project (http://www.mouseconnectome.org/) led be solved by a new hydrogel-based tissue transformation and by Hong-Wei Dong at UCLA has a goal of imaging 500 brains, clearing method termed CLARITY21 (see Perspective about this with each brain injected with four tracers. Whereas the CSHL methodology in this Focus22). The spatial resolution of LSFM and UCLA projects rely on automated wide-field fluorescence for the mouse brain also has been limited by the requirement for microscopy (Hamamatsu Nanozoomer 2.0 and Olympus VS110, large field-of-view objectives with low power and low numeri- respectively) to image manually sectioned brains, the Mouse cal aperture that were used to visualize the whole brain19,23. Connectivity project at the Allen Institute is being done entirely However, new objectives with long working distance and high by STP tomography15. The main strength of these efforts is in the numerical aperture, such as 8-millimeter working distance and broad range of tracers used. Given that each tracer has its own 0.9 numerical aperture objective from Olympus, promise to ena- advantages and problems33, the information derived from all three ble LSFM of the whole mouse brain at submicrometer resolution. projects will ensure generalizable interpretation of the projection If necessary, LSFM can also be combined with one of several results throughout the brain. The CSHL group uses a combination forms of structured illumination to reduce out-of-focus back- of traditional anterograde and retrograde tracers, fluorophore- ground fluorescence and improve contrast24–26. Taken together, conjugated dextran amine34 and cholera toxin B (CTB) subunit35, these modifications are likely to enhance the applicability of respectively, which are complemented by a combination of viral LSFM to anterograde tracing of thin axons at high resolution vector–based tracers, GFP-expressing adeno-associated virus in the whole mouse brain, as done by STP tomography in the (AAV)36 for anterograde tracing (Fig. 2a) and modified rabies AIBS Mouse Connectivity project (see below) and by fMOST virus27 for retrograde tracing. Although the virus-based methods in a recent report16. In addition, LSFM is well-suited for ret- are less well tested, they offer advantages in terms of the bright- rograde tracing in the mouse brain, which relies on detection ness of labeling and the possibility of cell type–specific targeting 37 © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature of retrogradely fluorescence-labeled neuronal soma that are using Cre recombinase–dependent viral vectors and trans- typically >10 micrometers in diameter. Such application was genic lines expressing Cre recombinase from cell type–specific recently demonstrated for mapping retrograde connectivity of promoters38–40. The AIBS team uses solely anterograde tracing 20 41

npg granule cells of the mouse olfactory bulb using rabies viruses by AAV-GFP viruses that label axonal arborizations with GFP that achieve high levels of fluorescent protein labeling27,28. (Fig. 2b), in many cases taking advantage of transgenic ‘driver’ mouse lines expressing Cre recombinase from cell type–specific Mesoscopic connectivity-mapping projects promoters to achieve anterograde tracing of specific neuronal cell The labeling of neurons and subsequent neuroanatomical types. Finally, the team at UCLA is using a strategy of two injec- tract tracing by LM methods has been used for over a century tions per brain, each with a mix of anterograde and retrograde to interrogate the anatomical substrate of the transmission of tracers42: CTB together with Phaseolus vulgaris leucoagglutinin43 information in the brain. Throughout those years, the credo of and FluoroGold44 together with biotinylated dextran amine42,45. neuroanatomy, ‘the gain in brain is mainly in the stain’, signified This approach has an added advantage of enabling direct visu- that progress was made mainly through the development of new alization of the convergence of inputs and outputs from across anatomical tracers. Yet despite the decades of neuroanatomical different areas in one brain42,46,47. research, the laborious nature of tissue-processing and data- The unprecedented amounts of data being collected in these visualization has kept the progress in our knowledge of brain projects means that the considerable person-hours historically circuitry at a disappointingly slow pace7. Today, neuroanatomy spent performing microscopy have largely shifted toward data stands to greatly benefit from the application of high-throughput analysis. The first step of such data analysis comprises the compi- automated LM instruments and powerful informatics tools for the lation of the serial section images for viewing as whole-brain data analysis of mouse brain data29,30. The high-resolution capacity sets at resolutions beyond the minimum geometric volume of the LM methods afford, and the fact that an entire brain data neuronal structures of interest: soma for retrograde tracing and set can be captured, makes these systems well-suited for the axons for anterograde tracing. All three projects offer a conven- systematic charting of the spatial profile and the connectivity of ient way to browse the data sets online, including high-resolution populations of neurons and even individual cells projecting over magnified views that in most cases are sufficient to visually deter- long distances. mine labeled soma and axons. All three projects use the Allen

nature methods | VOL.10 NO.6 | JUNE 2013 | 517 review FOCUS ON MAPPING THE BRAIN

a recombinase driver mouse lines in the AIBS project provide a unique feature of cell-type specificity for the interpretation of the anterograde projections. The main strength of the CSHL and UCLA efforts lies in the multiplicity of the anatomical trac- ers used. The use of multiple retrograde tracers in particular will yield useful information, as retrogradely labeled soma (>10 micrometers in diameter) are easier to quantify than thin (<1 micrometer) axon fibers. These experiments will also pro- vide an important comparison between the traditional CTB and FluoroGold tracers and the rabies virus tracer that is also being used in trans-synaptic labeling (see below) but is less well studied and may show some variation in transport affinity at different types of synapses. In summary, the LM-based mesoscopic map- ping projects are set to transform the study of the circuit wiring of the mouse brain by providing online access to whole-brain data b sets from several thousand injections of anterograde and retro- grade tracers. The informatics tools being developed to search the databases will greatly aid in parsing the large amounts of data and in accessing specific brain samples for detailed scholarly analyses by the neuroscience community.

Mapping connectivity using trans-synaptic tracers In contrast to electron microscopy methods, which provide a readout of neuronal connectivity with synapse resolution over small volumes of tissue, the whole-brain LM methods permit the assessment of projection-based connectivity between brain regions and in some cases between specific cell types in those regions but without the option of visualizing the underlying syn- aptic contacts. Trans-synaptic viruses that cross either multiple or single synapses can help to circumvent the requirement to con- Figure 2 | Primary motor cortex projection maps. (a) Mouse Brain Architecture data of AAV-GFP injected into the supragranular layers firm connectivity at resolution achieved by electron microscopy and AAV–red fluorescent protein injected in the infragranular layers because such connectivity may be inferred from the known direc- (F. Mechler and P. Mitra; unpublished data). Front (left) and lateral (right) tion and mechanism of spread of the trans-synaptic tracer. Trans- views of the volume-rendered brain (top); and coronal section image from synaptic tracers based on rabies virus, pseudorabies virus and

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature the area marked by the dashed line (center) with magnification of the herpes simplex virus, which repeatedly cross synaptic connections lower boxed region showing axonal fibers in the cerebral peduncle (left) in a retrograde or anterograde direction, are powerful tools for and magnification of the upper boxed region showing projections to the studying multistep pathways upstream and downstream from the midbrain reticular nucleus (right). Scale bars, 1,000 µm (top) and 20 µm 49–51 npg (bottom). (b) Mouse Connectivity data of a similar AAV-GFP injection starter cell population . Furthermore, modified trans-synaptic show the primary motor cortex projectome reconstructed in the Allen rabies viruses have been developed that are restricted in their Brain Explorer48 (H. Zeng; unpublished data). Inset, magnified view and spread to a single synaptic jump and thus can be used to iden- coronal section overview of projections in the ventral posteromedial (VPM) tify monosynaptic connections onto and downstream of specific nucleus of the thalamus. neuronal populations and even individual cells27,52–58 (Fig. 3). Rabies virus spreads from the initially infected cells in a trans- synaptic retrograde manner49,59. Rabies virus infection does not Mouse Brain Atlas for the registration of the coronal sections, occur via spurious spread or uptake by fibers of passage and, which will help in the cross-validation of results obtained from because it cannot cross via electrical synapses, it is an effective the different tracers. The Allen Mouse Brain Connectivity Atlas tool for unidirectional anatomical tracing60. In a modified rabies website also offers the option to view the data after projection virus system, the infection can also be cell type–targeted by encap- segmentation, which selectively highlights labeled axons, as well sulating glycoprotein-deficient rabies virus with an avian virus as in 3D in the Brain Explorer registered to the Allen Mouse Brain envelope protein (referred to as ‘SAD-∆G-EnvA’). This restricts Atlas48 (Fig. 2b). infection to only those cells that express an avian receptor pro- The second step of data analysis requires the development tein TVA that is natively found in birds but not in mammals61,62. of informatics methods for quantitation of the data sets, which Thus, the delivery of vectors driving the expression of both TVA will facilitate the interpretation of the data available online. The and rabies virus glycoprotein (RV-G) into a single cell28,54,56 Allen Mouse Brain Connectivity Atlas online tools allow the user (see below) or a specific population of cells55,63, ensures that only to search the projections between injected regions and display the targeted cell or cells will (i) be susceptible to initial infection the labeled pathways as tracks in three dimensions in the Brain and (ii) provide the replication-incompetent virus with RV-G Explorer. The CSHL and UCLA connectomes can currently be required for trans-synaptic infection64. In this system, the virus can viewed online as serial section data sets. The data from the Cre spread from the primarily infected cell or cells to the presynaptic

518 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN review

Figure 3 | Mapping the function and a b c connectivity of single cells in the mouse brain Action potential response 2 Hz Synaptic response 24 V × s in vivo. (a) Experimental setup for combined 0 single-cell physiology and trans-synaptic 10 mV 315 45 100 ms connectivity mapping. Patch pipettes have 270 90 Rabies virus TVA RV-G solutions containing DNA vectors used to drive 225 135 –78 mV the expression of the TVA and RV-G proteins. Synaptic tuning Intrinsic properties Local synaptic (b) Patch pipettes are used to perform a connectivity whole-cell recording of the intrinsic and 20 mV V2ML 330° 200 ms V1 sensory-evoked synaptic properties of a V2L single layer-5 neuron in primary visual cortex. 300° 240° Synaptic responses are averages of five traces. 210° DLG (c) Then the encapsulated modified rabies virus 150° is injected into the brain in close proximity 120° DLG to the recorded neuron. After a period of up 60° 30° to 12 days to ensure retrograde spread of the Retrograde labeling of local Synaptic properties modified rabies from the recorded neuron, the and long-range inputs brain is removed and imaged for identification of the local and long-range presynaptic inputs underlying the tuning of the recorded neuron to the direction of visual motion (polar plot). Fluorescence image on the right shows injection site with the starter cell (yellow) in the middle. Scheme (bottom left) and imaged data (bottom right) of long-range retrograde tracing. V1, primary visual cortex; V2L, secondary visual cortex (lateral); V2ML secondary visual cortex (medio-lateral). Inset shows long-range inputs from the dorsal lateral geniculate nucleus (DLG). Scale bars, 300 µm (top) and 50 µm (bottom), respectively. Images modified from ref. 54.

input cells, which become labeled by expression of a fluorescent data sets (up to several terabytes per brain), which necessitates protein. However, as the presynaptic cells do not express RV-G28, automated analytical pipelining. STP tomography is currently the virus cannot spread further. This approach thus allows the dis- the most broadly used method among the whole-brain LM covery of the identity and location of the upstream input network approaches, and there are freely available informatics tools for relative to a defined population of neurons57,58. compiling STP tomography image stacks and viewing them as Brain-region specificity and cell-type specificity for map- three-dimensional data, including algorithms that automate ping connectivity by the modified rabies virus system can be seamless stitching15. Another key challenge for charting the dis- achieved by using a Cre recombinase–dependent helper virus tribution of the labeled elements in the whole mouse brain is driving expression of TVA and RV-G and transgenic driver the process of accurate registration of the individual brain data mouse lines that express Cre recombinase in specific cell types sets onto an anatomical reference atlas. To this end, scientists at or cortical layers38,39,63. This strategy is particularly useful for AIBS have generated the open-source segmented Allen Mouse brain regions comprising many different cell types that could Brain Atlas for the adult C57BL/6 mouse29,31,32,48, which is also not be otherwise selectively targeted. Moreover, the engineering available for registration of data sets generated by STP tomogra-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature of other neurotropic trans-synaptic viruses is adding new tools phy (Figs. 2b and 4b). In addition, the so-called Waxholm space for anatomical tracing, including Cre recombinase–dependent for standardized digital atlasing67 allows comparisons of regis- anterograde tracers based on a modified H129 strain of herpes tered mouse brain data using multiple brain atlases, including the 65 66

npg simplex virus and vesicular stomatitis virus , and retrograde Allen Mouse Brain Atlas, the digital Paxinos and Franklin Mouse tracers based on a modified pseudorabies virus (H. Oyibo and Brain Atlas68, and several magnetic resonance imaging reference A. Zador, personal communication). The use of retrograde and mouse brains. The continuing development of the Waxholm space anterograde trans-synaptic viruses, in combination with whole- and other online data-analysis platforms30,69,70 will be essential brain LM methods, thus promises to afford unprecedented for standardized comparisons of mouse brain data collected in access to the upstream and downstream connectivity of specific different laboratories using different instruments. cell types in the mouse brain. The completion of the three mesoscopic connectome projects in the next several years will yield a comprehensive map of point- Present challenges and opportunities for whole-brain LM to-point connectivity between anatomical regions in the mouse As highlighted above, LM instruments for whole-brain imaging are brain7. Determining the cell-type identity of the neurons send- expected to make a substantial contribution in large-scale projects ing and receiving the connections in the brain regions will be that focus on anatomical connectivity at the level of the whole essential for interpreting the function of the brain-wide neural mouse brain. It has also become clear that the use of these instru- circuits. Immunohistochemical analyses of labeled circuits have ments will have an impact in many experimental applications in proven invaluable for ascertaining the identity of specific classes different neuroscience laboratories. It is therefore imperative that of neurons71–73 and synaptic connections52,74. The combination there exist broadly applicable image-processing, warping and ana- of immunohistochemical analysis by array tomography75,76 and lytical tools that will facilitate data sharing and across-laboratory anatomical tracing by the whole-brain LM instruments promises collaboration and validation in future neuroscience studies to be particularly powerful, as it will bring together two largely focusing on, for example, mapping whole-brain anatomical automated methodologies with complementary focus on synaptic changes during development and in response to experience. and mesoscopic connectivity, respectively. STP tomography out- One practical problem arising from the choice to scan entire puts sectioned tissue (typically 50-micrometer-thick sections15), mouse brains at high resolution relates to the handling of large which can be further resectioned, processed and reimaged by array

nature methods | VOL.10 NO.6 | JUNE 2013 | 519 review FOCUS ON MAPPING THE BRAIN

Figure 4 | Imaging induction of c-fos as a a b c means to map whole-brain activation. (a) A 3D visualization of 367,378 c-fos–GFP cells detected in 280 coronal sections of an STP tomography data set of a mouse brain after the mouse was allowed to explore a novel object for 90 seconds. (b) Examples of anatomical segmentation of the brain volume with the Allen Mouse Brain Reference Atlas labels48 modified for the 280-section STP tomography data sets: hippocampus (blue), prelimbic cortex (aqua), infralimbic cortex (orange) and piriform cortex (green). (c) Visualization of c-fos–GFP cells in the hippocampus (38,170 cells), prelimbic cortex (3,305 cells), infralimbic cortex (3,827 cells) and piriform cortex (10,910 cells) (P.O., Y. Kim and K. Umadevi Venkataraju; unpublished data).

tomography for integrating cell type–specific information into the its intracellular nature permits recording the intrinsic biophysical whole-brain data sets. Industrial-level automation of slice capture profile of the target cell, which, in turn, may reflect its func- and immunostaining can be developed to minimize manual han- tional connectivity status in the local network84. In addition, dling and enhance the integration of immunohistochemistry and by recording sensory-evoked inputs, it is possible to compare STP tomography. In addition, sectioning and immunostaining single-cell synaptic receptive fields and anatomical local and can also be applied to LSFM-imaged mouse brains20. long-range connectivity traced by LM methods54. This combi- A related, cell type–focused application of whole-brain LM natorial approach, involving single-cell electrophysiology and imaging will be to quantitatively map the distribution (the cell genetic manipulation designed for connection mapping, makes counts) of different neuronal cell types in all anatomical regions it possible to test long-standing theories regarding the extent to in the mouse brain. Several such cell count–based anatomical which emergent features of sensory cortical function manifest via studies have been done previously at smaller scales, revealing, specific wiring motifs85. for example, cell densities with respect to cortical vasculature77 As has recently been achieved for serial electron microscopy– or the density of neuronal cell types per layer in a single cortical based reconstruction86,87, it will also be valuable to functionally column78–80. Using the whole-brain LM methods, a comprehen- characterize larger local neuronal populations for registration sive anatomical atlas of different GABAergic inhibitory interneu- against LM-based connectivity data. In this sense, genetically rons81 can now be generated by imaging cell type–specific Cre encoded calcium indicators, which permit physiological charac- recombinase–mediated knock-in mouse lines38,39 crossed with terization of neuronal activity in specific cell types88–90, along with

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature Cre recombinase–dependent reporter mice expressing nuclear viral vectors for trans-synaptic labeling and LM-based tracing, GFP. These and similar data sets for other neuronal cell types will have critical complementary roles. Large-volume in vivo two- will complement the mesoscopic brain region connectivity data photon imaging of neuronal activity before ex vivo whole-brain

npg and help the interpretation of the immunohistochemistry data by imaging will establish the extent to which connectivity patterns providing a reference for total numbers of specific cell types per relate to function91 at the level of single cells, and local and long- anatomical brain region. range circuits. Interpolation of such experiments will rely on the ability to cross-register in vivo functional imaging with complete Integrating brain anatomy and function ex vivo LM connectivity data. Preliminary experiments, which The anterograde, retrograde and trans-synaptic tracing approaches already hint at the spatial spread of monosynaptic connectivity described above will yield the structural scaffold of anatomical of individual principal cortical cells, suggest that combination projections and connections throughout the mouse brain. of functional imaging and traditional anatomical-circuit recon- However, such data will not be sufficient to identify how struction may only be feasible at the local network level where specific brain regions connect to form functional circuits driving connection probability is the highest92–94. Given the broad, sparse different behaviors. Bridging whole-brain structure and func- expanse of connectivity in most brain regions and especially in tion is the next frontier in systems neuroscience, and the cortical areas, high-throughput whole-brain LM methods will development of new technologies and methods will be crucial in be imperative for complete anatomical-circuit reconstruction of achieving progress. functionally characterized local networks. The structure-function relationship of single neurons can be The amalgamation of whole-brain LM and physiological meth- examined by in vivo intracellular delivery of the DNA vectors ods for single neurons and small networks offers a powerful means required for targeting and driving trans-synaptic virus expression to study the mouse brain. An exciting application of this approach via patch pipettes in loose cell-attached mode for electroporation56 will be to trace the synaptic circuits of neurons functionally char- or via whole-cell recording54 (Fig. 3). Used in combination with acterized in head-fixed behaving animals engaged in tasks related two-photon microscopy, this single-cell delivery technique may to spatial navigation, sensorimotor integration and other com- also be targeted at fluorescently labeled neurons of specific cell plex brain functions95–97. This research will lead to the genera- types56,82,83. The whole-cell method is particularly informative, as tion of whole-brain structure-function hypotheses for specific

520 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN review

behaviors, which can then be tested for causality by optogenetic Acknowledgments methods targeted to the identified cell types and brain regions98. We thank P. Mitra, H. Zeng and Christian Niedworok for comments on the manuscript and J. Kuhl for the graphics. P.O. is supported by the US National Furthermore, LM, physiological and optogenetic methods can be Institute of Mental Health grant 1R01MH096946-01, McKnight Foundation, applied to interrogate entire brain systems in large-scale projects, Technological Innovations in Neuroscience Award and Simons Foundation for as is currently being done for the mouse visual cortex in an effort Autism Research grants 204719 and 253447. T.W.M. is supported as a Wellcome led by C. Koch and R.C. Reid at AIBS99. Trust Investigator and by the Medical Research Council MC U1175975156.

Finally, a discussion in the neuroscience community has been COMPETING FINANCIAL INTERESTS initiated regarding the feasibility of mapping activity at cellular The authors declare competing financial interests: details are available in the resolution in whole brains and linking the identified activity pat- online version of the paper. terns to brain anatomy100. Today, such experiments are possible in small, transparent organisms, as was demonstrated by two- Reprints and permissions information is available online at http://www.nature. com/reprints/index.html. photon microscopy and LSFM-based imaging of brain activity in larval zebrafish expressing the calcium indicator GCaMP89,101,102. 1. Golgi, C. Sulla Struttura Della Sostanza Grigia del Cervello. Gazz. Med. Ital. Understandably, LM-based approaches will not be useful for (Lombardia) 33, 244–246 (1873). in vivo whole-brain imaging in larger, nontransparent animals, 2. Ramón y Cajal, S. Textura del Sistema Nervioso del Hombre y de los and the invention of new disruptive technologies will likely be Vertebrados. Vol. 2 (Moya, 1904). needed to achieve the goal of mapping real-time brain activity 3. Felleman, D.J. & Van Essen, D.C. Distributed hierarchical processing in the primate cerebral cortex. Cereb. Cortex 1, 1–47 (1991). at cellular resolution in, for example, the mouse. In contrast, LM 4. Rockland, K.S. & Pandya, D.N. Laminar origins and terminations of methods can be used to map patterns of whole-brain activation cortical connections of the occipital lobe in the rhesus monkey. Brain indirectly, by post-hoc visualization of activity-induced expres- Res. 179, 3–20 (1979). 5. Craddock, R.C. et al. Imaging human connectomes at the macroscale. sion of immediate early genes, such as mouse Fos (c-fos), Arc or Nat. Methods 10, 524–539 (2013). Homer1a (ref. 103). Transgenic fluorescent immediate early gene 6. Helmstaedter, M. Cellular-resolution connectomics: challenges of dense reporter mice, such as c-fos–GFP or Arc-GFP mice104–106, can be neural circuit reconstruction. Nat. Methods 10, 501–507 (2013). trained in a specific behavior, their brains subsequently can be 7. Bohland, J.W. et al. A proposal for a coordinated effort for the determination of brainwide neuroanatomical connectivity in model imaged ex vivo, and the exact distribution of GFP-positive neurons organisms at a mesoscopic scale. PLoS Comput. Biol. 5, e1000334 (2009). can be mapped and analyzed by computational methods (Fig. 4). This paper describes the rationale for mapping connectivity in the In this approach, a statistical analysis of the counts of GFP-labeled whole mouse brain at the mesoscale level by LM. 8. Odgaard, A., Andersen, K., Melsen, F. & Gundersen, H.J. A direct method neurons can be used to identify brain regions and cell types acti- for fast three-dimensional serial reconstruction. J. Microsc. 159, 335–342 vated during behaviors but without providing information on the (1990). temporal sequence of brain region activation or the firing patterns 9. Ewald, A.J., McBride, H., Reddington, M., Fraser, S.E. & Kerschmann, R. of the activated cells. However, the development of more sensitive, Surface imaging microscopy, an automated method for visualizing whole embryo samples in three dimensions at high resolution. Dev. Dyn. 225, for instance, fluorescent RNA–based methods, may allow calibra- 369–375 (2002). tion of the cellular signal with respect to the temporal window 10. Tsai, P.S. et al. All-optical histology using ultrashort laser pulses. and the pattern of activity related to the induction of immedi- Neuron 39, 27–41 (2003). This study pioneered the approach of serial imaging by two-photon © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature ate early genes. Such calibration would considerably enhance microscopy and tissue sectioning for ex vivo collection of the power of LM-based whole-brain mapping of the induction neuroanatomical data. of immediate early genes, which, in combination with the con- 11. Sands, G.B. et al. Automated imaging of extended tissue volumes using confocal microscopy. Microsc. Res. Tech. 67, 227–239 (2005). npg nectomics data, could then be used to begin to build cellular- 12. Ragan, T. et al. High-resolution whole organ imaging using two-photon resolution models of function-based whole-brain circuits. tissue cytometry. J. Biomed. Opt. 12, 014015 (2007). 13. Mayerich, D., Abbott, L. & McCormick, B. Knife-edge scanning microscopy Conclusions for imaging and reconstruction of three-dimensional anatomical structures The advances in automated LM methods, anatomical tracers, of the mouse brain. J. Microsc. 231, 134–143 (2008). 14. Li, A. et al. Micro-optical sectioning tomography to obtain a high- physiological methods and informatics tools have begun to trans- resolution atlas of the mouse brain. Science 330, 1404–1408 (2010). form our understanding of the circuit wiring in the mouse brain. 15. Ragan, T. et al. Serial two-photon tomography for automated ex vivo The focus on the mouse as an animal model is, of course, not acci- mouse brain imaging. Nat. Methods 9, 255–258 (2012). This study introduces the method of STP tomography and dental. In addition to the generation of cell type–specific driver demonstrates its use for anterograde and retrograde tracing in the 38–40 mouse lines that allow the study of specific neuronal popula- mouse brain. tions in the normal brain, mouse genetics are used in hundreds of 16. Gong, H. et al. Continuously tracing brain-wide long-distance axonal laboratories to model gene mutations linked to heritable human projections in mice at a one-micron voxel resolution. Neuroimage 74, 87–98, (2013). disorders, including complex cognitive disorders such as autism This study demonstrates the first long-range tracing of individual and schizophrenia. Without a doubt, understanding the relation- axons in the mouse brain by fMOST. ships between brain structure and function in the genetic mouse 17. Denk, W., Strickler, J.H. & Webb, W.W. Two-photon laser scanning models will be crucial to understanding the underlying brain cir- fluorescence microscopy. Science 248, 73–76 (1990). 18. Huisken, J., Swoger, J., Del Bene, F., Wittbrodt, J. & Stelzer, E.H. Optical cuit mechanisms of these disorders. The toolbox of LM methods sectioning deep inside live embryos by selective plane illumination described here, and the continuing development of new methods, microscopy. Science 305, 1007–1009 (2004). promise to transform the study of brain circuits in animal models 19. Dodt, H.U. et al. Ultramicroscopy: three-dimensional visualization of neuronal networks in the whole mouse brain. Nat. Methods 4, 331–336 and to decipher the structure-function relationships essential to (2007). the understanding of complex brain functions and their deficits in This study is the first to use LSFM for imaging the entire mouse human brain disorders. brain.

nature methods | VOL.10 NO.6 | JUNE 2013 | 521 review FOCUS ON MAPPING THE BRAIN

20. Niedworok, C.J. et al. Charting monosynaptic connectivity maps by two- 46. Hintiryan, H. et al. Comprehensive connectivity of the mouse main color light-sheet fluorescence microscopy. Cell Rep. 2, 1375–1386 (2012). olfactory bulb: analysis and online digital atlas. Front. Neuroanat. 6, 30 21. Chung, K. et al. Structural and molecular interrogation of intact (2012). biological systems. Nature doi:10.1038/nature12107 (2013). 47. Conte, W.L., Kamishina, H. & Reep, R.L. Multiple neuroanatomical tract- 22. Chung, K. & Deisseroth, K. CLARITY for mapping the nervous system. tracing using fluorescent Alexa Fluor conjugates of cholera toxin subunit Nat. Methods 10, 508–513 (2013). B in rats. Nat. Protoc. 4, 1157–1166 (2009). 23. Leischner, U., Zieglgansberger, W. & Dodt, H.U. Resolution of 48. Sunkin, S.M. et al. Allen Brain Atlas: an integrated spatio-temporal ultramicroscopy and field of view analysis. PLoS ONE 4, e5785 (2009). portal for exploring the central nervous system. Nucleic Acids Res. 41, 24. Mertz, J. & Kim, J. Scanning light-sheet microscopy in the whole mouse D996–D1008 (2013). brain with HiLo background rejection. J. Biomed. Opt. 15, 016027 49. Ugolini, G. Advances in viral transneuronal tracing. J. Neurosci. Methods (2010). 194, 2–20 (2010). 25. Kalchmair, S., Jahrling, N., Becker, K. & Dodt, H.U. Image contrast 50. Ekstrand, M.I., Enquist, L.W. & Pomeranz, L.E. The alpha-herpesviruses: enhancement in confocal ultramicroscopy. Opt. Lett. 35, 79–81 (2010). molecular pathfinders in nervous system circuits. Trends Mol. Med. 14, 26. Keller, P.J. et al. Fast, high-contrast imaging of animal development 134–140 (2008). with scanned light sheet-based structured-illumination microscopy. 51. Song, C.K., Enquist, L.W. & Bartness, T.J. New developments in Nat. Methods 7, 637–642 (2010). tracing neural circuits with herpesviruses. Virus Res. 111, 235–249 27. Wickersham, I.R., Finke, S., Conzelmann, K.K. & Callaway, E.M. Retrograde (2005). neuronal tracing with a deletion-mutant rabies virus. Nat. Methods 4, 52. Callaway, E.M. Transneuronal circuit tracing with neurotropic viruses. 47–49 (2007). Curr. Opin. Neurobiol. 18, 617–623 (2008). 28. Wickersham, I.R. et al. Monosynaptic restriction of transsynaptic tracing 53. Wickersham, I.R. & Feinberg, E.H. New technologies for imaging synaptic from single, genetically targeted neurons. Neuron 53, 639–647 (2007). partners. Curr. Opin. Neurobiol. 22, 121–127 (2012). This study describes a genetically modified rabies virus designed to 54. Rancz, E.A. et al. Transfection via whole-cell recording in vivo: bridging specifically label direct presynaptic input onto a given cell single-cell physiology, genetics and connectomics. Nat. Neurosci. 14, population. 527–532 (2011). 29. Ng, L. et al. An anatomic gene expression atlas of the adult mouse This study was the first to combine intracellular neuronal recording brain. Nat. Neurosci. 12, 356–362 (2009). with DNA delivery. The authors use this method to map the synaptic 30. Jones, E.G., Stone, J.M. & Karten, H.J. High-resolution digital brain function of a single cell in vivo and then target rabies-based atlases: a Hubble telescope for the brain. Ann. NY Acad. Sci. 1225 retrograde labeling of the cells’ synaptic input. (suppl. 1), E147–E159 (2011). 55. Miyamichi, K. et al. Cortical representations of olfactory input by trans- 31. Lein, E.S. et al. Genome-wide atlas of gene expression in the adult synaptic tracing. Nature 472, 191–196 (2011). mouse brain. Nature 445, 168–176 (2007). 56. Marshel, J.H., Mori, T., Nielsen, K.J. & Callaway, E.M. Targeting single This study pioneered large-scale LM-based mouse brain anatomy and neuronal networks for gene expression and cell labeling in vivo. Neuron introduced the Allen Mouse Brain Atlas and online data portal. 67, 562–574 (2010). 32. Dong, H.W. The Allen Reference Atlas: A Digital Color Brain Atlas of the This paper describes an electroporation method for single-cell C57Bl/6J Male Mouse (John Wiley & Sons Inc., 2008). delivery of DNA for targeted infection of modified rabies virus. 33. Lanciego, J.L. & Wouterlood, F.G. A half century of experimental 57. Takatoh, J. et al. New modules are added to vibrissal premotor circuitry neuroanatomical tracing. J. Chem. Neuroanat. 42, 157–183 (2011). with the emergence of exploratory whisking. Neuron 77, 346–360 (2013). 34. Glover, J.C., Petursdottir, G. & Jansen, J.K. Fluorescent dextran-amines 58. Arenkiel, B.R. et al. Activity-induced remodeling of olfactory bulb used as axonal tracers in the nervous system of the chicken embryo. microcircuits revealed by monosynaptic tracing. PLoS ONE 6, e29423 J. Neurosci. Methods 18, 243–254 (1986). (2011). 35. Llewellyn-Smith, I.J., Martin, C.L., Arnolda, L.F. & Minson, J.B. Tracer- 59. Finke, S. & Conzelmann, K.K. Replication strategies of rabies virus. toxins: cholera toxin B-saporin as a model. J. Neurosci. Methods 103, Virus Res. 111, 120–131 (2005). 83–90 (2000). 60. Ugolini, G. Specificity of rabies virus as a transneuronal tracer of motor 36. Grinevich, V., Brecht, M. & Osten, P. Monosynaptic pathway from rat networks: transfer from hypoglossal motoneurons to connected second-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature vibrissa motor cortex to facial motor neurons revealed by lentivirus-based order and higher order central nervous system cell groups. J. Comp. axonal tracing. J. Neurosci. 25, 8250–8258 (2005). Neurol. 356, 457–480 (1995). 37. Atasoy, D., Aponte, Y., Su, H.H. & Sternson, S.M. A FLEX switch targets 61. Federspiel, M.J., Bates, P., Young, J.A., Varmus, H.E. & Hughes, S.H. Channelrhodopsin-2 to multiple cell types for imaging and long-range A system for tissue-specific gene targeting: transgenic mice susceptible npg circuit mapping. J. Neurosci. 28, 7025–7030 (2008). to subgroup A avian leukosis virus-based retroviral vectors. Proc. Natl. 38. Taniguchi, H. et al. A resource of Cre driver lines for genetic targeting of Acad. Sci. USA 91, 11241–11245 (1994). GABAergic neurons in cerebral cortex. Neuron 71, 995–1013 (2011). 62. Young, J.A., Bates, P. & Varmus, H.E. Isolation of a chicken gene that 39. Madisen, L. et al. A robust and high-throughput Cre reporting and confers susceptibility to infection by subgroup A avian leukosis and characterization system for the whole mouse brain. Nat. Neurosci. 13, sarcoma viruses. J. Virol. 67, 1811–1816 (1993). 133–140 (2010). 63. Wall, N.R., Wickersham, I.R., Cetin, A., De La Parra, M. & Callaway, E.M. 40. Madisen, L. et al. A toolbox of Cre-dependent optogenetic transgenic Monosynaptic circuit tracing in vivo through Cre-dependent targeting and mice for light-induced activation and silencing. Nat. Neurosci. 15, complementation of modified rabies virus. Proc. Natl. Acad. Sci. USA 107, 793–802 (2012). 21848–21853 (2010). 41. Harris, J.A., Wook Oh, S. & Zeng, H. Adeno-associated viral vectors for 64. Etessami, R. et al. Spread and pathogenic characteristics of a G-deficient anterograde axonal tracing with fluorescent proteins in nontransgenic and rabies virus recombinant: an in vitro and in vivo study. J. Gen. Virol. 81, cre driver mice. Curr. Protoc. Neurosci. 1.20, 21–18 (2012). 2147–2153 (2000). 42. Thompson, R.H. & Swanson, L.W. Hypothesis-driven structural 65. Lo, L. & Anderson, D.J.A. Cre-dependent, anterograde transsynaptic viral connectivity analysis supports network over hierarchical model of brain tracer for mapping output pathways of genetically marked neurons. architecture. Proc. Natl. Acad. Sci. USA 107, 15235–15239 (2010). Neuron 72, 938–950 (2011). 43. Gerfen, C.R. & Sawchenko, P.E. An anterograde neuroanatomical tracing 66. Beier, K.T. et al. Anterograde or retrograde transsynaptic labeling of CNS method that shows the detailed morphology of neurons, their axons and neurons with vesicular stomatitis virus vectors. Proc. Natl. Acad. Sci. USA terminals: immunohistochemical localization of an axonally transported 108, 15414–15419 (2011). plant lectin, Phaseolus vulgaris leucoagglutinin (PHA-L). Brain Res. 290, 67. Hawrylycz, M. et al. Digital atlasing and standardization in the mouse 219–238 (1984). brain. PLoS Comput. Biol. 7, e1001065 (2011). 44. Naumann, T., Hartig, W. & Frotscher, M. Retrograde tracing with Fluoro- 68. Paxinos, G. & Franklin, K.B. The Mouse Brain in Stereotaxic Coordinates Gold: different methods of tracer detection at the ultrastructural level (Gulf Professional Publishing, 2004). and neurodegenerative changes of back-filled neurons in long-term 69. Swanson, L.W. & Bota, M. Foundational model of structural connectivity studies. J. Neurosci. Methods 103, 11–21 (2000). in the nervous system with a schema for wiring diagrams, connectome, 45. Reiner, A. et al. Pathway tracing using biotinylated dextran amines. and basic plan architecture. Proc. Natl. Acad. Sci. USA 107, 20610–20617 J. Neurosci. Methods 103, 23–37 (2000). (2010).

522 | VOL.10 NO.6 | JUNE 2013 | nature methods FOCUS ON MAPPING THE BRAIN review

70. Moene, I.A., Subramaniam, S., Darin, D., Leergaard, T.B. & Bjaalie, J.G. 89. Akerboom, J. et al. Optimization of a GCaMP calcium indicator for neural Toward a workbench for rodent brain image data: systems architecture activity imaging. J. Neurosci. 32, 13819–13840 (2012). and design. Neuroinformatics 5, 35–58 (2007). 90. Mank, M. et al. A genetically encoded calcium indicator for chronic 71. Klausberger, T. & Somogyi, P. Neuronal diversity and temporal dynamics: in vivo two-photon imaging. Nat. Methods 5, 805–811 (2008). the unity of hippocampal circuit operations. Science 321, 53–57 91. Ko, H. et al. Functional specificity of local synaptic connections in (2008). neocortical networks. Nature 473, 87–91 (2011). 72. O’Rourke, N.A., Weiler, N.C., Micheva, K.D. & Smith, S.J. Deep molecular 92. Holmgren, C., Harkany, T., Svennenfors, B. & Zilberter, Y. Pyramidal cell diversity of mammalian synapses: why it matters and how to measure it. communication within local networks in layer 2/3 of rat neocortex. Nat. Rev. Neurosci. 13, 365–379 (2012). J. Physiol. 551, 139–153 (2003). 73. DeFelipe, J. et al. New insights into the classification and nomenclature of 93. Thomson, A.M., West, D.C., Wang, Y. & Bannister, A.P. Synaptic cortical GABAergic interneurons. Nat. Rev. Neurosci. 14, 202–216 (2013). connections and small circuits involving excitatory and inhibitory neurons 74. Emes, R.D. & Grant, S.G. Evolution of synapse complexity and diversity. in layers 2–5 of adult rat and cat neocortex: triple intracellular Annu. Rev. Neurosci. 35, 111–131 (2012). recordings and biocytin labelling in vitro. Cereb. Cortex 12, 936–953 75. Micheva, K.D., Busse, B., Weiler, N.C., O’Rourke, N. & Smith, S.J. Single- (2002). synapse analysis of a diverse synapse population: proteomic imaging 94. Song, S., Sjostrom, P.J., Reigl, M., Nelson, S. & Chklovskii, D.B. Highly methods and markers. Neuron 68, 639–653 (2010). nonrandom features of synaptic connectivity in local cortical circuits. 76. Micheva, K.D. & Smith, S.J. Array tomography: a new tool for imaging PLoS Biol. 3, e68 (2005). the molecular architecture and ultrastructure of neural circuits. Neuron 95. Harvey, C.D., Collman, F., Dombeck, D.A. & Tank, D.W. Intracellular 55, 25–36 (2007). dynamics of hippocampal place cells during virtual navigation. Nature 77. Tsai, P.S. et al. Correlations of neuronal and microvascular densities in 461, 941–946 (2009). murine cortex revealed by direct counting and colocalization of nuclei This study introduced the method of physiological recording in head- and vessels. J. Neurosci. 29, 14553–14570 (2009). restrained mice on a spherical treadmill while the mice performed 78. Oberlaender, M. et al. Cell type-specific three-dimensional structure of spatial tasks in a virtual environment. thalamocortical circuits in a column of rat vibrissal cortex. Cereb. Cortex 96. Harvey, C.D., Coen, P. & Tank, D.W. Choice-specific sequences in parietal 22, 2375–2391 (2012). cortex during a virtual-navigation decision task. Nature 484, 62–68 (2012). 79. Meyer, H.S. et al. Inhibitory interneurons in a cortical column form hot 97. Huber, D. et al. Multiple dynamic representations in the motor cortex zones of inhibition in layers 2 and 5A. Proc. Natl. Acad. Sci. USA 108, during sensorimotor learning. Nature 484, 473–478 (2012). 16807–16812 (2011). 98. Fenno, L., Yizhar, O. & Deisseroth, K. The development and application of 80. Meyer, H.S. et al. Number and laminar distribution of neurons in a optogenetics. Annu. Rev. Neurosci. 34, 389–412 (2011). thalamocortical projection column of rat vibrissal cortex. Cereb. Cortex 99. Koch, C. & Reid, R.C. Neuroscience: observatories of the mind. Nature 20, 2277–2286 (2010). 483, 397–398 (2012). 81. Ascoli, G.A. et al. Petilla terminology: nomenclature of features of 100. Alivisatos, A.P. et al. Neuroscience. The brain activity map. Science 339, GABAergic interneurons of the cerebral cortex. Nat. Rev. Neurosci. 9, 1284–1285 (2013). 557–568 (2008). 101. Ahrens, M.B., Orger, M.B., Robson, D.N., Li, J.M. & Keller, P.J. 82. Margrie, T.W. et al. Targeted whole-cell recordings in the mammalian Whole-brain functional imaging at cellular resolution using light-sheet brain in vivo. Neuron 39, 911–918 (2003). microscopy. Nat. Methods 10, 413–420 (2013). 83. Kitamura, K., Judkewitz, B., Kano, M., Denk, W. & Hausser, M. Targeted 102. Ahrens, M.B. et al. Brain-wide neuronal dynamics during motor patch-clamp recordings and single-cell electroporation of unlabeled adaptation in zebrafish. Nature 485, 471–477 (2012). neurons in vivo. Nat. Methods 5, 61–67 (2008). 103. Herrera, D.G. & Robertson, H.A. Activation of c-fos in the brain. Prog. 84. Angelo, K. et al. A biophysical signature of network affiliation and Neurobiol. 50, 83–107 (1996). sensory processing in mitral cells. Nature 488, 375–378 (2012). 104. Barth, A.L., Gerkin, R.C. & Dean, K.L. Alteration of neuronal firing 85. Reid, R.C. From functional architecture to functional connectomics. properties after in vivo experience in a FosGFP transgenic mouse. Neuron 75, 209–217 (2012). J. Neurosci. 24, 6466–6475 (2004). 86. Briggman, K.L., Helmstaedter, M. & Denk, W. Wiring specificity in the 105. Grinevich, V. et al. Fluorescent Arc/Arg3.1 indicator mice: a versatile tool

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature direction-selectivity circuit of the retina. Nature 471, 183–188 (2011). to study brain activity changes in vitro and in vivo. J. Neurosci. Methods 87. Bock, D.D. et al. Network anatomy and in vivo physiology of visual 184, 25–36 (2009). cortical neurons. Nature 471, 177–182 (2011). 106. Reijmers, L.G., Perkins, B.L., Matsuo, N. & Mayford, M. Localization of a 88. Wallace, D.J. et al. Single-spike detection in vitro and in vivo with a stable neural correlate of associative memory. Science 317, 1230–1233 npg genetic Ca2+ sensor. Nat. Methods 5, 797–804 (2008). (2007).

nature methods | VOL.10 NO.6 | JUNE 2013 | 523 © 2013 Nature America, Inc. All rights reserved. 524 eiw review | O.0 O6 JUNE 2013 | NO.6 VOL.10 tion func about information include also to grown have the human brain, conceptualizations of the connectome tome. At in studies and ofthe macroscale, particularly ated, each provides unique perspectives on the connec of scales resolution all Although are intimately associ New York, USA. and Randolph Cowen Institute for Pediatric Neuroscience, StudyChild Center, New York University Langone Medical Center, New York, Correspondence Correspondence should be addressed to S.C. ( Orangeburg, New York, USA. Maryland, Maryland, USA. and functional perspectives on brain architecture. of development and inquiry and emphasize the importance of integrating structural measurements of functional and structural connectivity. We highlight emerging areas brain. In this Review, we provide a survey of magnetic resonance imaging–based cataloging of neurophenotypes promise to transform our understanding of the human Annotation of phenotypic associations with variation in the connectome and areas, the structural pathways connecting them and their functional interactions. At macroscopic scales, the human connectome comprises anatomically distinct brain RECEIVED RECEIVED 22 FEBRUARY; ACCEPTED 22 APRIL; 1 and mapping for positioned best are resolution scale neurons 80–120 (commonly referred to as of micro- or mini-columns) columns vertical encompasses mediate resolution is the mesoscale, which, in humans, inter The larger). or centimeter cubic a (commonly tissues cortical encompasses which macroscale, the individual neurons and their synaptic connections, and encompasses which microscale, the are extremes the resolutions.At at varying examined be can that struct tions in the brain, the connectome is a con multiscale connec neural of map complete a as defined Initially entific innovation and reflects an agenda for a new era. neurosci of century a over of advances the embodies First introduced in 2005 (ref. R Cameron Craddock the macroscale Imaging human connectomes at and their functional interactions. and functional their connections anatomical their areas, brain to refer to F F Xavier Castellanos Duke Duke University, Durham, North Carolina, USA. Stan Colcombe Center Center for the Developing Brain, MindChild Institute, New York, New York, USA. At present, methodologies for analysis at macro at analysis for methodologies present, At 3 . . In this Review, we will use the term connectome |

nature methods 5 8 Department Department of Statistical Duke Science, University, Durham, North Carolina, USA. Siemens Medical Solutions USA, Charlestown, Massachusetts, USA. 2 & Michael P Milham 3 2 The The Oxford Centre for Functional MRI of the Brain, University of Oxford, Oxford, UK. , 4 1 , , Adriana Di Martino , 2 , , Saad Jbabdi 1), the term ‘connectome’ publi [email protected] 7 Institute for Data Intensive Engineering and JohnSciences, Hopkins University, Baltimore, sh ed on 3 1 , , 9 2 li , Chao-Gan , YanChao-Gan n e 30 m 1 , 4 2 - - - ­ - - - - , , Clare Kelly . ay ) ) or M.P.M. ( 2013; tively respec connectivity, functional and structural ring for infer used are widely (fMRI) MRI and functional (dMRI) MRI Diffusion-weighted resolution. spatial and safety dominant, availability, widespread is of because (MRI) partly imaging resonance magnetic model organisms and neurosurgical patients. neurosurgical and organisms model to limited are currently studies microscale-resolution only available for analyses at the macroscale; macroscale; the at analyses for available only for tools tional and demands. analytical Moreover, noninvasive lower-resolution macroscale, owing to lower computa- the at feasible most is connectome the of annotation and mapping Comprehensive studies. brain-imaging findings is most amenable to guidance from lesion and tive processes. Interpretation of macroscale-resolution and affec relate to regulatory, cognitive most directly resolution, representations captured at the macroscale behavioral associations. The higher-order, and albeit cognitive lower- with connectomes human annotating portrayals of white-matter tracts and insights into into and orientation their that insights guide principles organizing and tracts white-matter of portrayals Among the modalities used for macroconnectomics, d Focus onmappingthebrain o 4 i . dMRI provides cubic-millimeter-resolution cubic-millimeter-resolution provides dMRI . :10.1038/nm 1 2 [email protected] in vivo in Nathan Kline Institute for Psychiatric Research, , 2 4 9 , These These authors contributed equally to this work. , , Keith Heberlein 4 , 9 , , Joshua T Vogelstein imaging the human connectome are are connectome human the imaging et h.248 6 Institute for Brain Sciences, 2

8 ,

). 4 The The Phyllis Green 1 , 5 – 7 ,

in

vivo - - -

© 2013 Nature America, Inc. All rights reserved. and most enforce specific properties on resulting brain areas brain resulting on properties specific enforce most individuals, and across information of pooling involve commonly areas based on homogeneity of functional time series time functional of homogeneity into on brain the based areas subdivide to used be can techniques clustering by dramatically different functional different dramatically by characterized are each that subregions contains region cingulate ok oe uig pta independence spatial using net define nodes to used work be also can techniques source-separation functional parcellations functional 400-unit and 200- from derived are atlases CC400 and CC200 The segmentations. myelo-architectonic and cyto- postmortem from derived represented as a single area in anatomical atlases anatomical in area single a as represented being For despite example, connectomes. represent to accurately capacity their limiting thus them, construct to used was mation infor connectivity–based or structural no functional roscience, neu to central are atlases resulting single these of Although individuals. morphology) cell example, (for architec measurements postmortem tonic used efforts prefrontal parcellation Early dorsolateral cortex). example, (for areas brain vox larger MRI to individual els in contained from cortex of range patches can small nodes the by represented the subunits brain investigation, the specific of scope the on Depending units. brain constituent the define to complex best how on a agreement is lack we as task connectome macroscale a of nodes the Defining D variables related to phenotypic meaningfully among variations individuals trajectories; fMRI reveals a universal functional architecture, with Figure ucinl r tutrl onciiy profiles connectivity structural or functional connectivity that are represented by edges in the connectome. the in by edges represented are that connectivity functional and structural of patterns quantify and map to used methodologies and analytic the imaging and review then efforts) ‘parcellation’ as to referred (here nodes by represented subunits discrete into brain the to subdividing challenges and approaches review We first nodes. between relationships pairwise represent edges and areas, brain of representations abstract are graph the in Nodes areas. brain among interactions of graph a as nectome con the treats that perspective mathematical a of terms in tion analysis of connectomes. Wemacroscale our structured presenta The EZ (Eickhoff–Zilles) EZ The of the human brain. AAL (automated anatomical labeling) anatomical (automated AAL brain. of human the views top and (right) view a lateral show schemes parcellation rows) two (bottom functional and rows) four (top anatomical using generated areas Oxford (HO) Oxford ity, accuracy, reproducibility or stability of the brain areas brain the of stability or reproducibility homogene on accuracy, ity, based estimated areas be of can number which the generated, be to prespecify to need the is drawback One vary substantially ( substantially conveyed vary details specific the another, one to bear similarity may gross data a parceling of strategies different using captured patterns brain human large-scale the Although patterns. tivity task-based fMRI (T-fMRI) studies (T-fMRI) fMRI Meta- task-based areas. of basis the on brain nodes define to used be can delineate approaches to analytic used be should information Focus onmappingthebrain efining In this Review, we focus on the mapping, characterization and characterization mapping, on the we focus Review, In this Ideally, both brain-function and structural-connectivity structural-connectivity and brain-function both Ideally,

1 | Different parcellations of the human brain. Atlases of brain of brain Atlases brain. of human the parcellations Different 5

(for example, behavioral and psychiatric). and behavioral example, (for nodes 11 0 are derived from anatomical landmarks (sulci and gyral). gyral). and (sulci landmarks anatomical from derived are Fig. Fig. 11 1 1 1 . and TT (Talariach Daemon) (Talariach TT and 1 ). 9 . Alternatively, data-driven data-driven Alternatively, . 7 and structural and 1 3 . These methods methods These . 11 2 atlases are atlases 8 , 10 6 11 , the anterior anterior the , 9 , and Harvard Harvard and 1 2 8 . Blinded Blinded . connec 10 , 1 10 1 , or or , , 1 1 1 1 ------. .

n dnrtc oncin aog neurons among connections axonal of dendritic collection and the encompasses connectivity Structural E single subjects and modalities. and subjects single require to for of fit of quality at cost the likely potentially modalities, and viduals is indi across connectomes information incorporate that strategies parcellation of comparison Optimal describe dMRI acquisition in more detail in in detail more in acquisition dMRI describe to Weused (tractography). matter of white tracts is large-scale reconstruct information this voxel; each through passing fibers of orientation(s) the estimate to necessary information the tain con images these combined, When to hundreds. six from varies acquired directions unique for images of number The direction. of images are acquired, each sensitive to along diffusion a specific matter.voxel of each white a In series dMRI, through passing ers the motion of water, it is to possible map of the orientation(s) fib in By equally all (almost) spinal directions. occurs fluid following cerebral and matter gray in diffusion water In contrast, of axons. path the along primarily occurs and hindered is matter white in diffusion water that is data dMRI from connectivity structural Acquisition. connectivity. structural human of studies for but the noninvasive nature of dMRI makes it the connectivity, remain the for gold standard establishing odologies meth These nonhumans. in tracing invasive or humans in tion dissec blunt and lesions from derived primarily was knowledge noninvasive, with measure to difficult is connectivity structural simplicity, tional stimating

CC400 CC200 TT HO EZ AAL

structural The basic principle underlying the inference of of inference the underlying principle basic The nature methods in in vivo

connectivity imaging approaches. Before dMRI, our dMRI, Before approaches. imaging

| O.0 O6 JUNE 2013 | NO.6 VOL.10 Box Box 1 Dsie defini Despite . de de facto 1 and and review standard Table |

525 1 . ------

© 2013 Nature America, Inc. All rights reserved. 526 review Image resolution. root of the increase number of samples hasless impact onSNR (reduced by square than the head to avoid excessive ghosting. Increasing the proportional decrease inthe SNR;the FOV must remain larger samples along adimension. Decreasing the FOV produces a reducing the field of view (FOV) orincreasing the number of maps of the spatialvariation inthe magnetic field spatial distortions canbemathematically corrected using and image dataproduce reconstruction errors misalignment (for example, head motion) between these scans calibration scans for proper reconstruction, and any and image acquisition rate SNR. Parallel imaging techniquesincrease effective bandwidth bandwidth canminimizespatialdistortions, though costing it particularlysusceptible to spatialdistortions. Increasing is afraction of the bandwidth inthe readout direction, making bandwidth spatially shifts the signalinproportion to the inverse of the frequency resulting from magnetic field imperfections the number of frequencies allocated to avoxel; any shiftin increases in-plane smoothing. Bandwidth corresponds to that result inprolonged spin-spin(T2)relaxation, which readout times and asymmetric bandwidth between directions Spatial distortion. General considerations forEPI on parameter reduce traded-off greater improves for demonstrates These samples number dimensions sented image Signal-to-noise SNR and bipolardiffusion–encoding gradients are applied. an fMRI and collected each acquired afterasingle excitation. EPIvolumes are diffusion MRI(dMRI)data.AnEPIvolume isasetof slices, acquiring functional magnetic resonance imaging (fMRI)and Echo-planar imaging (EPI)isthe most common method for oPTImizng Box 1 |

O.0 O6 JUNE 2013 | NO.6 VOL.10 additional 180°rephasing (spinecho) radiofrequency pulse the example,

diffusion

× × × × × × × × ∝ only

parameters contrast

in temporal spatial

I

than of ( 0

∆ SNR in

terms

involves to optimizations t 118 samples

)

( ∆ ∆ ∆

sequence,

∆

and acquiring required optimize

. The bandwidth inthe phase-encoding direction

c z y x acq z y x x directions how by

distortions). ,

determine

ratio of

∆

signal parallel

y

are

,

parameters

different

2 Image resolution can beincreased by in excitation ∆ The EPIreadout scheme leads to extended 117

. (SNR),

z

SNR each described ),

for

). However, increasing the number of

other representing and in

number

in must

imaging

image most fMRI

for

119 dimension

dMRI Beyond averaging

spatial

parameter imaging |

. These techniquesrequire

modern such and N N N N be

nature methods

or imaging

in quality. of

data.

evaluated factor

diffusion

readout. Table

image distortions f as

spatial time MRI and

aspects

( two available imaging

N Gradient settings

( x applications Image

, P 1 points acquisitions

f N . volumes ) image

However, y signal Equation in (equation ,

(for

N

and

119

terms technology SNR z

signal

), affect echo in

. Alternatively, considerations,

example,

time

in artifacts,

fMRI ( 15 can

N

d

of

dMRI. and

. in (1)

EPI acq

( (

the

(1)) N I

MRI image acquisition

between

0 impact P be

acq dMRI

1 ), data

= f

can in

11

), SNR; repre voxel

2) is to

and 7

.

be (1)(1)

- ceptibility Temporal resolution. to mize prevent M than gradient echo techniques hence have betterspatialspecificity, although less sensitivity sequences, whichare sensitive to T2rather thanT2*and seconds at3tesla).BOLDcanalso bemeasured withspin-echo time equalto the T2*of gray matter(echo time of ~30milli sensitive to the BOLDeffectand isoptimizedusing anecho BOLD contrast. Considerations forfMRI smoothness rienced during sliceacquisition and leads to greater in-plane readout time, whichincreases the amount of T2decay expe samples enable high minimize echo time high lel imaging and using stronger diffusion gradients ( blurring. Increasing bandwidth, partialFourier imaging, paral resonance signaland image SNR,and increases in-plane image T2 decay, isessential to dMRI.T2reduces overall magnetic Diffusion contrast. Considerations fordMRI ing and parallel imaging canresult inmarginal speedgains acquisition time. Increasing bandwidth, partialFourier imag temporal resolution. Echo time fundamentally limits slice- errors and motion-correlated imaging artifacts acquired dataand calibration scans induce reconstruction requires calibration scans can improve acquisition time, resolution and bandwidth but magnetic field inhomogeneity and motion time isused; long readouts increase the interaction between volume motions remain problematic unless along repetition minimize the impact of within-volume motion, butbetween- motion during sliceacquisition of the neighboring slice. Spiral sequences are less sensitive to allowing the signalinasliceto fullyrelax before acquisition time spin-history effects induces Motion sensitivity. artifacts tibility echoes and be contrast, inimizing dropout.

the avoided.

z-shim 125

signal b

values require longer diffusion gradient pulses) can

experiment.

and interleaved slicesminimizespin-history effects by with effects.

slices fMRI in but

and 12

the techniques

dropout 1 118 Specialty

different b is .

increasing signal values (desirable for fiber tracking) inshorter from Slice

; parallel imaging mitigatesthese effects

much Alternatively, phase Spinphase(T2*)contrast isthe most Focus onmappingthebrain

Reducing intersecting positioning

39 118 resulting

Beyond image misalignment, head motion Minimizing echo time, and therefore fluctuations

less

sequences . Lower flipangles dephasing

encoding 12 Numberof slicesacquired determines Tilting . Higher-strength gradient systems

3 resolution,

) sensitive 119

result

. Misalignments between the

echo

spin-echo and from

should 120

126 air-tissue direction (for

characteristics in

owing

positioning . time,

. The 3Dimaging techniques to

differences

the which

example,

BOLD

focus

acquisition EPI which to 124

substantially interfaces

decreases

127 , longer repetition partial contrast avoids

on

. Parallel imaging

imaging spiral-in/out decreases in

119 to areas

magnetic

. reduce susceptibility

voluming of

12

can SNR,

b 0

most

multiple values; . continued slices

increases

mini 119 BOLD

suscep should

vital 12

sus .

to 2 - 118 - and

- - -

- . -

© 2013 Nature America, Inc. All rights reserved. Repetition Repetition time Bandwidth (1/ Flip Flip angle ( N Echo time P T Preprocessing. ( Spatial resolution Parallel imaging ( the the magnetic field, known as eddy currents. These can artifacts be field and the currents induced by rapid switching of gradients with of spatial distortion is the the between interaction static magnetic corrected using estimates of the field variations of the field estimates using corrected mathematically or techniques imaging parallel through reduced be can which distortions, spatial generate field magnetic the in particularly prone to such inhomogeneities. These local variations are one with interfaces) (for another air-tissue example, interface MRI) the in achieved of magnetization extent is, (that ceptibility tor. Areas where materials that differ with respect to magnetic sus are a major contribu inhomogeneities field magnetic particular, In of preprocessing. focus are a central which distortions, spatial distortions. image Correcting field. the in consensus than rather impact complex its and preprocessing of difficulties the reflect ref. data (see preprocessing dMRI regarding exists debate Little performed. be can tography trac and extracted be can information directional before tioned ∆ able arameter acq Focus onmappingthebrain x b simultaneously, strategies. Multiband Emerging acquisitionsequences volume repetition time. The number of diffusion directions, Scan time. condition and shifting the echo train earlierintime Echo shifting also reduces echo time, by relaxing the spinecho time, though require ahardware upgrade for most systems. , , oPTImizng Box 1 values and averages acquired determine total scantime. ∆ y

, , 1 ∆ | z , , Magnetic resonance imaging parameters α N ) x , ,

∆

and N Echo time and number of slicesdetermine dMRI In t ) y P )

After acquisition, dMRI data must be precondi be must data dMRI acquisition, After f multiband )

multi-echo

substantially (sampling period) The time between acquisitions of adjacent fMRI volumes The range of frequencies mapped to a voxel rate-limited rate-limited by echo time required for BOLD contrast k repetition time and reduce spatial distortions by sampling Methods such as GRAPPA and SENSE can decrease subsequently subsequently averaged Number of acquisitions that are acquired and of Time between slice excitation and acquiring the center excitation pulse The amount of rotation applied to proton spins by the sampled sampled in a slice determined by field of view and the number of points The volume of tissue sampled in a given voxel; -space -space lines in parallel; decreases in repetition time are k space

imaging, 1

imaging 4 for this an may though exception),

In vivo improving

multiple f

are MRI and dMRI data bydMRI are plagued

emerging

Definition temporal

slices 1 5 . Another cause . Another

acquisition are

resolution, d 118

acquired MRI image acquisition (ctinued) .

- - - -

levels if the data areif the levels submitted to standard reconstruction noise increase can techniques imaging parallel Such combined. subsequently are that channels multiple through parallel in data acquire that antennas with equipped often are scanners modern correct for eddy-current distortions and subject head motion head subject and distortions eddy-current for correct to preprocessing in used commonly is scans, brain of alignment tant details are often overlooked (see ref. ref. (see overlooked often are details tant visual inspection, constitute standard preprocessing. A few impor Overlooked issues. tant issue tant struction method has recently been proposed to address this impor data and generate false positive connections. An alternative recon the overfit to approaches tractography of tendency the increases can artifactually lower diffusivity estimates along the axons, which reduced reduced using bipolar gradients acquisition are emerging as the preferred option preferred the as emerging are acquisition image during currents eddy of effects the for that account approaches explicitly Model-based gradients. diffusion strong using very acquired images for ineffective is method this However, practically and are (heartbeat, echo each In tially directions ment at or

the increases multi-echo

Impacts Impacts the signal available for imaging ( (T2*) on image Determines the impact of spin-spin relaxation (T2) and spin dephasing of slices; longer repetition time durations reduce motion sensitivity direction (for example, ‘scalloping’) shimming or andsusceptibility distortions in the phase-encoding Lower bandwidth settings can increase artifacts owing to inadequate or or pulsatile effects increase temporal noise resulting from head motion, respiration 1/parallelization factor, holding all other factors constant; will Can allow faster image acquisition but at a reduction of SNR by for dMRI Improves SNR; does not make sense for fMRI but is commonly used contrast of images. sensitivity and in-flow effects and improve spin-lattice relaxation (T1) Impacts Impacts given given repetition time will reduce factors constant has only ~30% of the SNR of a 3-mm isovoxel, holding all other higher spatial resolution have lower SNR, for example, a 2-mm isovoxel Signal-to-noise ratio (SNR) is substantially impacted by voxel volume; measured removed

acquisition time enables improving

cost 1

9 optimized per

no I

. Second, any correction of distortion must account account must distortion of correction any Second, . of

respiration, 0 in

. . Flip angles larger or smaller that the Ernst angle for a

from

in higher SNR SNR fMRI, b

spatial -value

the estimation (one nature methods

the and cost. The above preprocessing steps, combined with

two

first fMRI for

with

acquisitions resolution second, spatial

head

BOLD echoes echo

sampling

short of

motion

smoothness. imaging (with macroscale enabling Impact 1

of I 6

0 . . Image coregistration, or spatial echo per . . Low flip angles may reduce motion |

O.0 O6 JUNE2013 | NO.6 VOL.10 a in

rates

slice the

unit and dMRI

13 time I

signal 1 0 short

), ), impacted by the number ).

and scanner are 1 connectomes time

12

4 This Systematic and for examples). First, examples). for 9

acquired ,

more

12 echo denoising substan

one improve 8 1

, 8 instability)

though .

diffusion

review time), with

at - 13 noise -

0

with an . 1

9

. This . This |

527 1 7 - - -

. © 2013 Nature America, Inc. All rights reserved. etheses.nottingham.ac.uk/1164 ( Sotiropoulos of S. courtesy Images of tract. the presence the about confidence indicates bar Color image). anisotropy fractional on superimposed (results approaches (right) probabilistic and (left) deterministic on based principal diffusion direction. diffusion principal the with peak-aligned or function delta a using approximated is fODF the DTI, of case the in speaking, Formally voxel. a within orientation of fiber estimate best as the is taken direction, fusion dif principal the as to referred diffusion, maximal of direction The (eigenvalues). direction each along occurs that diffusion of and minimum water as motion (eigenvectors) well as the amount is the first step in estimating structural connectivity. structural estimating in step first the is voxel each (fODF)—at function density orientation fiber the as therefore be applied to the concurrent orientationgradient must and volume each to applied is that component rotation a ages. Finally, image coregistrations contain is largely ignored by current software pack between different gradient orientations, but magnitude and direction in vary tortions dis where dMRI, for important ticularly par is This coregistration. image during expansion or compression voxel-wise for Figure 528 to that in in that to relative 4× magnified is yellow. Image in is streamline image. Sample anisotropy fractional a on superimposed callosum of corpus region a from lines) (red estimates orientation fiber ( red, medial-lateral. and posterior; anterior- green, superior-inferior; Blue, front. the from viewed brain of human the orientations tensor of map principal based ( brain. a human for maps connectivity Estimating fiber orientation. fiber Estimating coverage a within voxel, heterogeneity require they greater though angular tion. More complex approximationsfunc of delta fODF can simple account a for by such for accounted not heterogeneity to lead and voxel a single within happen can of which one another)—all to adjacent run (temporarily kiss and merge dis cross, to (fan), known perse are Fibers case. the always not is this However, voxel. a within aligned homogeneously are axons when tations ( directions the regarding information yields matrix this of decomposition Mathematical voxel. given a representation of ellipsoid abstract profile the for water-diffusion a uses 3 (DTI) × imaging 3 to tensor an provide matrix Diffusion fusion profile that provides a simple approximation to the fODF relative proportions relative their and present orientations fiber different the captures which voxel, for each function a probability is to infer analysis of dMRI goal the one, just each not As fibers, axonal of voxel. thousands contains each voxel at orientation fiber of estimation the ena bles direction) and rate as (such properties water-diffusion to signal dMRI the of Sensitivity individually. voxels for white-matter inferred be must orientation(s) fiber tractography, via dles review The diffusion tensor provides a good estimate of fiber orien fiber of estimate good a provides tensor diffusion The The diffusion tensor The is diffusion a but simplistic viable model for the dif | O.0 O6 JUNE 2013 | NO.6 VOL.10

2 | Diffusion imaging of structural of structural imaging Diffusion a 23 . . ( , 2 c 4 ) Pyramidal tract streamlines streamlines tract ) Pyramidal (that is, more directions) and models that either either that models and directions) more is, (that 2 1 . Estimation of this function— referred to to referred function— this of Estimation . / ). Before delineating tracts and tracts bun Before delineating | b

http:// ) ) DTI nature methods a ) ) dMRI- x , , - - - y and and a c z ) of maximum maximum of ) 2 0 . 2 ------. assumption that diffusion is least hindered along axons. along hindered least is diffusion that assumption our given bundles, axon of trajectories average the of estimates depict they axons; do not actual represent streamlines individual ( bundles into grouped curves thin of renderings 3D as visualized cal pathways. In practice, however, inference of point-to-point point-to-point of inference however, anatomi of practice, In points pathways. end cal the and trajectories the both estimate Wecan areas. gray-matter between connections all measure to streamline. any given for probabilities of estimation for allowing orientations, fiber local of estimates in uncertainty for account approaches tractography probabilistic deterministic, are traditional approaches Whereas present. are fibers hetero when geneous crucial is which voxel, same the through pass to sion This direction. allows streamlines with orientations differing diffu principal single a than rather voxel each at available tions orienta peak multiple with though principle, same the follows streamlining at line voxel. For each more fODF models, complex for to stream tangent the the a it candidate provides specifically, streamline; the of formation the guides voxel each at direction approximations available. With fODF modeling, diffusion-tensor the principal diffusion the of complexity the on depending ies Estimating edges. Estimating us to trace major white-matter bundles white-matter major trace to us fODF, the allowing along streamlines of construction the guides information orientation fiber trace Local paths. white-matter to putative used are ‘streamlines’, as to referred trajectories, Three-dimensional (3D) nodes. connectome between connectivity structural establish to used are approaches tractography tions, entation and the diffusion signal (see ref. or ori explicitly fiber account implicitly for between interactions Complex fODF models better estimate fiber trajectories, trajectories, fiber DTI to invisible pathways nondominant of recovery estimate better and allow intersect tracts white-matter several when particularly models fODF Complex Fig. Fig. The specific process by which streamlines are developed var developed are streamlines which by process specific The Using streamlining methodologies, it is theoretically possible possible theoretically is it methodologies, streamlining Using 2 ), reminiscent of postmortem dissection photographs. The ), of dissection reminiscent postmortem 2 cm Focus onmappingthebrain After estimation of voxel-wise fiber orienta fiber voxel-wise of estimation After b 2 2 1 6 for example methods). . Results are typically typically are Results . 2 5 . High Low ------

© 2013 Nature America, Inc. All rights reserved. such as partial volume (for example, voxels containing a mix dispersion a and axonal matter) and gray matter of ture white containing voxels example, (for volume partial as such factors confounding to sensitive extremely are but complexities, and diffusivity, serve as common proxies for these microstructural anisotropy such measures, as Related of fractional connectomics. nepeain n considerations. and Interpretation connectivity using streamlining is imprecise and error-prone and imprecise is streamlining using connectivity complexities. Future approaches may benefit from semiglobal semiglobal from benefit may approaches Future complexities. from a single voxel cannot be used to unambiguously resolve these dispersion and crossing) happen in the same voxel. Diffusion data bending, example, (for configurations these all where situations complex more even imagine easily can One dispersion. fiber of pattern that may diffusion from not that be distinguishable easily a create will voxel a in For bend that axons of profile. collection a diffusion instance, the from recovered easily be always not may and crossing simple a than complex more be can axons of tract for models tiple directions in a voxel (crossing fibers) are replacing tensor tensor replacing are fibers) (crossing voxel a in directions tiple not align do necessarily which axons, of thousands of hundreds contains voxel white-matter A modeling. inaccurate of because microstructure regarding inferences erroneous make still physiologi can we or yet noise, cal instrument-based no is there case, ideal the In of from measurements water infer axonal organization diffusion. for other types of statistical analyses statistical of types other for and groups use compare between to quantify, threshold, difficult are connections structural As tractography-based a tance. result, dis with decrease inevitably probabilities connection spatially, uncertainty propagating by operates streamlining Because tract. that of is length the with increases streamlining path streamline’s the in to uncertainty specific issue Another effects. volume partial and ratio signal-to-noise as such factors nonrelevant by affected also is larger Uncertainty cross uncertainty. greater have that bundles), those example, (for pathways nondominant locally instance, For break. easily can however, approximation, This trajectories. their in uncertainty lower therefore and data tions are expected to have a more discernible trace in the diffusion is Strong used to strength. jectories, connection quantify connec tra streamline of uncertainty the in estimate an provides which tractography, probabilistic Often, properties. above of any of the relevant. more are determine potentially though noninvasively, to harder much are efficacy synaptic and terminals axon at synapses of number densities, spine densities, dendrite to as such factors caution anatomical Other with strength. used connection quantify be should measures these Accordingly, ments are emerging experi dMRI complex more on based features microstructural transfer. and of information Measures hence of potentials, action propagation the on consequences important have myelination, and length size, density, as such axons, these of features axons; up are of made connections Anatomical as well. strengths (edge) connection estimate but nodes, between connections of absence connectomes. structural accurate more yield to needed are modeling and quality data both in improvements precision (reproducibility) and precision (correctness) accuracy categories: two into divided be can Focus onmappingthebrain Unfortunately, tractography does not result in quantification quantification in result not does tractography Unfortunately, or existence the infer to able be only not should we Ideally, 3 o 0 graphy. However, the subvoxel organization organization subvoxel the However, graphy. . The fODF models, which account for mul for account which models, fODF The . 2 8 , , and may become an important component 1 4 . Accuracy refers to our ability to to ability our to refers Accuracy . 27 , Pitfalls of tractography tractography of Pitfalls 2 9 . 27 , 2 2 7 9 ------. ;

Uncertainty in voxelwise fiber orientation can be quantified process. entire the in uncertainty the estimating by errors these to quantify try algorithms tractography ment error. Probabilistic measure increase continuous) are tracts (when voxels of nature of in variations anatomy) (despite local and streamlines the discrete stream generation the in size generated step fixed the a of use in Additionally, lines. variations spurious inducing water- tion, estima fODF inadequate and tensor compromise and can modeling diffusion physiological) or instrument-based adjacent features. multiple subvoxel infer to voxels across data diffusion aggregate that models different aspects of the functional architecture functional the of aspects different probe analyses (for these rest), versus task example, comparisons tories into spatial histograms of their locations ( locations their of histograms spatial into tories trajec of streamline estimates 3D point turns process This lines. stream of location the regarding uncertainties into propagated remote brain areas brain remote spatially between events neurophysiological of synchronization the as it defining connectivity, functional of perspective logical to challenging more be define. can The macroconnectomics field has adopted a neurophysio counterpart functional its brain areas, between connections physical of presence the given tive Although the concept of connectivity is structural relatively intui - E or stimuli, referred to (iFC) as connectivity intrinsic functional spontaneous synchronized activity of occurring in the absence of using experimenter-controlled tasks detection Approaches the events. on of focus types R-fMRI specific to response in or performance task of period entire the across quantified be can and aspects of iFC obtained during one state may not necessarily one state may during not of necessarily iFC obtained and aspects obtained using one task will not to generalize another,necessarily when eFC is assessed using meta-analytic techniques meta-analytic using assessed is eFC when iFC and can eFC patterns be Although similar, notably especially nectivity (eFC) or nectivity coactivation con functional evoked extrinsic as here to to referred tasks, or responses stimulation synchronous of detection the on focus task-free or ‘resting state’ fMRI (R-fMRI)). Task-based approaches versus T-fMRI is, (that task a of absence or presence the of basis connectivity. functional inferring for technique used widely most the is fMRI BOLD-based analyses), metabolism flow and blood glucose cerebral (BOLD), dependent level– oxygenation blood example, (for function physiological to related indices different and magnetoencephalography) and fMRI tomography, emission positron example, (for modalities neuroimaging of variety a using noninvasively measured be can (ref. 1993 in fMRI and tomography emission positron for adopted were connec analyses tivity functional studies, recording multiunit and lography eled through diffusion and biases in cortical projections. cortical in biases and diffusion through eled result ties from a mod lack of architecture in white-matter detail identifying where they project at into gray matter project they where (yet) identifying good not but matter white deep in bundles of location the estimating at good very are which algorithms, tractography for a challenge remains This cortex. throughout terminate tracts where of knowledge requires entirety their in fibers of fication stimating With regard to precision, measurement noise (for example, example, (for noise measurement precision, to regard With eod ocrs eadn acrc ad rcso, identi precision, and accuracy regarding concerns Beyond Studies of functional connectivity may on be dichotomized connectivity the Studies of functional

functional nature methods 3 2 . First quantified in early electroencepha early in quantified First .

connectivity 3 2 ). Although functional connectivity connectivity functional Although ). 3 3 . Evoked functional connectivity . connectivity Evoked functional | O.0 O6 JUNE2013 | NO.6 VOL.10 2 7 3 Fig. 2 Fig. . These difficul . These 3 . eFC patterns patterns eFC . review 3 c 5 or broad broad or ). 3 | 1

and 529 3 4 - - - - - ­ ------.

© 2013 Nature America, Inc. All rights reserved. sity neces its question some Although slices. different at obtained series in time shifts effective creating times, at different acquired correction. timing Slice ( preferred is sources noise of impact the minimize to strategies acquisition of However,optimization The preprocessing steps described below are post- benchmarks. objective of lack a to owing part in elusive, remain analyses iFC and for eFC implications their and strategies essing motion and respiration). Comprehensive comparison of preproc ables can be contaminated by the same noise signals (for example, spurious findings given that the independent and dependent vari task-activation of the risk ofapproaches because greater with functional-connectivity for originated greater are implications and use their However, steps approaches. preprocessing Most set of preprocessing steps, their ordering and their implementation. for. Despite unaccounted considerable left effort,if we measurements lack consensus connectivity regarding thefunctional optimal Structured nuisance signals and anatomical variation can obscure ing from variation data and across comparison subjects. facilitate Preprocessing. in quality R-fMRI). We acquisition of imaging BOLD determinants discuss awake; while (typically scanner the in quietly rests or (T-fMRI) task experimental an performs either participant a as minutes, 5–30 in obtained typically are fMRI using analysis connectivity an measure indirect of Data neural activation. sets for functional- Acquisition. generalize to another (for example, wakefulness and sleep; see see Box sleep; and wakefulness example, (for another to generalize 530 denoising and time-series extraction from brain areas. brain from extraction time-series and denoising signal on lags these of impact deleterious for potential the avoid moglobin, which is diamagnetic is which moglobin, oxyhe to contrast in signal), is, resonance (that magnetic the dephases paramagnetic is which deoxyhemoglobin, of trations concen relative to sensitive are that sequences imaging ultrafast using measured is BOLD technique). flow–based blood cerebral ref. (see connectivity functional of studies in review connectivity) during sleep, suggesting decreased integration tivity among network components (thatis, within-network lateral networks each exhibit decreased functional connec reveals state-related iFCchanges. For example, the default and those observed during wakefulness, though direct comparison iFC patterns detected inthese statesare grossly similarto vegetative syndrome and minimallyconscious state and hypnosis) and pathological states(for example, coma, (for example, sleep), induced states(for example, anesthesia attention, withmany studies examining physiological states strated inthe literature. Consciousness hasreceived particular states The impact of cognitive Box 2 3

8 | 2 O.0 O6 JUNE2013 | NO.6 VOL.10 , correction by temporal interpolation is recommended to to recommended is interpolation temporal by correction , for a discussion of states other than wakeful rest). wakeful than other states of discussion a for 133 Box Box onfunctional connectivity hasbeenrecently demon BOLD is the predominant fMRI technique used used technique fMRI predominant the is BOLD 1 The aim of preprocessing is to remove confound i . FC andConsciou 97 The slices of an fMRI volume are are volume fMRI an of slices The , physiological |

nature methods 3 7 ; the resulting measurement is is measurement resulting the ; Box Box 132 1 ). and pathological s States 3 6 hoc for alternative alternative for corrections. 133 ).

------

sary sary to account for motion in group-level analyses tion). Regardless of the motion-correction scheme used, it is neces dynamics, spectral analysis and estimation of temporal autocorrela promising analyses that rely on this structure (for example,timepoints alters temporalthe temporal structure of the data, thereby com throughout the brain the throughout prior slice excitations (spin-history effects) magnetizationresidualfromvoluming) andcomposition (partial tualfMRI signal fluctuations resulting from changes inslice tissue registration techniques. Additionally, head motion inducesimage- artifac 3D using for accounted typically volumes between areas Motion correction. ing coregistration taining predictors calculated from motion parametersare estimatedtypically modeled durand removed in a regression framework, con via removalvia logical correction (for example, Corsica example, (for correction logical physio of means another provide techniques source-separation signal do not completely remove motion-related fluctuations in the fMRI Physiological noise correction. noise Physiological improves the specificity of iFC of specificity improves the signal is considered a nonspecific measure of noise, whose removal brain is commonly regressed from the data. In this model, regression. signal theGlobal global white-matter signal provides superior denoising (for example, anatomy-based correlationexample, corrections (ANATICOR) (for denoising superior provides signal white-matter Incorporating series. spatial variation in the noise captured by the respiration and and from cardiac regressed effects, the fMRI time for surrogates as taken are fluid cerebrospinal and matter white in present signals Instead, performed. commonly not is it ideal, sequently remove their impact their remove sequently sity fluctuations in fMRI images inten producing field, in changes induce the magnetic abdomen effects. Respiration and pulse can be recorded to model and sub to model can be recorded and pulse Respiration effects. longer-term create of breathing depth and rate as well as rhythm criticisms attributing iFC to these physiological signals rather rather than signals neural signals physiological these to iFC attributing criticisms early to led had which fluctuations, signal fMRI induce can tion recovery from coma. iFC, such asimproving recognition of consciousness after These studies also suggest potential clinicalapplications of cortical circuitry inconsciousness remains underexplored. sleep during anesthesia, non–rapid eye movement (non-REM) Of note, changes inthalamocortical connectivity are reported comparison of findings across statesmaybeproblematic. ble to examination inwakeful states(for example, toddlers), sleep-based studies maybeuseful inpopulations not amena segregation aswell iFC between these networks and others maysuggest decreased of information. Complementary findings of decreased negative 41– and 4 3 . To address this issue, the ‘scrubbing’ of offending volumes vegetative 4 1 or spikeregressionor Focus onmappingthebrain 4 0 . Although effective, modeling-based approaches 4 134 Head motion results in a misalignment of brain 6 . . The cardiac cycle generates pulsatile motion states 4 . These results suggest thatalthough 7 . Respiratory movement of the chest and and chest the of movement Respiratory . The mean time series across the whole whole acrossthe series time mean The 135 , 5 but 4 0 7 4 , decreases motion effects motion decreases , 4 . Additionally, changes in cardiac 7 Cardiac pulsation and respira and pulsation Cardiac 4 . Although this is accepted as as accepted is this Although . has been proposed.Excludingbeen has the specific 4 3 9 9 ). . These motion artifacts role 42 , 4 of 5 . 4 thalamo­ 8 ). ). Blinded

4 4 - , and , ------­ - - -

© 2013 Nature America, Inc. All rights reserved. Spatial normalization and smoothing. and normalization Spatial consideration. additional merits of filtering inclusion low-pass ferent perspective ferent dif a providing each measures, dependency of range wide a for between activities in observed brain areas and is an umbrella term dependency statistical a implies simply connectivity Functional connectivity. functional define they which with stringency the in primarily differing relationships, functional define to used be Estimating edges. terizing functional interactions from a different perspective. different a from interactions functional terizing connectivity, with each graph charac from be functional derived differences inter-individual globally synchronous neural signals in gray matter not absolute. Additionally, electrophysiological demonstrations of relativearevalues, regression signal global after obtained ficients this regard, it is important to note that functional correlation coef regularization (shrinkage) methods (shrinkage) regularization using reduced be can which relationships, of statistical estimates is removing valuable signal valuable removing is 0.1 forhertz brain several areas, that than suggesting low-pass filtering greater frequencies at connectivity functional of strations demon recent and series time the for freedom of degrees the in aliasing) example, (for fMRI of low-temporal-resolution the by induced plete removal of physiological noise is unlikely because of artifacts matically precise (directional) description of the interactions interactions the areas brain of between description (directional) precise matically connectivity functional with associated ally low-frequency scanner drift and frequencies above those tradition from the fMRI time series. This frequency range targets removal of remove below 0.001 and hertz frequencies greater than 0.08 connections negative introduces thus zero,atand distribution correlation the centers regression signal global that awareness However, effects. intersite and intersession removes tionships Pearson’swhereas to relationships, rela linear sensitive is primarily correlation nonlinear and linear to sensitive is distribution and function joint-probability the from dependency statistical Temporal filtering. Temporal tion the interpretation of the signal as global simply noise. simulations approaches is that they do not account for information from from information for account not do they that is approaches individuals and increases the signal-to-noise ratio signal-to-noise the increases and individuals across areas brain of correspondence the improves additionally used to increasingly optimize Spatial correspondence. smoothing are templates study-specific and common population- a space; to stereotactic data subject transforming by individuals across variation morphological addresses normalization Spatial jects. sub across comparison for data the conditioning is processing time series time of pairing possible every between coherence) spectral and tion (for example, dependency Pearson’s correlation, mutual statistical informa for tests bivariate from estimated commonly is tivity connec functional Intrinsic area). brain parcellation-defined is, tion of the mean time across series voxels in each brain node (that Focus onmappingthebrain Estimating iFC from R-fMRI data typically begins with extrac iFC from begins R-fMRI Estimating data typically 4 5 6 7 . Concerns about temporal filtering include reductions about . filtering temporal Concerns 5 . Effective connectivity, in contrast, requires a mathe requires in contrast, connectivity, . Effective 5 9 9 . Although these approaches perform well in simple simple in well perform approaches these . Although , , the limited number of results observations in noisy 3 Several mathematical modeling techniques can 2 . For example, mutual information measures measures information mutual . For example, 5 Bandpass filtering is usually performed to to performed usually is filtering Bandpass 8 . This leads to a plurality of to that can graphs a . leads plurality This 5 5 5 2 . Thus, despite historical precedent, precedent, historical despite Thus, . , has made its use controversial. In controversial. use its made has , 6 0 . A limitation of bivariate bivariate of limitation A . Another aspect of pre of aspect Another 34 , 5 4 . However,com . 5 5 3 1 5 call into ques and can alter can and 6 .

hertz hertz - - - - ­ ------

pooled across individuals to optimize estimation parameters estimation optimize to individuals across pooled needn cmoet analysis component independent statistical maps generated from data in the literature in the from data maps generated statistical across coactivation of patterns of detection enabling tasks, often approaches provide a means of eFC across and studies measuring appropriate for identifying nodes of connectome graphs than than exist exceptions graphs although edges, connectome of nodes identifying for appropriate can can be to used a find solution larization techniques (for example, graphical lasso and elastic net) regu cases, these In correlations. partial of specification unique brain areas commonly exceeds the degrees of freedom, preventing of the number approach to this is approaches, preferred bivariate measures akin to mutual information mutual to akin measures using values two the of distribution joint the from compared be using using nonparametric tests of such significance as wavestrapping accounting for interactions with every other area other every with interactions for accounting two brain after areas, between dependency linear the conditional correlation of estimates in Partial results matrix) covariance areas). inverse the to (related other with by (mediated relationships interactions common indirect from direct distinguish to used be cannot they Hence simultaneously. areas brain multiple findings suggests the brain’s intrinsic functional architecture architecture functional intrinsic brain’s the suggests findings iFC and eFC patterns. Examples include self-organizing maps self-organizing include Examples patterns. eFC and iFC correlation or partial correlation partial or correlation task analysis. Regression coefficient series are then compared using statistics, care must be taken to adjust the degrees of freedom freedom of degrees for Alternatively, autocorrelation. temporal this can be addressed the adjust to taken be must care statistics, parametric using When edge. each to significance statistical of test a applying by accomplished be can but straightforward not is selection Threshold not). or present is connection a determine whether to is, (that binarized or thresholded be to it require substantially greater substantially is networks of number actual the suggested have studies though networks, distinct functionally and spatially 8–20 of definitions T-fMRI that these and R-fMRI have studies converged on similar serve (for example, cognitive, affective or visceral). It is impressive domain sub they annotated in terms functionality of the specific set of spatially and distinct functionally networks that can each be T-fMRI and R-fMRI studies is to considerations. fractionate the connectome and into a Interpretations used. be also can methods to regression coefficients) time series principal component analysis, normalized cut clustering cut normalized analysis, component principal Psychophysiological interaction Psychophysiological entire the spans task series time the that assumption the on based are approaches such iFC; examine to used approaches borrowed ( other the not but direction one in dependency obtain can one words, other in try; symme enforce approaches these of all, not but some, Note that r iclrbok bootstrap circular-block or ted ted regression coefficients have Others findings. eFC measured from ‘coactivation’ using fit of specificity greater offering potentially design, stimulus mental actions between patterns of functional connectivity and the experi A A variety of data-driven techniques are also used for identifying Once functional connectivity is estimated, some applications applications some estimated, is connectivity functional Once Many approaches exist for estimating eFC. Several authors have 6 3 r octntd lcs f pcfc ak conditions task specific of blocks concatenated or nature methods Box Box 1 3 . The concordance of T-fMRI and R-fMRI R-fMRI and T-fMRI of concordance The . 3 ). 6 5 or binarized (by applying a threshold 5 7 9 2 . . Additionally, can be information . Sparse covariance estimation estimation covariance Sparse . 3 5 9 9 33 | analyses directly model inter model directly analyses . O.0 O6 JUNE 2013 | NO.6 VOL.10 7 , 6 0 6 5 . These methods are more more are methods These . 6 . Binarized time series can can series time Binarized . generated from a first-level 6 6 A common pursuit of of pursuit common A . Finally, meta-analytical meta-analytical Finally, . review 6 35 1 . Although Although . , 6 7 . 6 9 | and and

531 6 6 6 7 4 2 8 1 ------, . .

© 2013 Nature America, Inc. All rights reserved. external world. As summarized in ref. As world. summarized external the to responses moment-to-moment for framework a provides 532 another. Consistent with this notion, eFC studies have noted noted have studies eFC notion, this with Consistent another. to state or act cognitive one of from another one patterns with their interaction change to appear assemblies neural dis tributed Specifically, interactions. functional many of nature sient tran the asserted long have models neurophysiological graphs, static with represented is often connectivity functional Whereas at ‘rest’.” when ‘active’ even dynamically and continuously is action in brain the by utilized networks functional of repertoire ity ( ity review When considering the visualization of functional connectiv functional of visualization the considering When because ly. and prediction past effective DCM firmatory analysis, exploratory implementations of SEM tions they canmodel efficiently. Primarilyintended for con approaches are limitedinthe number of nodes and interac a large populationof putative models feasibility of successfully identifying a‘bestfit’model from to identify the ‘best-fit’model. Concerns exist regarding the models representing competing hypotheses regarding causality ated individually; however, common practice isto compare to produce the downstream fMRIactivity. Models canbeevalu based onagiven biophysical model and uses aforward model approach, namic (BOLD or both endogenous variable (activity ispredicted by another node) exogenous variable (predicting the activity of another node), covariance-based approach thatrepresents each node asan actions among nodes withmeasured fMRIdata.SEMisa evaluate the fitof hypothesized models of directional inter­ eling (SEM)and dynamiccausal modeling (DCM)approaches directionality of information flow. Structural equationmod not beinterpreted asindicating causality butrather the we note thatdespite the nomenclature, their findings should Before reviewing statisticaleffective connectivity approaches, Statistical techniques view of noninvasive approaches to establishing directionality. not generally applicable inhumans. Here we provide anover directional relationships innonhuman populations ing and stimulation techniquesare powerful tools for mapping conditions (for example, epilepsy control systems aswell asectopic foci leading to pathological tification of neural drivers canfacilitate ourunderstanding of information flow infunctional-connectivity studies of directional relationships isessential for characterizing Commonly referred to aseffective connectivity, the mapping Box 3 Fig. Fig. | Granger

O.0 O6 JUNE 2013 | NO.6 VOL.10 Application

136 DCM, values are emerging for use withR-fMRI. 3 139

of

), an important question is: ‘what are we missing?’. missing?’. we are ‘what is: question important an ),

connectivity

GCA

; effective connectivity ismodeled atthe hemody modeling effective connectivity atthe neuronal level

causality of

(i) of

the response)

Direction

limitations the

can of

current

GCA

time be

analysis

used

approach to

series level.

R-fMRI value imposed

to

(GCA) |

In

assess for nature methods in 14

ality infuction 14 contrast,

2 one

137 4

another

by

that is is

). Although invasive trac many

6 a node the particularly 140 7

model-free, investigates , , it that “the appears full . SEMand DCM

sluggish DCM 14

nodes could 3 .

In

is

contrast a simultaneous improve

controversial and

generative data-driven 138

136 whether

, they are variable . Iden 141

to the

and -

- SEM - - -

------al connect means of explaining observed behavioral variability. More and and More variability. behavioral observed explaining of means a be to eFC prove may changing performance task of of course the over detection patterns the (ii) configurations and iFC populations, various the between in spent time of distribution different a reflect may they as reassessed, be to need would ies stud population in hyper-connectivity or hypo- of findings (i) multifold: be would implications the If true, time. over con ‘mean’ nectivity the capturing only incomplete, are iFC of metrics 5-minute scan a 5-minute over patterns iFC in changes dynamic observed have studies ing at the same regions same the at ing look when even findings, their in task-dependency substantial commonly limitedto scans afterlesion, taken afternatural interhemispheric connectivity. Lesionstudies are, however, before and aftercallosotomy revealed disruptions specific to example are obtained before and afteroccurrence of agiven lesion. For taken to infercausal influences. Inthe ideal paradigm, scans activity of anode afterdisruptionof inputs from another are a controlled manner. From thisperspective, changes inthe be compelling inestablishing causality when carriedoutin Lesion studies directly perturbthe system and therefore can Lesion studies using breath-hold GCA rehabilitate HRF (for hemodynamic is alternatives and approaches tion and studies suppression used evolving. techniques (that tions—are ing corticocortically but demonstration The Brain stimulation studies cannot beconsidered definitive. testing predictions orgenerating hypotheses, findings of such mality, neoplasm and seizure focus). Although useful for occurrence of lesions (for example, stroke, congenital abnor ­

challenging

of 13

not

gold example, are the

characteristics and

to

7 is,

GCA electrical ,

have accounting produce

emerging the 146 visual

in

standard transcranial Transcranial

with

, R-fMRI scans from the brain of a6-year-old child

powerful

neurosurgical with

deliver

GCA scans converse.

in to

revealed

repetition 15

response

that system

targeted 7 TMS,

0 R-fMRI

3

Focus onmappingthebrain signals . temporary for a evoked omics . These findings suggest that commonly used used commonly that suggest findings These .

as and for

broader

weak

are for direct

R-fMRI tools though less

alternating-current

14

establishing causal In

faster

magnetic

regional remains constant

function 9

from time cortical

33 . potentials—which

currents this expensive

14

patients). Transcranial

stimulation ,

3 range

7 include and 9

,

delivering sampling . Perhaps most exciting, recent iFC iFC recent exciting, most Perhaps .

relationships stimulation ≥ though regard,

2

problematic. variation

reversible

stimulation

across areas. (HRF), s) of

to directional 5

HRF

9 and applications

Noninvasive

targeted

and intraoperative

with

of

rates

direct-current Concurrent nodes

focal

(ii) technically deconvolution

node

in (iii) sites

stimulation

neural in limited

14

slow involve the

(TMS)

5 stimulation 5

motor 9

. brain assumptions influences

A

. to

Nonetheless,

HRF

Attempts

impacts

sampling are

stimulation excitation other

applicability TMS-fMRI

can

areas,

the

demanding circuits studies slowly using

stimula

before

be loca track

node is

to

rate

that the

14

- of or

- - 8

B

- - - © 2013 Nature America, Inc. All rights reserved. from ref. ref. from permission with (reproduced hemispheres to corresponds halves right or color, by left and differentiated are lobes lines; with connections and layers separate in metrics connectivity differing band, circular the in columns as ( ref. from reproduced (image connectome of macaque the atlas of CoCoMac the from ref. ref. from reproduced (image 25) (BA cortex cingulate subcallosal scACC25, cortex; cingulate posterior ventral vPCC, 24); (BA cortex midcingulate MCC24, 10); (BA cortex prefrontal medial dorsal dMF10, 11); (BA cortex orbitofrontal OFC11, 9); (BA cortex prefrontal dorsolateral dlPFC9, 24); (BA cortex cingulate anterior right rACC24, insula; anterior aINS, 25); and 24 (BA cingulate subgenual sgCg24|25, 10); (BA) area (Brodmann cortex prefrontal medial ventral vMF10, thalamus; dorsomedial dmThalamus, depression. in status of diagnostic predictive connections brain of functional representation ref. from reproduced (image subdivisions cortex posteromedial for of iFC short-term iFC alterations reflect experience-induced plasticity, plasticity, experience-induced reflect alterations iFC short-term If hypothesis. this supports recall) example, (for behavior quent of iFC and subse modulations task-related between correlations paring iFC during R-fMRI scans collected before and after after and before collected performance scans task R-fMRI during iFC paring com By acquisition. data R-fMRI before demands cognitive by impacted systematically be can iFC that suggested have studies findings. for clarity greater providing and time over boundary the in changes mapping by efforts, such inform may dynamics temporal of Examination between functional areas in the brain the in areas functional between zones transition examining of value tial including median nerve stimulation nerve median including the to time in protocols by iFC is modulated direct-stimulation measurements, close performed tasks to related plasticity iting exhib to addition In stimuli. the for memory subsequent with correlated modulated was iFC which to extent the participants, more studies are highlighting the poten the highlighting are studies more Figure tion category of stimuli viewed before the R-fMRI scan R-fMRI the before viewed stimuli of category the on depending varied areas the visual and gyrus between frontal inferior connectivity functional For R-fMRI–based manner. example, task-dependent a in altered is networks between ( magnetic stimulation magnetic memory consolidation memory and learning in role a has activity intrinsic correlated corollary: a suggests also this However, areas. among coactivation his of a tory reflects iFC that hypothesis the for support strong vides of their efficacy (for example, in the context of the treatment treatment the of depression context of the in example, (for efficacy their of as as assessment well protocols for stimulation of targets fication d a Focus onmappingthebrain ) Classical anatomical-tracing–style depiction depiction anatomical-tracing–style ) Classical ) Connectogram depicts brain areas (nodes) (nodes) areas brain depicts ) Connectogram In addition to naturally occurring variations in iFC over time, time, over in iFC variations occurring to naturally In addition This iFC-based evidence of experience-induced plasticity pro plasticity of experience-induced evidence iFC-based This 7 9

. This suggests that R-fMRI may have utility in the identi the in utility have may R-fMRI that suggests . This 3 | 11 9 Visualizing the connectome. connectome. the Visualizing 0 5 ). ( ). ). c 8 ) Flatmap-based representation representation ) Flatmap-based 0 ). 7 4 , studies show that iFC strength within and and within strength iFC that show studies , 7 8 8 and transcranial direct current stimula current direct transcranial and 1 11 . The demonstration of brain-behavior brain-behavior of demonstration The . 3 ). ( ). b ) Matrix ) Matrix

11 7 4 6 , heat pain , heat ). ).

5 - . c a 7 7 V1 , transcranial , transcranial V2 7 5 . Across Across . PHC MACV TCV MACD ------TCI

TCC PCIP CCR PCM PCI PCS A2 or training. Several studies suggest this is the case the is this suggest studies Several training. or then enduring changes would be after expected extended practice structural connections can be thought of as pathways along along pathways as of thought be can connections structural within an individual across time or treatments or or time across individuals of individual an within population a that across analysis connectivity statistical compares dimensional or categorical a forming variables cogni phenotypic other or higher-order diagnoses to neuropsychiatric relevance processes, tive their of terms in them tate is goal to next the anno are graphs estimated, connectome Once S e negtd n ae tity nonnegative strictly are and unweighted to tend be graphs connectivity Structural negative. or positive be can and weighted typically are connections functional example, For connectivity. functional and structural between differences idiosyncratic the for account always must results the of interpre tation the However, sets. data structural or functional with of were constructed they of whether analysis regardless graphs connectome the for appropriate are approaches statistical same who received no treatment no received who children versus interventions reading by remediated were who dyslexia with children between observed were in iFC differences study, retrospective a in preliminary For example, interventions. treatment of efficacy the index to even or function of recovery our understanding of mechanisms involved in remediation-based long-term training–induced plasticity have the potential to inform TCS CPO A1 CCP S1 tatistical S2 IP IA M1 MCV PMCM MCD FCV FEF CCA 5 FCOR8 . These associations are most often inferred by per by inferred often most are associations These . PFCD

analysis FCD FCC PFCM CCS FCPCL nature methods Hippocampus scACC25

of d b Amygdala

the 8 vPCC

MCC24 3 connectome . dMF10 Nucleus accumbens OFC11 | dlPFC9 O.0 O6 JUNE2013 | NO.6 VOL.10 rACC24 Hypothalamus aINS sgCg24|25 dmThalamus vMF10 2 7 8 . Additionally, Additionally, .

4 . Many of the the of Many . review 8 2 . Studies of of Studies .

–6.30 –5.91 –5.51 –5.12 –4.72 –4.33 –3.94 –3.54 –3.15 –2.76 –2.36 –1.97 –1.58 –1.18 –0.79 –0.39 0.00 0.39 0.79 1.18 1.58 1.97 2.36 2.76 3.15 3.54 3.94 4.33 4.73 5.12 5.51 5.91 6.30 |

533 Connection strength (z) strength Connection - - - - © 2013 Nature America, Inc. All rights reserved. their role in connecting other pairs of nodes (betweenness nodes of pairs other connecting in role their Node and graph-level statistics: invariants. statistics: graph-level and Node when comparing graphs comparing when information topological preserve that measures graph-distance using performed be also can they structure, graph ignore which representations, bag-of-edges to applied be to tend techniques power, while maintaining control of false positives. false of control maintaining power, while statistical increase to graphs connectome of structure group the of direct connections (degree centrality (degree connections direct of sure a node’s on the of importance basis the number and strength a graph. Several different centrality metrics are available that mea connections functional but flow, can information which 534 connections that must be traversed to connect any two nodes captures numberof the efficiency interconnected,global whereas Local efficiency assesses the extent to which neighbors are densely efficiency. global and local their of terms in assessed commonly are graphs particular, In connectivity. functional and structural multiple-comparison correction multiple-comparison if desired, requires a return to tests, edge-specific and the need for this information, the involvement Extracting of edges. individual about information obscure they relationships, tome-phenotype edges may result in overly liberal or conservative corrections conservative or liberal overly in result may edges discovery rate false discovery as such techniques correction Standard positives. false of number the for control adequately to comparisons multiple for correction require which tests, the statistical many in perform to need results approach this However, weights. edge and ables vari dimensional or categorical between relationships pretable F neighboring nodes (eigenvector nodes neighboring cannot be interpreted in the same manner same the in interpreted be cannot understanding of connectome-phenotype relationships connectome-phenotype of understanding role in the graph and, when combined, can lead to a more holistic node’s a on perspectives different provide measures various The tic Alternate correction techniques such as the statis network-based edgewise is, (that statistics) them between relationships or interactions account into taking without time, a at one edge each at analyses statistical of and as treat them edges perform a or bag, collection, edges. of bag A h ‘ml-olns’ f graph a of ‘small-worldness’ the random graphs with similar properties cancombined be to assess relationships of these two measures to what would be obtained from are centrality measures that a indicate measures are node’scentrality in relative influence sentations of the graph. The most commonly used node invariants repre particular to unique not are they because parlance theory network- in measures’ ‘topological or parlance graph-theory in ‘invariants’ called are measures These statistics. graph-level and which connections, can using be a described variety of node-level of bivariate strength and presence the beyond architecture brain of sentations contain a connectomes of wealth about information graphs and their associated phenotypic variables with a single single test a statistical with variables phenotypic associated their and graphs connectome entire the between relationship the evaluate niques -tests or regression) allow researchers to identify easily inter easily identify to researchers allow regression) or -tests review Alternatively, multivariate regression and classification tech classification and regression multivariate Alternatively, Similarly, a range of graph-level invariants is used for studyingfor used invariantsis graph-levelSimilarly, of range a 8 8 or group Benjamini-Hochberg | O.0 O6 JUNE 2013 | NO.6 VOL.10 8 4 . Such univariate approaches (for example, example, (for approaches univariate Such . 90 8 , 6 9 The simplest approach to compare graphs is to to is graphs compare to approach simplest The that do not model the dependencies between between dependencies the model not do that 1 . Although powerful for of the connec analysis powerful . Although 9 2 . |

nature methods 9 9 9 4 0 3 or Page Rank . Although these multivariate multivariate these Although . . Small-world graphs balance balance graphs Small-world . 8 9 leverage information leverage about 9 3 ), the importance of of importance the ), 8 5 . Graphical repre Graphical 9 5 centrality) and centrality) 9 5 . t 9 -tests, -tests, 3 . The 8 8 5 7 ). ). ­ ------.

processes and cognition) and processes that are putatively ‘human’ (for example, language, self-referential func tional networks supporting an array of including functions, those homologous for macaques and humans in observed iFC the between correspondence Initial impressive demonstrated has work studies. neuroanatomy functional comparative for tool powerful a as emerging is R-fMRI Simultaneously, dMRI. using ated using histology approaches with longitudinal atlases obtained complement cross-sectional atlases of the developing brain gener Mouse direct comparison of module membership between graphs allows and modules interconnected sparsely but intraconnected quantifies the extent to which a graph can be segregated into densely edges in the graph of number the and efficiency global between difference the from Predictive modeling has primarily focused on invariants and and invariants on bag-of-edges–style focused primarily has modeling Predictive associated phenotypes associated their and graphs of connectivity consisting set training the with arising by chance arising relationships of set a of improbability inferential evaluate to which contrast statistics, in pre individual, to an of pattern phenotype connectivity the a dict of ability the assess directly to P ­integration and segregation to obtain fast and cost-efficient propa ewr iiitv ( initiative Network Research Informatics Biomedical Mouse recent The chemicals). sive (for techniques example, animal and killing of injection toxic inva with associated examinations longitudinal to barriers ing remov rapidly by well, as models are animal in neuroscience research transforming connectomics to approaches MRI-based T nosis connectivity diag identify example, (for variable to phenotypic a of aim predictive patterns frequently researchers Finally, manipulations that are impossible in humans. in impossible are that manipulations genetic and molecular pharmacological, structural, direct with imaging of noninvasive combination the through function brain ani mal aremodels in to poised provide of a understanding mechanistic studies connectomics macroscale tools, imaging-based metric statistical tests are preferred are tests statistical metric nonpara characterized, are poorly of invariants most properties racy via cross-validation via racy single-node failures to robustness as well as graph the through information of gation rnltoa potential this underscores translational rats, as such mammals, lower in iFC of terns nepeain f findings of interpretation confound and groups or individuals between differ sys tematically can that edges) of number example, (for properties graph in differences of potential impact the consider to important it is graphs, between invariants comparing When interpret. to ficult dif more be can relationships resulting the comparisons, tiple of mul number the by decreasing power statistical invariants increase can Although phenotypes. dimensional and categorical with relationships identify to evaluated statistically be can ants ranslational redictive ah f h peiul dsrbd oe n gah invari graph and node described previously the of Each 9 0 / , , age ) provides an initial demonstration of the potential to to potential the of demonstration initial an provides )

modeling 9 1 or brain state brain or

connectomics Focus onmappingthebrain 9 93 8 9 90 . Predictive modeling is typically supervised, supervised, is typically modeling . Predictive . An additional invariant is modularity, which 6 . The cost-efficiency of a graph can be inferred , 10 htt 10 9 0 9 approaches. p://www.loni.ucla.ed 9 2 . One can assess the predictive accu predictive the assess can One . 10 9 . Armed with increasingly powerful powerful increasingly with Armed . or other model-selection techniques. techniques. model-selection or other 9 1 9 3 . Evidence of homologies with pat with homologies of Evidence . 7 . Additionally, as the distribution distribution the as Additionally, . ). Predictive modeling can be used used be can modeling Predictive ). 9 3 . u/BIRN/Projects/ 85 . ------

© 2013 Nature America, Inc. All rights reserved. ac.uk UK UK Biobank Imaging ( Initiative Initiative (INDI) (global) International Neuroimaging Data-sharing ( 1000 Functional Connectomes Project (FCP) United States) consortium consortium ( Connectome Project: MGH-Harvard-UCLA US National Institutes of Health Human conducted under anesthesia—in particular, using the general general the isoflurane anesthetic using particular, anesthesia—in under conducted ( Saint Louis–University of Minnesota consortium Connectome Project: Washington University in US National Institutes of Health Human (European Union) Developing Human Connectome Project regarding these initiatives and others). through the generation and sharing of large-scale imaging data setswithphenotyping orinnovation of data-acquisition and/or analysis techniques(seeref. Large-scale initiatives from around the worldthatare promising toaccelerate the paceof macroscale connectomics research through either the advancement of macroconnectomics research ( Exploration of Brain Connectivity and Tracts Consortium of Neuroimagers for the Noninvasive China) Brainnetome ( Brain Genomics Superstruct (United States) Initiative (weblink whenavailable; location) T gic agonist medetomidine may be preferable as it avoids such such avoids it confounds as preferable be may medetomidine agonist gic adrener alpha-2 sedating The excitability. neural on effects its identical to those developed in humans in developed those to identical rats approaches and analytical monkeys, and have on mice relied preprocessing in studies translational Initial essential. are and n h lu MI environment MRI loud the in restraint to habituated been have that rats awake in examined be can iFC example, For addressed. be also must issues odological R-fMRI techniques in humans also apply to animal studies. animal to apply also humans in techniques R-fMRI and of dMRI interpretation the regarding raised many questions the that note we Finally, strategies. optimal at arrive to explored example, and cardiac (for activity respiration) and must imaging parameters be physiological in differences encouraging, is cess http://fcon_ http://humancon http://w able Focus onmappingthebrain Despite the rich promise of translational connectomics, meth connectomics, of promise translational rich the Despite /

; ; UK) 2 | ww.brain-connect.eu Initiatives Initiatives promising to accelerate macroscale connectomics research http:// 10 1000.projects.nitrc.org http://www.brain 5 Ds-epne tde o aetei ae few are anesthesia of studies Dose-response . nectome.org humanconnectomeproje h ttp://www.ukbiobank. 10 4 , which can confound findings owing to to owing findings confound can which , / / ; ; European Union) ; ; United States) netome.org / ; ; global) 10

3

. However, most studies are are studies most However, .

/ ct.org

;

10 / 3 ; . Although their suc their Although . limited limited to R-fMRI and dMRI), as well as cardiac MRI and rich phenotyping. Extension aims to resample 100,000 of the cohort using multimodal neuroimaging (including but not blood samples and lifestyle information from a cohort of 500,000 subjects, the UK Biobank Imaging Building on an existing long-term prospective epidemiological study that has collected genetics, available through the FCP and INDI efforts combined as well as a growing number of dMRI data sets. Sample a model for prospective, prepublication data sharing (major release: the Nathan Kline Institute-Rockland Brain Imaging Data Exchange (ABIDE; data releases: ADHD-200 Consortium ( the functional neuroimaging community to include phenotypic data beyond age and sex Second(major FCPINDI initiative that was founded in an attempt to (i) expand the scope of open data sharing in community via the Neuroimaging Informatics Tool Resources Clearinghouse ( data sets from 33 independent sites around the world and released them openly to the scientific Grass-roots data-sharing initiative that brought together over 1,200 previously collected R-fMRI resolution, quality and speed of acquisition. of conventional systems). Efforts to optimize dMRI technology will focus on increasing the spatial which is designed to carry out diffusion using ultrahigh gradient strength (4–8 times the strength Initiative focusing on unraveling the full connectivity map using the first ‘Connectome Scanner’, All All data and tools developed through the initiative will be openly shared. dMRI (high spatial resolution), which it has refined and is currently distributing to interested centers. The project uses multiband imaging sequences for R-fMRI (high spatial and temporal resolution) and 300 families) to provide insights into relationships between brain connectivity, behavior and genetics. that makes use of a twin design (1,200 healthy adults, including twin pairs and their siblings from State-of-the-art multimodal imaging initiative (R-fMRI, T-fMRI, dMRI and magnetoencephalography) in in utero Initiative to comprehensively map and model the human connectome for 1,000 babies, including connectivity connectivity atlas. dMRI. Target deliverables include optimized acquisition protocols, analytic tools and a Consortium focused on studying the brain’s microstructure, tracts and connectivity using glioma and 2,000 healthy controls collected from 11 hospitals and imaging centers. Alzheimer’s disease and mild cognitive impairment, 120 patients who had a stroke, 50 patients with behavioral and blood data from more than 1,000 patients with schizophrenia, 300 patients with (electroencephalography, fMRI and dMRI). R-fMRI and diffusion-imaging data sets, along with microscale (microtechnique, ultramicrotomy, staining and visualization techniques) to the macroscale Attempts to characterize brain networks with multimodal neuroimaging techniques, from the containing 1,500 completed, quality-pass data sets is expected to be publicly available in 2013. with comprehensive phenotyping data (cognition, personality and lifestyle), and the resultant repositoryinfluences. The initiative has collected R-fMRI, dMRI and saliva samples from over 3,000 adults, along Aims to collect a large-scale imaging data set to explore brain–behavior relationships and their genetic http://fcon_10 and in in vivo 10 imaging (20–44 weeks after conception). 4 - - -

00.projects.nitrc.or to phenotypic profiles to phenotypic them relate and neurophenotypes catalog to is connectomics of goal A central individual’s neurophenotype. of that specification the to contribute subgraphs its and connectome al’smacroscale individu An investigators. from enthusiasm increasing despite An overarching goal of the connectomics era is the derivation derivation of ‘neurophenotypes’ the is era connectomics the of goal overarching An T cificity of findings will depend on their nature, granularity of of granularity nature, node definitions and their quality of neuroimagingon data used. Similarly,depend will findings of cificity spe the connectomes, macroscale upon based neurophenotypes behavioral, neurological or psychiatric variables. When cataloging affective, cognitive, of combination some of consists typically it though application, the on depending vary can phenotyping of neurophenotypes based on their phenotypic profiles. The breadth of populations differentiating by or connectomes in distinctions driven approaches focused on the detection of commonalities and oward

neurophenotypes http://fcon_100 htt p://con_1000.projects.nitrc.org/indi/abi g/indi/enhanced nature methods Goals 10 4 0.projects.nitrc.org/indi/adhd200 6 . This can be accomplished through data- through accomplished be can . This , a concept that remains poorly specified specified poorly remains that , a concept /

and ). More than 5,000 R-fMRI data sets are

clinical | O.0 O6 JUNE2013 | NO.6 VOL.10

116 for additional information applications http://www.nitrc.or de / )) and (ii) provide /

review ), the Autism

g |

/ 535 ).

- - © 2013 Nature America, Inc. All rights reserved. to rapidly aggregating the necessary data necessary the aggregating rapidly to dedicated initiatives data-sharing large-scale several has supported community macroconnectomics the regard, this In lation. of human the popu representative sets data large-scale attaining consider consider the many complexities could this jeopardize burgeoning and tion, preprocessing data. of Failure analysis brain-imaging to the many that remain broadly surveyed challenges in the acquisi we Review this In methodologies. invasive traditional more by clinical faced progress to and barriers overcoming by research translational neuroscience basic, transforming are macroscale the at connectome the annotating and mapping to approaches MRI-based innovation. methodological cen and a conceptual of than tury more of culmination the is era connectomics The C pro phenotypic of basis the on neurophenotypes sorting when and DTI libraries; (http://www.mendeley neuropsychiatric diagnosis. Calculatedfrom the ChildMind Institute Librarian (Resting State connectivity. Determined bythe number (count) of R-fMRIand dMRIpapersdedicated tothe Neuropsychiatric disorders most commonly studied using macroscale functional and structural Schizophrenia Neuropsychiatric diagnoses T 536 treatment response may prove to be a more fruitful goal than than goal on diagnosis focusing fruitful more a be to prove may response treatment and prognosis risk, disease upon based individuals of stratifying capable of tools attainment the that suggested have recently field the in leaders However, literature. the in espoused increasingly (see indices brain reliable vidual-relevant is the promise of of utility because clinical the ability to obtain indi connectomics surrounding excitement the for reason a major cognition, and behavior for implications its and architecture neurophenotypes. of perspectives dimensional and categorical between balance a find to need will work Future analysis. statistical to available phenotyping the of hensiveness (that is, number and breadth of independent features) by and compre the precision be determined will specificity files, Bipolar Bipolar disorder Amyotrophic lateral sclerosis Obsessive compulsive disorder Coma and vegetative state Anxiety disorders Stroke Sleep disorders Traumatic brain injury Parkinson’s disease Multiple sclerosis Autism spectrum disorders Attention deficit hyperactivity disorder Substance dependence Other neurological disorders Mild cognitive impairment Epilepsy and seizures Depression Alzheimer’s disease ( and dMRI R-fMRI with studied disorders and psychiatric logical Recent years have witnessed an explosion in the number research). of neuro macroconnectomics of pace the accelerating are that Table able onclusion review Beyond Beyond the derivation of a fundamental understanding of brain |

O.0 O6 JUNE 2013 | NO.6 VOL.10 3 3 | ). Hopes of attaining clinically useful diagnostic tools are ). diagnostic Hopes useful of attaining clinically Connectomics studies in clinical populations 10 7 . Regardless, a . key remains: requirement Regardless, .com/profiles/cmi-librarian/). | nature methods R-fMRI count 100 10 10 10 12 15 15 15 17 18 19 22 26 32 40 45 58 62 81 89 Table 8 . 2 for initiatives initiatives for dMRI count 182 259 211 115 179 101 201 283 61 84 27 16 73 86 37 94 50 9 7 ------com/reprints/index.htm R online version of The authors declare competing financial interests: details are available in the COMPETING We acknowledge our colleagues who allowed us to reproduce their figures. as well as Z. Shehzad, Z. Yang and S. Urchs for their helpful comments. D. Lurie for his assistance in the preparation of the manuscript and references by a gift from Joseph P. Healey to the Child Mind Institute (M.P.M.). We thank US National Institutes of Health R01ES017436. Additional support was provided from the London Institute for Mathematical Sciences HDTRA1-11-1-0048 and (R.C.C.) and the Leon Levy Foundation (C.K. and A.D.M.). J.T.V. receives funding Niarchos Foundation (M.P.M.), the Brain and Behavior Research Foundation (BRAINS R01MH094639 to M.P.M. and K23MH087770 to A.D.M.), the Stavros This work was supported by grants from US National Institute of Mental Health A cataloged. variations their and annotated meaningfully be can connectomes which at pace the accelerate optimized preprocessing and can to analytic methodologies serve with combined data, high-quality of acquisition the to attention associated with suboptimal Conversely,methodologies. increased field through the introduction of spurious, irreproducible findings 18. 17. 16. 15. 14. 13. 12. 11. 10. 9. 8. 7. 6. 5. 4. 3. 2. 1. eprints cknowledgments

, 144–153 (2012). 144–153 22, Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. 20th Annual Meeting and Exhibition and Meeting Annual 20th (ISMRM) Medicine in Resonance Magnetic for Society International the of in images. diffusion in movements and distortions correcting (2009). PICA. group order model high cortex. cingulate using resting functional connectivity MRI. connectivity functionalresting using fMRI. (Thieme Classics, 1988). Classics, (Thieme Med. coefficient. diffusion apparent the calculate to used images gradients. bipolar of use the images: echo-planardiffusion-weighted in artifacts imaging. tensor diffusion to application images: echo-planar spin-echo in distortions MRI. diffusion with vivo in clustering. spectral constrained spatially via generated atlas fMRI brain whole A specialization. functional to relation its and cortex cingulate human of parcellation pitfalls. and promise connectome:functional the in variation Characterizing (2001). 229–239 integration.large-scale and synchronization phase brain. human the of description Anderson, J. Anderson, echo-planar of distortion for Correction J.R. Moore, & J.C. Haselgrove, current eddy of Elimination Parker,D.L. & J.S. Tsuruda,Alexander, A.L., susceptibility correct to How Ashburner,J. & S. Skare, J.L.R., Andersson, connectivity brain quantifying of limitations and Challenges D.K. Jones, V.Kiviniemi, A.L. Cohen, H.S. Mayberg, X.P.& Holtzheimer,P.E.,Hu, G.A., James, R.C., Craddock, P.Bellec, N.U. Dosenbach, Connectivity-basedM.F.S. Rushworth, & H. Johansen-Berg, M., Beckmann, D.S. Margulies, M.P.Milham, & X.F.Castellanos, R.C., Craddock,B.B., Biswal, Kelly,C., connectomics.Human O. Sporns, & T.E.J.Behrens, B.B. Biswal, brainweb: The Martinerie,J. & E. J.P.,Rodriguez, Lachaux, F.,Varela, structural A connectome: human The Kötter,R. & G. Tononi,O., Sporns, Talairach, J. & Tournoux,P.& Talairach,J. sets. (2008).

and , 960–964 (1996). 960–964 36, Neuron Neuroimage F

permissions INANCIAL INANCIAL INTERESTS Trends Cogn. Sci. Cogn. Trends Identification of large-scale networks in the brain using brain the in networks large-scale Identification of al. et Magn. Reson. Med. Reson. Magn. the pape Hum. Brain Mapp. Brain Hum. Defining functional areas in individual human brains individual human in areas functional Defining al. et Toward discovery science of human brain function. brain human of science discovery Toward al. et , 799–812 (2006). 799–812 50, A comprehensive Gaussian Process framework for framework Process Gaussian comprehensive A al. et using cortex brain the of segmentationFunctional al. et Focus onmappingthebrain J. Neurosci. J. Mapping the functional connectivity of anterior of connectivity functional the Mapping al. et Neuroimage A core system for the implementation of task of implementation the for system core A al. et Neuroimage , 1231–1243 (2006). 1231–1243 29, l

r . information .

, 181–188 (2012). 181–188 16, , 4734–4739 (2010). 4734–4739 107, , 1175–1190 (2009). 1175–1190 29, Imaging Co-planar Stereotaxic Atlas of the Human Brain Human the of Atlas Stereotaxic Co-planar , 870–888 (2003). 870–888 20,

, 579–588 (2007). 579–588 37, , 1914–1928 (2012). 1914–1928 33, , 1016–1021 (1997). 1016–1021 38, Hum. Brain Mapp. Brain Hum.

PLoS Comput. Biol. Comput. PLoS is

(Melbourne, 2012). (Melbourne, available , 341–355 (2010). 341–355 2, Neuroimage

online

, 3865–3886 30, Nat. Rev. Neurosci. Rev. Nat. Curr. Opin. Neurobiol. Opin. Curr.

at , e42 (2005). e42 1,

http://ww , 45–57 41, Magn. Reson. Magn. Proceedings

w.nature.

2,

© 2013 Nature America, Inc. All rights reserved. 44. 43. 42. 41. 40. 39. 38. 37. 36. 35. 34. 33. 32. 31. 30. 29. 28. 27. 26. 25. 24. 23. 22. 21. 20. 19. Focus onmappingthebrain doi:10.1002/mrm.24623 (7 February 2013). February (7 doi:10.1002/mrm.24623 resonance imaging analyses. imagingresonance magneticfunctional resting-state seed-based, in effects normalization USA Sci. Acad. Natl. Proc. imaging.resonance magnetic with mapping brain functionalstimulation: MRI. perfusion (2002). 577–582 heterogeneity.fiber matter white intravoxel Neuroimage brain. human the in fasciculi matter white of dissection interactive Johansen-Berg, H. and Behrens, T.E.J.) 333–352 (Academic Press, 2009). Press,(Academic 333–352 T.E.J.)Behrens, and H. Johansen-Berg, Neuroanatomy vivo in to Measurement Quantitative From motion on intrinsic functional connectivity MRI. connectivity functionalintrinsic on motion motion. subject from arise networks MRI connectivity functional in correlationssystematic but Spurious (1996). 346–355 time-series. fMRI in effects Movement-related brain. MRI. planar echo- using brain human resting of cortex motor the in connectivity Metab. Flow Blood Cereb. J. sets. data (PET) large of analysis principal-component the connectivity: microscopy.electron for brain mouse Neuroimage MRI. diffusion of don’ts and do’s The fallacies: other and count, Med. Reson. Magn. MRI. diffusion from distribution diameter axon measuring for method Connect. Brain gain? we can What orientations: fibre multiple with tractography diffusion Probabilistic imaging.resonance magnetic spectrum diffusion with architecture tissue complex Mapping 2009). Press,(Academic 55–72 T.E.J.)Behrens, and H. Johansen-Berg, (eds., Neuroanatomy tensor.in data. DTI in motion subject for Satterthwaite, T.D.Satterthwaite, T.D.Satterthwaite, head of influence The Buckner,R.L. & M.R. Sabuncu, K.R.A., Dijk, Van S.E. Petersen, & B.L.Schlaggar, Snyder, A.Z., K.A., Barnes, Power,J.D., Turner,R. & R.S. Frackowiak, R., Howard, S., Williams, K.J., Friston, K.J. Friston, C.W.Wu, S. Ogawa, K.-H. Chuang, human the of map coactivation FunctionalT.Paus, P.T. Fox,& R., Toro, Functional J.S. Hyde, V.M.& Haughton,F.Z., Yetkin, B., Biswal, M.P.Milham, & F.X.Castellanos, S., Colcombe, Kelly,C., M., Mennes, Functional R.S. Frackowiak,P.F.Liddle, & C.D., Frith, K.J., Friston, tractography.in diffusion MR S. Jbabdi, & T.Behrens, whole the embedding and W.Staining Denk, & J. Binding, S., Mikula, integrity,fiber matter White Turner,R. & T.R.Knosche, D.K., Jones, a P.J.AxCaliber: Basser, & Y.Blumenfeld-Katzir,Yovel,Y.,T., Assaf, here? from go we do Tractography:where H. Johansen-Berg, & S. Jbabdi, Virtual D.K. Jones, & S. Pajevic, R.J., Howard, M., Catani, M.W.Woolrich, & M.F. Rushworth, S., Jbabdi, H.J., Berg,T.E., Behrens, R.M. Weisskoff, W.Y.,& T.G.P.,Tseng,Reese,V.J.,Hagmann, Wedeen, D.S. Tuch, effective the of Estimation D. LeBihan, & J. P.J.,Mattiello,Basser, diffusion the beyond fibers: Multiple Alexander,D. & K. Seunarine, correcting when rotated be must B-matrix The D.K. Jones, & A. Leemans, S.N. Sotiropoulos, are not equivalent. not are brain human the of architecturesfunctional intrinsic and extrinsic The 247–254 (1994). 247–254 echo. spin NMR the from tensor self-diffusion (2013). data. connectivity functionalresting-state of preprocessing the in artifact motion of control for filtering and youth. neurodevelopmentin of studies for Relevance connectivity: functional of measures neuroimaging. floor using SENSE. using floor noise thereducing MRI: diffusion multichannel from mappingorientation (2012). Cereb. Cortex Cereb. Empirical evaluations of slice-timing, smoothing, and smoothing, slice-timing, of evaluations Empirical al. et Diffusion MRI: From Quantitative Measurement to in vivo in to Measurement Quantitative From MRI: Diffusion Intrinsic signal changes accompanying sensory accompanying changes signalIntrinsic al. et High angular resolution diffusion imaging reveals imaging diffusion resolution angular High al. et , 77–94 (2002). 77–94 17, , 239–254 (2013). 239–254 73, Magn. Reson. Med. Reson. Magn. Psychophysiological and modulatory interactions in interactionsmodulatory andPsychophysiological al. et Neuroimage Mapping resting-state functional connectivity using connectivity functionalresting-state Mapping al. et Neuroimage , 169–183 (2011). 169–183 1, Neuroimage

Effects of image reconstruction on fibre on reconstructionimage of Effects al. et Magn. Reson. Med. Reson. Magn. , 1347–1354 (2008). 1347–1354 59, An improved framework for confound regressionconfound for framework improved An al. et multiple on motion head in-scanner of Impact al. et Cereb. Cortex Cereb. , 2553–2559 (2008). 2553–2559 18, Magn. Reson. Med. Reson. Magn.

, 144–155 (2007). 144–155 34, , 5951–5955 (1992). 5951–5955 89, , 218–229 (1997). 218–229 6,

, 1595–1605 (2008). 1595–1605 40, , 5–14 (1993). 5–14 13, Neuroimage Brain Connect. Brain

, 537–541 (1995). 537–541 34, Magn. Reson. Med. Reson. Magn. , 223–229 (2013). 223–229 23, Neuroimage advance online publication, online advance Nat. Methods Nat.

, 1377–1386 (2005). 1377–1386 54, , 623–632 (2012). 623–632 60, Neuroimage , 401–410 (2011). 401–410 1, Magn. Reson. Med. Reson. Magn. Magn. Reson. Med. Reson. Magn. J. Magn. Reson. B Reson. Magn. J. , 2142–2154 (2012). 2142–2154 59, Neuroimage

, 1198–1201 (2012). 1198–1201 9, , 1336–1349 2009). 1336–1349 61, Diffusion MRI: Diffusion , 240–256 64, (eds., in vivo in , 431–438 59,

48,

103, 35,

54. 53. 52. 51. 50. 49. 48. 47. 46. 70. 69. 68. 67. 65. 64. 63. 62. 61. 60. 59. 58. 57. 56. 55. 45. 72. 71. 66. 1253–1268 (2008). 1253–1268 (2009). rest. and activation during MRI. functional in investigation (2008). fMRI. in integrationfunctional of 875–891 (2011). 875–891 datasets. 35–46 (2007). 35–46 components. ICA identification of automatic connectomics.functional micromovementson head of impact Connect. regression.signal global after distorted become differences 1326–1333 (2001). 1326–1333 data. “resting-state” in cortex cerebral the basis of global resting-state fMRI activity. fMRI resting-state global of basis (2009). introduced?networks anti-correlated are correlations: state resting on regressionsignal global of impact The (2009). 3270–3283 101, networks. brain state restinganticorrelated observed removal. and detection artifact with fMRI, state resting in correlation of sources connectivity. noise? induced 76C clustering of resting-state fMRI data. fMRI resting-state of clustering Mapp. Brain Hum. algorithm. (SOM) map self-organizing a using fMRI in connectivity (2004). 752–763 task. cognitive a of stages distinct during connectivity data. fMRI of analyses coherence partial and coherence using connectivity connections. cortico-cortical fluctuations reflect BOLD low-frequency in Correlations PET. with measured connectivity images. PET in activation for (2011). networks. state resting of characteristics changes in brain connectivity for functional MRI. functional for connectivity brain in changes Mapp. Brain Hum. data.neuroimaging functional for resampling wavelet Spatiotemporal (2006). 267–276 brain. human the of connectivity determining Cordes, D. Cordes, Neural D.A. Leopold, & J.H. Duyn, Ye,F.Q.,Maier, A., M.L., Schölvinck, Z. Saad, P.a.Bandettini, T.B.& Jones,Handwerker, D.a., R.M., Birn, Murphy,K., and signal global The M.E. Raichle, & Snyder,A.Z. D., Zhang, M.D., Fox, V.Perlbarg, R.W.Mapping Cox, & L.A. W.K.,Milbury,Simmons, Z.S., Saad, H.J., Jo, functionalresting-state in noise physiological of role The R.M. Birn, cardiac- correlating or connectivity functionalfcMRI–mappingT.E. Lund, Beckmann, C.F. Modelling with independentcomponents. with ModellingC.F. Beckmann, group cut Normalized H. Pol,Hulshoff & R. Mandl, M., Heuvel, den van functionallow-frequency Detecting D.C. Noll, & T.A.Polk, Peltier,S.J., S.M. Smith, functionalMeasuring M. D’Esposito, & A. Gazzaley, J., Rissman, functionalinterregional Measuring M. D’Esposito, & F.T.,L.M. Miller,Sun, M.D. V.P.Phillips, Mathews,J.T., & Lurito, M., Dzemidzic, M.J., Lowe, G. Varoquaux, G. Marrelec, Poline,J.B. & A. Gramfort, G., Varoquaux, S.M. Smith, effective in changesdependent Time C.D. Frith, & K.J. Friston, G. Marrelec, space scale Searching A.C. Evans, P.& Neelin, S., Marrett, Worsley,K.J., Spectral S.M. Smith, & C.F.Beckmann, K., Miller, J., Xie, Niazy,R.K., Yan, C.G. Yan, Bellec, P., Marrelec, G. & Benali, H. A bootstrap test to investigate to test bootstrap A H. Benali, & G. P.,Marrelec, Bellec, L.M. Williams, P.& Das,E.T., Bullmore, Brammer, M.J., Breakspear,M., to approach Bayesian A J.K. Rilling, & F.D.Bowman, R.S., Patel, . (Vancouver, Canada, 2010). (Vancouver,Canada, Systems. Processing 891–901 (2012). 891–901 , 10238–10243 (2010). 10238–10243 107, , 183–201 (2013). 183–201 , Neuroimage Trouble at rest: how correlation patterns and group and patterns correlation how rest: at Trouble al. et , 25–32 (2012). 25–32 2, Neuroimage Neuroimage A comprehensive assessment of regional variation in the in variationregional of assessmentcomprehensive A al. et Frequencies contributing to functional connectivity in connectivity functional to contributingFrequencies al. et CORSICA: correction of structured noise in fMRI by fMRI in noise structured of correction CORSICA: al. et Correspondence of the brain’s functional architecturefunctional brain’s the ofCorrespondence al. et fMRI. for methodsmodelling Network al. et Partial correlation for functional brain interactivity brain functional for correlationPartial al. et measureshierarchical brain: the and systems,Regions, al. et Neuroimage Neuroimage A group model for stable multi-subject ICA on fMRI on ICA multi-subject stable for model group A al. et nature methods Magn. Reson. Med. Reson. Magn.

, 647–658 (2004). 647–658 21, , 220–226 (2003). 220–226 20, , 1–25 (2004). 1–25 23, , 571–582 (2010). 571–582 52, , 288–299 (2010). 288–299 51, , 582–587 (2000). 582–587 12, , 864–870 (2012). 864–870 62, Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. Hum. Brain Mapp. Brain Hum. Neuroimage Hum. Brain Mapp. Brain Hum.

Med. Image Anal. Image Med. , 628–629 (2001). 628–629 46, | O.0 O6 JUNE2013 | NO.6 VOL.10 PLoS ONE PLoS Neuroimage Prog. Brain Res. Brain Prog. Advances in Neural Information Neural in Advances Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. AJNR Am. J. Neuroradiol. J. Am. AJNR , 228–237 (2006). 228–237 32, Magn. Reson. Imaging Reson. Magn. Hum. Brain Mapp. Brain Hum.

, e2001 (2008). e2001 3, , 74–90 (1996). 74–90 4, Stat. Sin. Stat.

, 69–79 (1993). 69–79 1,

, 484–496 12,

, 893–905 44, J. Neurophysiol. J. , 13040–13045 106, Neuroimage Neuroimage

Neuroimage review , 259–276 193, Brain 18, Neuroimage

27,

25, |

23,

54,

22, 62, 537 © 2013 Nature America, Inc. All rights reserved. 98. 97. 96. 95. 94. 93. 92. 91. 90. 89. 88. 87. 86. 85. 84. 83. 82. 81. 80. 79. 78. 77. 76. 75. 74. 73. 538 review 15284–15293 (2011). 15284–15293 intrinsic functional connectivity approach. connectivity functionalintrinsic Cereb. Cortex Cereb. connectivity of resting-state networks during fMRI. during networks resting-state of connectivity (2011). pain. future predict and past reflect to activity J. Psychiatr. Res.Psychiatr. J. hallucinations.auditory of treatment rTMS of effects schizophrenia: in region temporo-parietal the of connectivity Functional A. Aleman, Ann. Appl. Stat. Appl. Ann. imaging. (2011). 616–626 graphs. connectivity fMRI from states brain Decoding networks. functional brain human small-world fractal reconfiguration of Adaptive connectome.brain human (2009). 186–198 systems.functional and structural of analysis subgraph. (2009). 1619–1628 62, connectivity.functional state resting from prediction Groups. networks. brain in differencesidentifying (2002). 870–878 15, rate. discovery false the usingneuroimaging functional in (2010). interpretations.and uses connectivity: 2013). April doi:10.1016/j.neuroimage.2013.04.007(11 subjects. across connectomes USA Sci. Acad. Natl. Proc. brain. human resting the of activity spontaneous the sculpts Learning (2010). experiences.recent for memory to related are rest Psychiatry Biol. cingulate.subgenual the with connectivity functionalintrinsic to related is depression for targets stimulation transcranial magnetic of Efficacy 2012). March (25 hbm.22074 nucleus. thalamic medial the of changes connectivity by mediated is plasticity (2010). 1997–2006 regions.category-preferentialvisual in experience recent by modulated are brain human resting the in fluctuations BOLD connectivity. functional with measured networks state resting on state cognitive Hansen, L.K. Multivariate strategies in functional magnetic resonancemagnetic functional in strategies Multivariate L.K. Hansen, D. Ville, De Van Vuilleumier,P.& S., Schwartz, H., Eryilmaz, J., Richiardi, E. Bullmore, & T.Duke, S., Achard, A., Meyer-Lindenberg, D.S., Bassett, X.N. Zuo, the of models graphical graphs: Brain D.S. Bassett, & E.T.Bullmore, theoretical graph networks: brain Complex O.Sporns, & E. Bullmore, common maximal the on based metric distance graph A H. Bunke, N.U.F. Dosenbach, state Disease H.S. Mayberg, X.P.& Holtzheimer,P.E.,Hu, R.C., Craddock, With Control Rate Discovery False H.H. Zhou, & H. Zhao, J.X., Hu, statistic: Network-basedE.T. Bullmore, & A. Fornito, Zalesky,A., combined? be problems testing hypothesis should When B. Efron, maps statistical of ThresholdingT.Nichols, & N.A. Lazar, C.R., Genovese, brain of measures network Complex O.Sporns, & Rubinov,M. functional comparing and Learning R.C. Craddock, & G. Varoquaux, M.S. Koyama, M. Corbetta, & G.L. Romani, G., Committeri, A., Baldassarre, C.M., Lewis, duringcorrelations brain Enhanced L. Davachi, & N. Ketz, A., Tambini, Pascual-Leone,A. & M.D. M.P.,Greicius,White, Buckner,R.L., M.D., Fox, Keeser,D. & J.A. Boer,den E.J., Liemburg, H., Knegtering, A., Vercammen, V.Riedl, O.W.Witte,Perceptual & Axer,H. S., Brodoehl, Hasler,C., Klingner,C.M., low-frequency Correlated D.L. Schacter, & W.D.,Buckner,R.L. Stevens, prior of Effect G.D. Jackson, & D.F. Abbott, A., Stanislavsky,A.B., Waites, S.M. Smith, fMRI. (2012). spontaneous brain activity.brain spontaneous (2010). | O.0 O6 JUNE 2013 | NO.6 VOL.10 Science J. Am. Stat. Assoc. Stat. Am. J. Hum. Brain Mapp. Brain Hum. Repeated pain induces adaptations of intrinsic brain intrinsic of adaptationsinduces pain Repeated al. et Brain Lang. Brain Network centrality in the human functional connectome.functional human the in centrality Network al. et Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. Pattern Recognit. Lett. Recognit. Pattern Prefrontal transcranial direct current stimulation changes stimulation current directtranscranial Prefrontal al. et Temporally-independent functional modes ofTemporally-independent modesfunctional al. et , 1862–1875 (2012). 1862–1875 22, Hum. Brain Mapp. Brain Hum. , 1358–1361 (2010). 1358–1361 329, Cortical signatures of dyslexia and remediation: an remediation:and dyslexia of signatures Cortical al. et

, 595–603 (2012). 595–603 72,

, 197–223 (2008). 197–223 2, , 725–731 (2010). 725–731 44, Prediction of individual brain maturity using maturity brainindividual of Prediction al. et , 186–191 (2007). 186–191 102,

, 17558–17563 (2009). 17558–17563 106, advance online publication, doi:10.1002/ publication, online advance Annu. Rev. Clin. Psychol. Clin. Rev. Annu.

, 1215–1227 (2010). 1215–1227 105, Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. advance online publication, online advance Neuroimage

| , 59–68 (2005). 59–68 24,

, 255–259 (1998). 255–259 19, nature methods

, 19518–19523 (2006). 19518–19523 103, Neuroimage Neuroimage PLoS ONE PLoS Nat. Rev. Neurosci. Rev. Nat. Neuroimage Cereb. Cortex Cereb.

Neuron , 1059–1069 52, , 113–140 (2011). 113–140 7, J. Neurosci. J.

, e55454 (2013). e55454 8, Magn. Reson. Med. Reson. Magn. , 3131–3136 109, , 1197–1207 53, Neuroimage , 280–290 65, Neuroimage , 206–213 57, 20, 31,

10,

56,

103. 102. 101. 100. 99. 109. 108. 107. 106. 105. 104. 118. 117. 116. 115. 114. 113. 112. 111. 110. 125. 124. 123. 122. 121. 120. 119. , 212–216 (2004). 212–216 51, monkey brain. monkey Interv.Assist. attention-deficit/hyperactivity disorder.for functional connectivity.functional resting-state and contrast BOLD task-induced using studiesextended for Med. Reson. Magn. cytoarchitectonic maps and functional imaging data. imagingfunctional and maps cytoarchitectonic Neuroimage interest. of regions based gyral into scans MRI on cortex cerebral human optimal method. optimal Methods Neurosci. J. models. brain scale large in dynamics global and local on structure (2009). monkeys. and humans anesthesia. isoflurane of levels different at network sensorimotor rat in fluctuations signal (BOLD) level-dependentoxygenation blood spontaneous and high magnetic fields. magnetic high and 1999). Design Sequence and Principles Physical Imaging: Resonance publication, doi:10.1016/j.neuroimage.2012.10.064 (30 October 2012). October doi:10.1016/j.neuroimage.2012.10.064(30 publication, experience.FCP/INDI the work: sharing Psychiatry Mol. it? about do to what and tests clinical develop to psychiatry biological (2012). function. brain auditory in differences (2011). e25701 6, brain. rodent awake the in networks state restingreproducible neuroimage.2013.04.002 (6 April 2013). April neuroimage.2013.04.002(6 disorders. Neuroimage fMRI. BOLD in selection angle flip the on effects noise Physiological Z-SAGA. z-shim: single-shot for spin-echo asymmetric artifacts. susceptibility reduced (2006). T. 1.5 and T 3 at analysis whole-brain a losses: sensitivity BOLD susceptibility-induced ofreduction for parameters Becerra, L., Pendse, G., Chang, P.-C., Bishop, J. & Borsook, D. Robust D. Borsook, & J. P.-C.,Bishop, Chang, Pendse,G., L., Becerra, C.P.Pawela, J.L. Vincent, C.Z. Zhu, J. Friedman, & R. Tibshirani,T., Hastie, Tzourio-Mazoyer,N. data M.P.Making Milham, & F.X.Castellanos,B.B., Biswal, M., Mennes, for long so taken it has WhyT.R. Insel, & A.G. Phillips, Kapur,S., inter-individualgenes: and variability On S.G. Stanton, P.& Sörös, C.P.Pawela, K. Wang, Bernstein, M.A., King, K.F. & Zhou, X.J. Zhou, & K.F.King, M.A., Bernstein, R. Venkatesan, & M.R. Bornw,R.W.,Thompson, E.M., Haacke, its and brain the understand to -ome new a Brainnetome:T. Jiang, J.D. Horn, Van S.A. Knock, D.S. Margulies, Lancaster,J.L. S.B.Eickhoff, R.S. Desikan, Grootoonk, S. Grootoonk, J. Bodurka, P.A.& V.,Bandettini, Roopchansingh, J., Gonzalez-Castillo, andgradient-echo of acquisition Simultaneous X. Hu, & K.A. Heberlein, and SNR increased for fMRI BOLD Spiral-in/out C.S. Law, & G.H. Glover, EPI Optimal R. Deichmann, & O.Josephs, C., Hutton, N., Weiskopf, Yacoub,E. Blaimer,M. (Springer,2001). Learning Phineas Gage. Phineas (Academic Press, 2004). Press,(Academic mapping. single-subject brain. single-subject MRI MNI the of parcellation anatomicalmacroscopic a using SPM in (2000). series. time fMRI ofrealignment the in effects (2005). 1325–1335 Discriminative analysis of brain function at resting-state at function brain of analysis Discriminative al. et Temporal scaling properties and spatial synchronization of synchronization spatial and properties scalingTemporal al. et Hum. Brain Mapp. Brain Hum. advance online publication, doi:10.1016/j. publication, online advance Neuroimage Spin-echo fMRI in humans using high spatial resolutions spatial high using humans in fMRI Spin-echo al. et NMR Biomed. NMR SMASH, SENSE, PILS, GRAPPA: how to choose the choose to how GRAPPA: PILS, SENSE, SMASH, al. et , 968–980 (2006). 968–980 31, , 2764–2778 (2011). 2764–2778 54, The effects of physiologically plausible connectivity plausible physiologically of effects The al. et A protocol for use of medetomidine anesthesia in rats in anesthesiamedetomidine of use for protocol A al. et brain. rat the of connectivity functional Resting-state al. et Intrinsic functional architecture in the anaesthetized the in architecturefunctional Intrinsic al. et An automated labeling system for subdividing the subdividing for system labeling automated An al. et A new SPM toolbox for combining probabilistic combining for toolbox SPM new A al. et , 468–475 (2005). 468–475 8, Focus onmappingthebrain Characterization and correction of interpolation of correction and Characterization al. et PLoS ONE PLoS Nature Automated Talairach atlas labels for functional brain functional for labels atlas TalairachAutomated al. et , 1174–1179 (2012). 1174–1179 17, Mapping connectivity damage in the case of case the in damage connectivity Mapping al. et Precuneus shares intrinsic functional architecture in architecturefunctional intrinsic shares Precuneus al. et Top. Magn. Reson. Imaging Reson. Magn. Top.

, 1021–1029 (2008). 1021–1029 59,

Automated anatomical labeling of activations of labeling anatomical Automated al. et , 86–94 (2009). 86–94 183, Neuroimage , 83–86 (2007). 83–86 447, Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. Neuroimage , e37454 (2012). e37454 7, , 61–67 (2011). 61–67 24, Magn. Reson. Med. Reson. Magn.

, 120–131 (2000). 120–131 10, Magn. Reson. Med. Reson. Magn. , 273–289 (2002). 273–289 15, , 1137–1147 (2009). 1137–1147 46, Front. Hum. Neurosci. Hum. Front. advance online advance Neuroimage The Elements of Statistical of Elements The Handbook of MRI Pulse Sequences Pulse MRI of Handbook Neuroimage Med. Image Comput. Comput. Comput. Image Med.

, 223–236 (2004). 223–236 15, , 655–664 (2003). 655–664 49, Neuroimage

, 20069–20074 106,

, 515–522 (2001). 515–522 46, , 493–504 33, Neuroimage Magn. Reson. Med. Reson. Magn. (Wiley-Liss, , 49–57 11,

, 150 6, Magnetic PLoS ONE PLoS 25,

© 2013 Nature America, Inc. All rights reserved. 138. 137. 136. 135. 134. 133. 132. 131. 130. 129. 128. 127. 126. Focus onmappingthebrain 11379–11387 (2010). 11379–11387 multi-slice echo planar imaging. planar echo multi-slice (2010). 1144–1153 fMRI. whole-brain temporal and spatial high to application with imaging parallel partial using acceleration (2010). imaging. diffusion fast and FMRI brain whole Brain Connect. Brain imaging. volumar echo , 295 (2012). 295 3, States. consciousnesspathological and pharmacological, physiological, in connectivity network state resting multiple electrophysiological validation.electrophysiological selection, causality and deconvolution. and causality selection, model fMRI: using brain the in networks interacting identification of the (2013). 526–537 data. echo short using fluctuations BOLD intrinsic from activation. cortical of imaging Scannell, J.W., Burns, G.A., Hilgetag, C.C., O’Neil, M.A. & Young, M.P.Young, & M.A. O’Neil, C.C., Hilgetag, G.A., J.W.,Burns, Scannell, O.David, on: Comments causality Granger and modeling causal Dynamic K. Friston, Spoormaker,V.I. S.G. Horovitz, M. Boly, L. Heine, artifacts physiological and motion Removing Murphy,K. & M.G. Bright, D.A. Feinberg, K. Setsompop, Moeller,S. C. Rabrait, MR K-space Schneider,W.Spiral & Meyer,C.H. J.D., Cohen, D.C., Noll, Cereb. Cortex Cereb. cat. the of system cortico-thalamic the of organizationconnectional The network during human non-rapid eye movement sleep. movement eye non-rapidhuman during network sleep. deep during Brain connectivity in disorders of consciousness.of disorders in connectivity Brain al. et Resting state networks and consciousness: alterations of alterationsconsciousness: and networks state Resting al. et Identifying neural drivers with functional MRI: an MRI: functional with drivers neuralIdentifying al. et Multiband multislice GE-EPI at 7 tesla, with 16-fold with tesla, 7 at GE-EPI multislice Multiband al. et High temporal resolution functional MRI using parallel using MRI functionalresolution temporal High al. et , 277–299 (1999). 277–299 9, Decoupling of the brain’s default mode network mode default brain’s the of Decoupling al. et Multiplexed echo planar imaging for sub-second for imaging planar echo Multiplexed al. et , 1–10 (2012). 1–10 2, Improving diffusion MRI using simultaneous using MRI diffusion Improving al. et Development of a large-scale functional brain functionallarge-scale a of Development al. et Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. J. Magn. Reson. Imaging Reson. Magn. J. J. Magn. Reson. Imaging Reson. Magn. J. PLoS Biol. PLoS Neuroimage Neuroimage Magn. Reson. Med. Reson. Magn. , 2683–2697 (2008). 2683–2697 6,

, 569–580 (2012). 569–580 63, , 11376–11381 (2009). 11376–11381 106, PLoS ONE PLoS

, 744–753 (2008). 744–753 27, , 303–305 (2011). 303–305 58, J. Neurosci. J.

Neuroimage , e15710 5, , 49–56 (1995). 49–56 5, Front. Psychol. Front.

63,

30, 64,

148. 147. 146. 145. 144. 143. 142. 141. 140. 139. 150. 149. (2004). study.potential evoked cortico-cortical a system: , 817–827 (2008). 817–827 18, TMS-fMRI. concurrent from evidence cortex: visual human on conventional rectangular pad. rectangularconventional versus electrode ring a using focality spatial improved stimulation: concurrent transcranial magnetic stimulation.magnetic transcranialconcurrent (1998). causality analysis. causality Granger to reference with issues current and findings past data: fMRI (2008). networks.default-mode andcentral-executive between switching in cortex fronto-insular methods.cross-spectral mapping. causality Granger and modelingautoregressive vector using data fMRI coordination. motor bimanual of study fMRI an modeling:equation structural with exploration modelling. causal dynamic Mapp. imaging. brain functional in analysis network to application Bohning, D.E. Bohning, R. Matsumoto, J.M. Johnston, from connectivity brain effective Investigating X. Hu, & G. Deshpande, right the for role critical V.A Menon, & D.J. Levitin, D., Sridharan, and modelseconometric by relations causal Granger,C.W.Investigating time-resolved in interactionscortical directed Investigating R. Goebel, Connectivity X. Hu, & K. Zhang, Peltier,S., S., LaConte, J., Zhuang, on comments Critical Turner,R. & Müller,K. K., Erfurth, G., Lohmann, its and modelingequation StructuralF. Gonzalez-Lima, & A.R. Mclntosh, Datta, A. Datta, C.C. Ruff, , 6453–6458 (2008). 6453–6458 28, callosum. corpus the of section complete after connectivity , 2–22 (1994). 2–22 2, Gyri-precise head model of transcranial direct current directtranscranial of model head Gyri-precise al. et Magn. Reson. Imaging Reson. Magn. Distinct causal influences of parietal versus frontal areas frontal versus parietal of influences causal Distinct al. et Echoplanar BOLD fMRI of brain activation induced by induced activation brain of fMRI BOLD Echoplanar al. et Functional connectivity in the human languagehuman the in connectivity Functional al. et nature methods Loss of resting interhemispheric functionalinterhemispheric resting of Loss al. et Brain Connect. Brain Neuroimage Proc. Natl. Acad. Sci. USA Sci. Acad. Natl. Proc. Econometrica Neuroimage Brain Stimulat. Brain

, 1251–1261 (2003). 1251–1261 21, , 462–470 (2005). 462–470 25, , 235–245 (2012). 235–245 2, | O.0 O6 JUNE2013 | NO.6 VOL.10 , 424–438 (1969). 424–438 37, , 2322–2329 (2012). 2322–2329 59, Invest. Radiol. Invest. , 201–207 (2009). 201–207 2,

Brain , 12569–12574 105, , 2316–2330 127, review J. Neurosci. J. Hum. Brain Hum. , 336–340 33, Cereb. Cortex Cereb.

|

539 RESOURCE

Improved tools for the Brainbow toolbox

Dawen Cai1,2, Kimberly B Cohen1,2, Tuanlian Luo1,2, Jeff W Lichtman1,2 & Joshua R Sanes1,2

In the transgenic multicolor labeling strategy called ‘Brainbow’, intensity, failure to fill all axonal and dendritic processes, and Cre-loxP recombination is used to create a stochastic choice disproportionate expression of the ‘default’ (that is, nonrecom- of expression among fluorescent proteins, resulting in the bined) XFP in the transgene. We have now addressed several of indelible marking of mouse neurons with multiple distinct these limitations, and we present here a new set of Brainbow rea- colors. This method has been adapted to non-neuronal cells in gents. In addition, we provide guidelines for imaging Brainbow- mice and to neurons in fish and flies, but its full potential has expressing tissue. yet to be realized in the mouse brain. Here we present several lines of mice that overcome limitations of the initial lines, RESULTS and we report an adaptation of the method for use in adeno- Design of Brainbow 3.0 transgenes associated viral vectors. We also provide technical advice As a first step in improving Brainbow methods, we sought XFPs about how best to image Brainbow-expressing tissue. with minimal tendency to aggregate in vivo, high photostability and maximal stability with respect to paraformaldehyde fixa- The discovery that recombinant jellyfish GFP fluoresces when tion. Because some XFPs that ranked highly in cultured cells expressed in heterologous cells1 has led to a vast array of powerful performed poorly in vivo, we generated transgenic lines from 15 methods for marking and manipulating cells, subcellular com- XFPs (Supplementary Table 1 and refs. 2–13). Of the XFPs tested partments and molecules. The discovery or design of numerous in this way, seven were judged suitable: mTFP1, EGFP, EYFP, spectral variants2–13 (fluorescent proteins collectively called XFPs; mOrange2, TagRFPt, tdTomato and mKate2. ref. 14) expanded the scope of the ‘GFP revolution’ by enabling From these XFPs, we chose three according to the criteria discrimination of nearby cells or processes labeled with contrast- of minimal spectral overlap and minimal sequence homology. ing colors. At least for the nervous system, however, two or three These XFPs were mOrange2 from a coral (excitation peak colors are far too few because each axon or dendrite approaches (Ex) = 549 nm, emission peak (Em) = 565 nm), EGFP from a

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature hundreds or thousands of other processes in the crowded neuropil jellyfish (Ex = 488 nm, Em = 507 nm) and mKate2 from a sea of the brain. anemone (Ex = 588 nm, Em = 635 nm)2,9,11. Our reason for mini- Several years ago, we developed a transgenic strategy called mizing sequence homology was to ensure that the XFPs would be 15

npg Brainbow that addresses this problem by marking neurons with antigenically distinct, in contrast to spectrally distinct but anti- many different colors. In this method, three or four XFPs are genically indistinguishable jellyfish variants (EBFP, ECFP, EGFP incorporated into a transgene, and the Cre-loxP recombination and EYFP). Exploiting this property, we generated antibodies to system16 is used to make a stochastic ‘choice’ of a single XFP to be the XFPs in different host species (rabbit anti-mCherry, anti- expressed from the cassette. Because multiple cassettes are inte- mOrange2 and anti-tdTomato; chicken anti-EGFP, anti-EYFP and grated at a single genomic site, and the choice within each cassette anti-ECFP; and guinea pig anti-mKate2, anti-TagBFP and anti- is made independently, combinatorial expression can endow indi- TagRFP; Supplementary Table 1). Tests in transfected cultured vidual neurons with 1 of ~100 colors, providing nearby neurons cells confirmed a lack of cross-reactivity (data not shown). with distinct spectral identities. Next we addressed the need to fill all parts of the cell evenly. If Cre recombinase is expressed transiently, descendants of Unmodified XFPs labeled somata so strongly that nearby proc- the initially marked cell inherit the color of their progenitor. esses were difficult to resolve, whereas palmitoylated derivatives, Accordingly, the Brainbow method has been adapted for use in which targeted the XFPs to the plasma membrane, were selectively lineage analysis in non-neural tissues of mice17–21. In addition, it transported to axons and labeled dendrites poorly15. We there- has been adapted for analyses of neuronal connectivity, cell migra- fore generated farnesylated derivatives27, which were trafficked tion and lineage in fish22,23 and flies24,25. In contrast, the method to membranes of all neurites (see below). has been little used in the mouse nervous system26. We believe On the basis of these results, we generated ‘Brainbow 3.0’ trans- that the main reasons for this are limitations of the initial set of genic lines incorporating farnesylated derivatives of mOrange2, Brainbow transgenic mice. These include suboptimal fluorescence EGFP and mKate2. We retained the Brainbow 1 format15, in which

1Center for Brain Science, Harvard University, Cambridge, Massachusetts, USA. 2Department of Molecular and Cellular Biology, Harvard University, Cambridge, Massachusetts, USA. Correspondence should be addressed to J.R.S. ([email protected]). Received 3 October 2012; accepted 30 March 2013; published online 5 May 2013; doi:10.1038/nmeth.2450

540 | VOL.10 NO.6 | JUNE 2013 | nature methods RESOURCE

Figure 1 | Brainbow 3 transgenic mice. a Brainbow 1.0 (a) Brainbow constructs and transgenic mice. Thy1 P 2 RFP pA P YFP pA 2 CFP pA Brainbow 1.0 is described in ref. 15. Brainbow Brainbow 3.0 3.0 incorporates farnesylated (‘f’), antigenically Thy1 2 N mO2f pA 2 EGFPf pA N mK2f pA

distinct XFPs: mOrange2f (mO2f), EGFPf and Brainbow 3.1 mKate2f (mK2f). Brainbow 3.1 incorporates a Thy1 P 2 N ∅NFPnls pA P mO2f pA 2 EGFPf pA N mK2f pA nuclear-targeted nonfluorescent XFP (ØNFPnls) Brainbow 3.2 in the first (default) position. Brainbow 3.2 Thy1 P 2 N ∅NFPnls pA P mO2f W pA 2 EGFPf W pA N mK2f W pA incorporates a woodchuck hepatitis virus post- transcriptional regulatory element (W) into Brainbow 3.1. P, loxP site; 2, lox2272; N, loxN; b c d pA, polyadenylation sequence. (b,c) Low- (b) and high-power (c) views of muscles from a Brainbow 3.0 (line D) Islet-Cre mouse, showing terminal axons and neuromuscular junctions in extraocular muscle. (d) Rotated image along dashed bar in c showing five motor axons labeled in distinct colors. The open circles show that farnesylated XFPs mark plasma membranes more than cytoplasm. (e) Cerebellum from a Brainbow 3.1 (line 3) L7-Cre mouse. The ten Purkinje cells in this field are labeled by at least seven distinct colors (antibody amplified and numbered i–vii). Because Cre is selectively e vii f expressed by Purkinje cells in the L7-Cre line, no other cell types are labeled. (f) Cerebellum v P from a Brainbow 3.2 (line 7) CAGGS-CreER iv mouse showing granule native fluorescence in iii vi red, pink, yellow, green, cyan, blue and brown.

P, parallel fibers in molecular layer. Purkinje i cell bodies, which are unlabeled, are outlined. ii ii Scale bars: 50 µm (b), 20 µm (c,e), 5 µm (d), iii i 10 µm (f). iv v vi vii incompatible wild-type and mutant loxP sites are concatenated so that Cre recombi- nase yields a stochastic choice among XFPs (Fig. 1a). We also retained two other features of the Brainbow 1 into a first position. With this modification, the three XFPs were

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature strategy. First, we used neuron-specific regulatory elements from expressed only in Cre-positive neurons, giving more spectral the Thy1 gene14 because it promotes high levels of transgene diversity. Second, we inserted a fourth XFP, Phi-YFP (from the expression in many, although not all, neuronal types; other pro- hydrozoon Phialidium)4, which is antigenically distinct from the

npg moter-enhancer sequences that we tested support considerably other three, into the stop cassette. We mutated Phi-YFP to elimi- lower levels of expression21. Second, we generated transgenic lines nate its endogenous fluorescence (PhiYFP Y65A), fused it to a by injection into oocytes because this method leads to integration nuclear localization sequence and generated antibodies against of multiple copies of the cassette and, thus, a broad spectrum of it in rat. In Brainbow 3.1 mice generated from this cassette outcomes15,28; by contrast, knock-in lines generated by homol- (Fig. 1a), one can screen sections with rat anti–Phi-YFP before ogous recombination contain one or two copies of the cassette Cre recombination (Supplementary Fig. 1). (as heterozygotes or homozygotes, respectively) and, consequently, Finally, we inserted a sequence that stabilizes mRNAs, called a smaller number of possible color combinations17–20. a woodchuck hepatitis virus post-transcriptional regulatory ele- ment (WPRE), into the 3′ untranslated sequences following each Design of Brainbow 3.1 and 3.2 transgenes XFP. The WPRE has been used in many cases to increase protein In Brainbow 1, 2 and 3.0 (Fig. 1a) and ref. 15, one XFP is expressed levels produced by viral vectors and transgenes30,31. We call lines ‘by default’ in Cre-negative cells. The presence of a default XFP has incorporating the WPRE ‘Brainbow 3.2’ (Fig. 1a). both drawbacks and advantages. In cases of limited Cre expres- sion, this XFP is expressed in a majority of cells, reducing spectral Characterization of Brainbow 3 mice diversity among recombined neurons. On the other hand, incorpo- We generated 31 lines of transgenic mice from the Brainbow 3.0, ration of a default XFP allows one to screen numerous lines in the 3.1 and 3.2 cassettes. Offspring were crossed with several Cre trans- absence of a Cre reporter to assess the number and types of cells in genic lines32–37, leading to multicolor spectral labeling of neuronal which XFPs could be expressed following recombination. populations in numerous regions including cerebral cortex, brain- To eliminate the default XFP while retaining the ability to assess stem, cerebellum, spinal cord and retina (Fig. 1b–f, Supplementary expression in the absence of Cre, we adopted the following strat- Table 2, Supplementary Fig. 2 and Supplementary Videos 1 egy. First, we incorporated three rather than two pairs of incom- and 2). The intensity of expression varied markedly among lines, patible loxP sites, which allowed insertion of a ‘stop’ cassette29 making quantitative comparison of dubious value, but the most

nature methods | VOL.10 NO.6 | JUNE 2013 | 541 RESOURCE

Figure 2 | Improved visualization of neurons in Brainbow 3 mice. (a,b) Cerebellum from a b Brainbow 1.0 (a) and Brainbow 3.1 (line 3) L7-Cre (b) mice. Insets show high-magnification views of boxed regions. The farnesylated XFPs clearly label the fine processes and dendritic spines (b, inset arrows), which are missing in the cytoplasmic labeling (a, inset). (c,d) Retina from a Brainbow 3.0 (line D) Islet-Cre mouse expressing EGFP (blue), mOrange2 (green) and mKate2 (red). Intrinsic XFP fluorescence c d (c) and a nearby section immunostained with chicken anti-GFP, rabbit anti-mOrange2 and guinea pig anti-mKate2 (d) are shown. (e–h) Immunostained cerebellum section from Brainbow 3.2 (line 7) parvalbumin-Cre mouse. Separate channels of region boxed in e are shown in f–h. (i) Fraction of all labeled neurons that express EGFP, mOrange2 (mO2) or mKate2 (mK2) (n = 2,037 neurons from 15 regions of two brains). Scale bars: 40 µm (a,b), 20 µm (c,d), 25 µm (e–g), 10 µm (insets). e f h

strongly expressing lines were those that incorporated the WPRE (Brainbow 3.2). Analysis of these lines confirmed their advantages over Brainbow 1 and 2 lines15. First, the use of farnesylated XFPs led the g i n = 2,037 XFPs to concentrate at the plasma mem- 50 brane (Fig. 1d). As a consequence, somata 40 were less intensely labeled in Brainbow 3 30 than in Brainbow 1 mice, so processes 20 could be visualized without saturation of 10 % of total neurons 0 the somata (Fig. 2a,b and Supplementary + + + EGFP mO2 mK2 Fig. 3). Use of farnesylated XFPs also improved labeling of fine processes and

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature dendritic spines (Fig. 2a,b). Second, the ability to immunostain combinatorial expression of XFPs in multiple neurons (Fig. 3b all three XFPs led to enhancement of the intrinsic fluorescence and Supplementary Fig. 2e). We were concerned that precocious without loss of color diversity (Fig. 2c,d and Supplementary Cre activation in the germ line might lead to loss of the cassette.

npg Fig. 4). Third, XFP fluorescence was visible only in Cre- We therefore established a line from the third founder and exam- positive cells, thereby correcting the color imbalance caused ined mice in the second and sixth generations. Color range was by default XFP expression in cells that were Cre negative or limited in this line, perhaps because only a few copies had been cells that expressed Cre at low levels in Brainbow 1 and 2 lines integrated into the genome, but the variety of colors and level of (Fig. 2e–i and Supplementary Fig. 5). Thus, Brainbow 3 lines expression were similar in both generations (Fig. 3c,d). Thus, are likely to be more useful than Brainbow 1 and 2 lines for Autobow transgenes can be stably maintained. multicolor labeling. Brainbow using Flp recombinase and FRT sites Brainbow with self-excising Cre recombinase A second recombination system, orthogonal to Cre-loxP, could In Brainbow 1–3 lines, the cassette encodes XFPs separated by be used to independently control expression of distinct XFPs in, loxP sites; Cre recombinase is supplied from a separate trans- for example, excitatory and inhibitory neurons. In this way, the gene. For analysis of connectivity in mouse mutants, breeding color of a neuron could denote cell identity, a feature lacking in mice with two transgenes (Brainbow and Cre) into an already currently available Brainbow lines15,17–20. We therefore tested a complex background is cumbersome. We therefore attempted to second recombination system, in which Flp recombinase acts on combine constructs encoding XFPs and Cre in a single cassette. Flp recombinase target (FRT) sites. The Flp-FRT system has been In this transgene, called ‘Autobow’, we substituted a self-excising used in conjunction with Cre-loxP in mice16 and in Brainbow-like Cre recombinase38 for the stop sequences in Brainbow 3.1 transgenes in Drosophila24,25. (Fig. 3a). The Thy1 regulatory elements lead to expression of Cre We tested previously described mutant FRT sites39,40 to find in differentiated neurons; Cre then simultaneously activates an incompatible sets (Supplementary Fig. 6) and used these sets to XFP and excises itself. construct ‘Flpbow’ lines (Fig. 4a). In one cassette, we fused the We generated three founder mice using this construct. XFPs to an epitope tag41, which allowed for discrimination of cells Two were analyzed as adults, and both of these exhibited labeled by Cre- and Flp-driven cassettes (Supplementary Fig. 7).

542 | VOL.10 NO.6 | JUNE 2013 | nature methods RESOURCE

a 1 2 3 Figure 3 | Autobow. (a) Schematic of the Autobow construct and the steps that lead Thy1 P 2 N Cre pA P Cerulean pA 2 PhiYFP pA N mKate2 pA Cre to expression of the three XFP combinations (outcomes 1–3). P, loxP site; 2, lox2272; 1 P Cerulean pA 2 PhiYFP pA N mKate2 pA Cerulean N, loxN; pA, polyadenylation sequence. (b) Labeling of hippocampal neurons by Autobow

2 P 2 PhiYFP pA N mKate2 pA PhiYFP in founder 3. The 20 large neurons (diameter <5 µm) in this section are labeled in 20 distinct colors. Antibody-amplified Cerulean, PhiYFP and pA 3 P 2 N mKate2 mKate2 mKate2 are in blue, green and red, respectively. (c,d) Cortical neurons of an Autobow mouse line in the second (c) and sixth (d) generations. b Antibody-amplified Cerulean, PhiYFP and mKate2 are as in b. Scale bars, 50 µm.

Cre-independent expression, and WPRE elements were added to increase expres- sion. In this design, recombination can lead to three outcomes from two XFPs: XFP1, XFP2 or neither. We generated two AAVs with two XFPs each, such that co-infection would lead to a minimum of eight hues (3 × 3 – 1; Fig. 5c). Because AAV can infect cells at high multiplicity, the number of possible colors is 8. An addi- tional feature is that excision of the non- expressed XFP in a second step (Fig. 5a) enhances and equalizes expression of the c d remaining XFP (Fig. 5d,e). We infected cortex, cerebellum and retina of Cre transgenic mice with these vectors. When examined 3–5 weeks later, neurons were labeled in multiple colors (Fig. 5f–j and Supplementary Video 3). Near injec- tion sites, high levels of infectivity led to

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature coexpression of all XFPs in single cells so that neurons appeared gray or white. The variety of colors increased with distance

npg from these sites and then decreased again in sparsely injected regions, presumably because each labeled neuron received only one virion (Supplementary Fig. 9).

When Flpbow mice were mated to Flp-expressing mice42,43, we Methods to optimize Brainbow imaging observed multicolor labeling (Fig. 4b,c). Although the few lines Obtaining high-quality images from of tissues expressing tested to date exhibit narrow expression patterns, these results Brainbow transgenes is challenging. Because colors are derived demonstrate that Flp- and Cre-based Brainbow systems can be by mixing images of multiple fluorophores over a wide range of used in combination. concentrations, factors that differentially affect the labels degrade the final image. In addition, it is usually necessary to image a Brainbow adeno-associated viral vectors tissue volume rather than a single section, so methods for taking In parallel to developing Brainbow transgenic lines, we gener- image stacks must be optimized. Here we summarize guidelines ated adeno-associated viral (AAV) vectors to provide spatial for imaging Brainbow tissue. and temporal control over expression and to make the method applicable to other species. Because the Brainbow 3.1 cassette Sample preparation. To minimize background, section thickness described above is >6 kilobase pairs (kb), but the capacity of AAV should be less than the working distance of the objective, gener- vectors is <5 kb, we re-engineered the cassette. On the basis of ally <100 µm for high–numerical aperture (high NA) lenses. It is results from initial tests (Supplementary Fig. 8), we devised a also important to match the refractive indices of the immersion scheme in which loxP sites with left or right element mutations44 medium and the sample because chromatic aberrations caused by were used for unidirectional Cre-dependent inversion (Fig. 5a,b). mismatches between these values lead to spatial offsets between Farnesylated XFPs were positioned in reverse orientation to prevent color channels (Supplementary Fig. 10). Commercial antifade

nature methods | VOL.10 NO.6 | JUNE 2013 | 543 RESOURCE

Figure 4 | Flpbow. (a) Schematics of Flpbow a transgenes. In Flpbow 1, the loxP sites of Flpbow 1 Brainbow 1.0 were replaced by the incompatible Thy1 5T2 545 tdTomato pA 5T2 PhiYFP pA 545 ECFP pA FRT sites Frt5T2 and Frt545. In Flpbow 3, the loxP sites of Brainbow 3.2 were replaced Flpbow 3 by incompatible FRT sites F3, Frt5T2 and Thy1 F3 5T2 545 ØNFPnls pA F3 S-mO2f W pA 5T2 S-EGFPf W pA 545 S-mK2f W pA Frt545. XFPs were fused to SUMO-Star (‘S-’). f indicates farnesylation; pA, polyadenylation b c sequence; W, woodchuck hepatitis virus post-transcriptional regulatory element. (b) Neurons in caudate putamen of a Flpbow 1 Wnt-Flp double transgenic. Neurons are labeled by at least nine colors (red, orange, yellow, yellow-green, green, cyan, blue, purple and pink). Native fluorescence of ECFP, tdTomato and antibody-amplified PhiYFP are in blue, red and green, respectively. (c) Mossy fibers in a Flpbow3 Wnt-Flp double transgenic. Fluorescence of antibody-amplified EGFP, mOrange2 and mKate2 are in blue, red and green, respectively. Scale bars: 50 µm (b), 20 µm (c).

mountants such as Vectashield (Vector Labs) or ProLong Gold power, photomultiplier-tube voltage or digital gain as a function (Invitrogen) that have refractive indices of ~1.47 are optimal for of depth. Imaging parameters can be adjusted to obtain images objectives that use glycerin as the immersion medium. Polyvinyl with similar signal ranges throughout the stack. alcohol mountants (such as Mowiol 4-88; Sigma-Aldrich) provide a better match for oil-immersion (refractive index of Image processing. Brainbow images must be postprocessed to ~1.52) objectives. maximize color information, but care is needed to avoid intro- ducing artifacts. Often one begins by reducing noise. Because Confocal laser scanning microscopy. Epifluorescence micros- confocal laser scanning of multicolor stacks is generally done at copy can be used for imaging thin sections (<10 µm) or mon- speeds of ~1 µs per pixel or less to save time, the small number olayer cultures, but confocal microscopes are preferable for of photons collected for each pixel gives rise to sufficient shot thick specimens because they decrease contamination by light noise to cause perceptible local color differences. This problem from outside the plane of focus. Newly developed two-photon can be minimized by slower scanning or averaging of multiple multi-XFP imaging techniques are also useful45–47. Apochromatic scans, but when this is infeasible, simple filtering and deconvo-

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature or fluorite microscope objectives that are corrected for three lution methods are helpful (Fig. 6a–c). For example, median or or more colors are strongly recommended. Most lens manu- Gaussian filters with radii of 0.5–2 pixels reduce color noise, but facturers specify preferred oils and coverslips. Using the wrong at the expense of resolution. Deconvolution algorithms (Online

npg oil or coverslip degrades the sharpness of focus and increases Methods) are more challenging to use than simple filters but can chromatic aberration. remove color noise without compromising spatial resolution Fluorophores with overlap in the excitation or emission spec- (Supplementary Fig. 11). tra should be imaged sequentially rather than simultaneously Subsequent processing steps can expand the detectable color to minimize fluorescence cross-talk and thereby optimize color range and correct for color shifts (Fig. 6d,e). To obtain easily separation. Laser power should be set as low as possible for sev- perceived color differences, pixel intensity values for each channel eral reasons. First, all planes are bleached as each image plane in each image are normalized to the same minimum and maxi- is scanned, so generation of stacks leads to gradual bleaching mum intensity values for that color in the whole image stack. and decreased signal through the stack. Second, because each This linearly stretches all channels and images to the full dynamic fluorophore bleaches at a different rate, colors may shift during range. Color shifts also arise because illumination strength is gen- imaging. Third, linear signaling requires that fluorophores emit erally uneven across the imaging field and differs among lasers. photons at submaximal rates; at higher excitation intensities, only This effect can be attenuated by intensity or shading correction for the out-of-focus signal is increased48. Fourth, if one fluorophore each channel46,49. The resulting composite RGB images provide species is saturated but another is not, small changes in laser maximum color separation for viewing by eye (Fig. 6f). power will affect the fluorophore intensities differently, leading to color change. With high-NA objectives, a laser power of just DISCUSSION a few milliwatts is saturating; this is generally a small percentage The goal of the work reported here was to design, generate and of the total power the laser can provide. characterize improved reagents for multicolor Brainbow imag- With laser power adjusted to a low level, the photomultiplier- ing of neurons in mice. First, we generated new transgenic lines tube voltages and digital gains must be set to relatively high val- that overcome some limitations of the Brainbow 1 and 2 lines ues. In some confocal microscopes it is possible to compensate for that are currently available15. The improvements were the sub- signal loss from deep layers through automatic adjustment of laser stitution of XFPs (especially red and orange fluorescent proteins)

544 | VOL.10 NO.6 | JUNE 2013 | nature methods

RESOURCE EYFPf

Figure 5 | Brainbow AAV. (a) AAV Brainbow a AAV no. 1 5′ITR EF1 TagBFPf W pA 3′ITR

constructs and recombination scheme.

mTFPf Farnesylated TagBFP and EYFP or mCherry and mChef AAV no. 2 5′ITR EF1 pAW 3′ITR mTFP were placed in reverse orientation between mutant loxP sites. EF1α, regulatory elements 1 1 from elongation factor 1α gene;

W, woodchuck hepatitis virus post-transcriptional 2 regulatory element; pA, polyadenylation mTFPf mChef sequence; mChef, mCherryf; triangles, loxP mTFPf mChef mutants and their recombination products 3 4 (dark and light green sectors indicate wild-type and mutant portions of loxP sites, respectively);

1–6, recombination events. (b) Outcomes of mChef mTFPf mTFPf recombination events numbered in a. Unfilled arrows, direction of intervening cDNA. An X 5 6 6 signifies that a fully mutant (light green) loxP site cannot serve as substrate for Cre. (c) Eight b c color outcomes resulting from pairs of Brainbow AAV no. 1 AAV no. 2 Output f AAVs following recombination as shown in a. 1 This is a minimum value because it does not account for differences in relative intensity 2 of the four XFPs. (d,e) Test of color balance. 3 The mTFPf-mChef AAV vector was injected at 4 low titer into the cortex of a Thy1-Cre mouse, 5 and neurons of each color were counted. An image from the cortex (d) and the fraction of 6 neurons expressing mTFPf (58%) and mChef d (42%; n = 1,523 neurons in four sections of 70 three mice; s.d. in red) are shown. (f–h) Cortex e g h 60 of parvalbumin (PV)-Cre mice injected with the 50 two Brainbow AAVs. High-magnification views of 40 adjacent planes from the yellow (g) and white 30 (h) boxed regions in f are shown. (i) Retina from 20

AAV-injected mouse expressing Cre in retinal % of total neurons 10 ganglion cells. Arrows, dendrites; arrowheads, 0 axons. (j) Cerebellum from AAV-injected PV-Cre mTFPf+ mChef+ j mouse. Fan-shaped dendrites of Purkinje cells (white arrows) and interneurons (yellow arrows) i are indicated. In f–j, antibody-amplified mTFP and EYFP are in green, TagBFP is in blue and © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature mCherry is in red. Scale bars: 100 µm (d), 20 µm (f), 10 µm (g,h), 50 µm (i,j). npg that are more photostable and less prone to aggregation than those used initially; use of XFPs with minimal sequence homology so they could be of a default XFP; inclusion of a nonfluorescent marker in the immunostained separately; farnesylation of the XFPs for even default position to facilitate screening of multiple lines; and staining of somata and the finest processes; insertion of a stop insertion of a WPRE to boost expression (Figs. 1 and 2). These cassette to increase color variety by eliminating broad expression lines, which incorporate regulatory elements from the Thy1 gene, enable marking of many but not all neuronal types. To date, a b c elements tested other than those from the Thy1 gene do not sup- port the high expression levels needed to image Brainbow 1 and 2 material. The ability to immunostain provided by Brainbow 3 cassettes may allow weaker promoters to be used. Second, we designed two additional transgenes and performed initial tests to demonstrate that they can be used effectively in vivo.

f e d Figure 6 | Processing a Brainbow image. (a) Original image from parvalbumin- Cre mouse cortex injected with mixed Brainbow AAVs. (b) Region boxed in a. (c) Deconvolution, to decrease noise without sacrificing spatial resolution. (d) Intensity normalization, to expand perceptible color range. (e) Color- shift correction (Supplementary Fig. 10). (f) Fully processed image. Yellow arrows indicate the sequence of image processing. White arrowheads indicate corresponding objects in the original and color shift–corrected images. Scale bars, 10 µm (a,f), 3 µm (b–e).

nature methods | VOL.10 NO.6 | JUNE 2013 | 545 RESOURCE

One construct, called Autobow, incorporates a self-excising Cre 1. Tsien, R.Y. The green fluorescent protein. Annu. Rev. Biochem. 67, recombinase. Autobow lacks the temporal and spatial control 509–544 (1998). 2. Heim, R. & Tsien, R.Y. Engineering green fluorescent protein for improved afforded by the use of specific Cre lines or ligand-activated Cre brightness, longer wavelengths and fluorescence resonance energy transfer. (CreER). However, because it does not require the generation of Curr. Biol. 6, 178–182 (1996). double transgenics, it may be useful for rapid screening of neu- 3. Rizzo, M.A., Springer, G.H., Granada, B. & Piston, D.W. An improved cyan fluorescent protein variant useful for FRET. Nat. Biotechnol. 22, 445–449 ronal morphology in mutant mice or mice submitted to various (2004). experimental interventions (such as drug treatments). The other 4. Shagin, D.A. et al. GFP-like proteins as ubiquitous metazoan superfamily: novel transgene, Flpbow, replaces loxP sites with FRT sites so that evolution of functional features and structural complexity. Mol. Biol. recombination can be controlled by Flp recombinase rather than Evol. 21, 841–850 (2004). 5. Ai, H.W., Henderson, J.N., Remington, S.J. & Campbell, R.E. Directed Cre recombinase. Flpbow 3 also incorporates an epitope tag so evolution of a monomeric, bright and photostable version of Clavularia that XFPs in Flpbow can be distinguished immunohistochemi- cyan fluorescent protein: structural characterization and applications cally from XFPs in Brainbow. By using Cre and Flp transgenic in fluorescence imaging. Biochem. J. 400, 531–540 (2006). lines with distinct, defined specificities, it should be possible to 6. Subach, O.M. et al. Conversion of red fluorescent protein into a bright blue probe. Chem. Biol. 15, 1116–1124 (2008). map separate sets of neurons in a single animal. 7. Ai, H.W., Shaner, N.C., Cheng, Z., Tsien, R.Y. & Campbell, R.E. Exploration Finally, we generated Brainbow AAV vectors. These, along of new chromophore structures leads to the identification of improved with recently described Brainbow herpes viral vectors50, may be blue fluorescent proteins. Biochemistry 46, 5904–5910 (2007). 8. Karasawa, S., Araki, T., Nagai, T., Mizuno, H. & Miyawaki, A. Cyan- more useful than Brainbow transgenic mice in some situations. emitting and orange-emitting fluorescent proteins as a donor/acceptor Similar to Autobow, they avoid the need for double-transgenic pair for fluorescence resonance energy transfer. Biochem. J. 381, 307–312 animals. Because the time of infection can be varied, these vectors (2004). provide an alternative to CreER for temporal control. Moreover, 9. Shaner, N.C. et al. Improving the photostability of bright monomeric orange and red fluorescent proteins. Nat. Methods 5, 545–551 (2008). localized delivery of AAVs enables the tracing of connections 10. Shaner, N.C. et al. Improved monomeric red, orange and yellow from known sites to multiple targets and discrimination of long- fluorescent proteins derived from Discosoma sp. red fluorescent protein. distance inputs from local connections. Nat. Biotechnol. 22, 1567–1572 (2004). 11. Shcherbo, D. et al. Far-red fluorescent tags for protein imaging in living The three most broadly useful Brainbow 3 lines (Brainbow 3.0 tissues. Biochem. J. 418, 567–574 (2009). line D, Brainbow 3.1 line 3, Brainbow 3.1 line 18 and Brainbow 3.2 12. Shcherbo, D. et al. Near-infrared fluorescent proteins. Nat. Methods 7, line 7) have been provided to Jackson Laboratories (http://www. 827–829 (2010). jax.org/; stock numbers 21225–21227) for distribution. The two 13. Shaner, N.C., Steinbach, P.A. & Tsien, R.Y. A guide to choosing fluorescent proteins. Nat. Methods 2, 905–909 (2005). AAVs shown in Figure 5 can be obtained from the University 14. Feng, G. et al. Imaging neuronal subsets in transgenic mice expressing of Pennsylvania Vector Core (http://www.med.upenn.edu/gtp/ multiple spectral variants of GFP. Neuron 28, 41–51 (2000). vectorcore/). Plasmids used to generate Brainbow 3.0, 3.1, and 15. Livet, J. et al. Transgenic strategies for combinatorial expression of 3.2, Autobow and Flpbow 1.1 and 3.1 mice are available through fluorescent proteins in the nervous system. Nature 450, 56–62 (2007). 16. Branda, C.S. & Dymecki, S.M. Talking about a revolution: the impact of Addgene (http://www.addgene.org/). site-specific recombinases on genetic analyses in mice. Dev. Cell 6, 7–28 (2004). Methods 17. Snippert, H.J. et al. Intestinal crypt homeostasis results from neutral competition between symmetrically dividing Lgr5 stem cells. Cell 143, © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature Methods and any associated references are available in the online 134–144 (2010). version of the paper. 18. Red-Horse, K., Ueno, H., Weissman, I.L. & Krasnow, M.A. Coronary arteries form by developmental reprogramming of venous cells. Nature 464, 549–553 (2010). npg Accession codes. Addgene plasmids: 45176, 45177 (Brainbow 19. Rinkevich, Y., Lindau, P., Ueno, H., Longaker, M.T. & Weissman, I.L. 3.0), 45178 (Brainbow 3.1), 45179 (Brainbow 3.2), 45182, 45187 Germ-layer and lineage-restricted stem/progenitors regenerate the mouse (Autobow), 45180 (Flpbow 1.1), 45181 (Flpbow 3.1). digit tip. Nature 476, 409–413 (2011). 20. Schepers, A.G. et al. Lineage tracing reveals Lgr5+ stem cell activity in Note: Supplementary information is available in the online version of the paper. mouse intestinal adenomas. Science 337, 730–735 (2012). 21. Tabansky, I. et al. Developmental bias in cleavage-stage mouse Acknowledgments blastomeres. Curr. Biol. 23, 21–31 (2013). This work was supported by grants from the US National Institutes of Health 22. Gupta, V. & Poss, K.D. Clonally dominant cardiomyocytes direct heart (5U24NS063931) and the Gatsby Charitable Foundation and by Collaborative morphogenesis. Nature 484, 479–484 (2012). Innovation Award no. 43667 from the Howard Hughes Medical Institute. We 23. Pan, Y.A., Livet, J., Sanes, J.R., Lichtman, J.W. & Schier, A.F. Multicolor thank S. Haddad for assistance with mouse colony maintenance; X. Duan, Brainbow imaging in zebrafish. Cold Spring Harb. Protoc. 2011, L. Bogart and J. Lefebvre for testing Brainbow mice and AAVs; R.W. Draft for pdb.prot5546 (2011). valuable discussions and advice; R.Y. Tsien (University of California, San Diego) 24. Hampel, S. et al. Drosophila Brainbow: a recombinase-based fluorescence for mOrange2 and TagRFPt; and D.M. Chudakov (Institute of Bioorganic Chemistry labeling technique to subdivide neural expression patterns. Nat. of the Russian Academy of Sciences) for TagBFP, PhiYFP, mKate2 and eqFP650. Methods 8, 253–259 (2011). 25. Hadjieconomou, D. et al. Flybow: genetic multicolor cell labeling for neural circuit analysis in Drosophila melanogaster. Nat. Methods 8, AUTHOR CONTRIBUTIONS 260–266 (2011). D.C., K.B.C. and T.L. performed experiments. D.C., J.W.L. and J.R.S. designed 26. Lang, C., Guo, X., Kerschensteiner, M. & Bareyre, F.M. Single collateral experiments, interpreted results and wrote the manuscript. reconstructions reveal distinct phases of corticospinal remodeling after spinal cord injury. PLoS ONE 7, e30461 (2012). COMPETING FINANCIAL INTERESTS 27. Badaloni, A. et al. Transgenic mice expressing a dual, CRE-inducible The authors declare no competing financial interests. reporter for the analysis of axon guidance and synaptogenesis. Genesis 45, 405–412 (2007). Reprints and permissions information is available online at http://www.nature. 28. Lichtman, J.W., Livet, J. & Sanes, J.R. A technicolour approach to the com/reprints/index.html. connectome. Nat. Rev. Neurosci. 9, 417–422 (2008).

546 | VOL.10 NO.6 | JUNE 2013 | nature methods RESOURCE

29. Lakso, M. et al. Targeted oncogene activation by site-specific 40. Schlake, T. & Bode, J. Use of mutated FLP recognition target (FRT) sites recombination in transgenic mice. Proc. Natl. Acad. Sci. USA 89, for the exchange of expression cassettes at defined chromosomal loci. 6232–6236 (1992). Biochemistry 33, 12746–12751 (1994). 30. Paterna, J.C., Moccetti, T., Mura, A., Feldon, J. & Büeler, H. Influence 41. Peroutka, R.J., Elshourbagy, N., Piech, T. & Butt, T.R. Enhanced protein of promoter and WHV post-transcriptional regulatory element on AAV- expression in mammalian cells using engineered SUMO fusions: secreted mediated transgene expression in the rat brain. Gene Ther. 7, 1304–1311 phospholipase A2. Protein Sci. 17, 1586–1595 (2008). (2000). 42. Farley, F.W., Soriano, P., Steffen, L.S. & Dymecki, S.M. Widespread 31. Madisen, L. et al. A robust and high-throughput Cre reporting and recombinase expression using FLPeR (flipper) mice. Genesis 28, 106–110 characterization system for the whole mouse brain. Nat. Neurosci. 13, (2000). 133–140 (2010). 43. Awatramani, R., Soriano, P., Rodriguez, C., Mai, J.J. & 32. Hippenmeyer, S. et al. A developmental switch in the response of Dymecki, S.M. Cryptic boundaries in roof plate and choroid plexus DRG neurons to ETS transcription factor signaling. PLoS Biol. 3, e159 identified by intersectional gene activation. Nat. Genet. 35, 70–75 (2005). (2003). 33. Srinivas, S. et al. Cre reporter strains produced by targeted insertion of 44. Araki, K., Okada, Y., Araki, M. & Yamamura, K. Comparative analysis of EYFP and ECFP into the ROSA26 locus. BMC Dev. Biol. 1, 4 (2001). right element mutant loxP sites on recombination efficiency in embryonic 34. Guo, C., Yang, W. & Lobe, C.G. A Cre recombinase transgene with mosaic, stem cells. BMC Biotechnol. 10, 29 (2010). widespread tamoxifen-inducible action. Genesis 32, 8–18 (2002). 45. Entenberg, D. et al. Setup and use of a two-laser multiphoton microscope 35. Zhang, X.M. et al. Highly restricted expression of Cre recombinase in for multichannel intravital fluorescence imaging. Nat. Protoc. 6, cerebellar Purkinje cells. Genesis 40, 45–51 (2004). 1500–1520 (2011). 36. Rossi, J. et al. Melanocortin-4 receptors expressed by cholinergic neurons 46. Mahou, P. et al. Multicolor two-photon tissue imaging by wavelength regulate energy balance and glucose homeostasis. Cell Metab. 13, 195–204 mixing. Nat. Methods 9, 815–818 (2012). (2011). 47. Wang, K. et al. Three-color femtosecond source for simultaneous 37. Campsall, K.D., Mazerolle, C.J., De Repentingy, Y., Kothary, R. & excitation of three fluorescent proteins in two-photon fluorescence Wallace, V.A. Characterization of transgene expression and Cre recombinase microscopy. Biomed. Opt. Express 3, 1972–1977 (2012). activity in a panel of Thy-1 promoter-Cre transgenic mice. Dev. Dyn. 224, 48. Conchello, J.A. & Lichtman, J.W. Optical sectioning microscopy. 135–143 (2002). Nat. Methods 2, 920–931 (2005). 38. Bunting, M., Bernstein, K.E., Greer, J.M., Capecchi, M.R. & Thomas, K.R. 49. Ducros, M. et al. Efficient large core fiber-based detection for multi- Targeting genes for self-excision in the germ line. Genes Dev. 13, channel two-photon fluorescence microscopy and spectral unmixing. 1524–1528 (1999). J. Neurosci. Methods 198, 172–180 (2011). 39. McLeod, M., Craft, S. & Broach, J.R. Identification of the crossover site 50. Card, J.P. et al. A dual infection pseudorabies virus conditional reporter during FLP-mediated recombination in the Saccharomyces cerevisiae approach to identify projections to collateralized neurons in complex plasmid 2 µm circle. Mol. Cell Biol. 6, 3357–3367 (1986). neural circuits. PLoS ONE 6, e21141 (2011). © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature npg

nature methods | VOL.10 NO.6 | JUNE 2013 | 547 ONLINE METHODS in bacteria, purified using His Fusion Protein Purification Kits Brainbow constructs. cDNA encoding the following fluorescent (Thermo Scientific), concentrated to >3 mg/ml, and used as proteins were used: EGFP1, EYFP2, ECFP2, mCerulean3, PhiYFP4 immunogens to produce rat anti-mTFP, chicken anti-EGFP, rat (Evrogen), mTFP5 (Allele Biotechnology), TagBFP6 (Evrogen), anti-PhiYFP, rabbit anti-mCherry and guinea pig anti-mKate2. EBFP2 (ref. 7) (Addgene), Kusabira-Orange8, TagRFPt9, mOr- (Covance). Chicken anti-EGFP IgY was purified from chicken ange2 (ref. 9), tdTomato10, mCherry10, mKate2 (ref. 11) (Evrogen) egg yolks using Pierce Chicken IgY Purification Kit (Thermo and eqFP650 (ref. 12) (Evrogen). A HRAS farnesylation sequence51 Scientific). Other sera were used without purification. Other was used to tether XFPs to the cell membrane. A nuclear locali- antibodies used were: rabbit anti-GFP (ab6556, Abcam), rabbit zation signal (NLS, APKKKRKV) was added to the N-terminal anti-PhiYFP (AB604, Evrogen), chicken anti-SUMOstar (AB7002, end of PhiYFP(Y65A). WPRE sequence was used to stabilize LifeSensors), DyLight 405–conjugated goat anti-rat (Jackson mRNA and enhance nuclear mRNA export30,31. Polyadenylation ImmunoResearch) and Alexa fluorescent dye–conjugated goat signals were from the SV40 T antigen for mouse transgenes and anti-rat 488, anti-chicken 488 and 594, anti-rabbit 514 and 546, from the human growth hormone gene for AAV. Brainbow con- and anti–guinea pig 647 (Life Technologies). structs were assembled by standard cloning methods. A cloning scaffold containing concatenated loxP mutant sequences and Histology. Mice were anesthetized with sodium pentobarbital unique restriction digestion sites was synthesized (DNA2.0, Inc.) before intracardiac perfusion with 2%–4% paraformaldehyde to facilitate cloning. in PBS. Brains were sectioned at 100 µm using a Leica vt1000s Brainbow modules were cloned into the pCMV- N1 mamma- vibratome. Muscle and retina were sectioned at 20 µm in a Leica lian expression vector (Clontech) for transient mammalian cell CM1850 cryostat or processed as whole mounts. For immunos- expression. Brainbow mouse constructs were cloned into a unique taining, tissues were permeabilized by 0.5% Triton X-100 with XhoI site in a genomic fragment of Thy1.2 containing neuron- 0.02% sodium azide in StartingBloc (Thermo Scientific) at room specific regulatory elements14. Brainbow AAV constructs were temperature for 2 h and then incubated with combinations of cloned into vectors provided by the University of Pennsylvania anti-XFPs (see above) for 24–48 h at 4 °C. After extensive washing Virus Core. Constructs were tested by expression in HEK293 cells in PBST (0.01 M PBS with 0.1% Triton X-100), all secondary anti- (ATCC) before generation of mice or AAV. bodies (1:500) were added for 12 h at 4 °C. Finally, sections and tissues were mounted in Vectashield mounting medium (Vector Mice. Transgenic mice were generated by pronuclear injection at the Labs) and stored at –20 °C until they were imaged. Harvard Genome Modification Core. Mice were maintained on C57B6 Antibody combinations used in figures are as follows. In or CD-1 backgrounds. Brainbow mice were crossed to mice that Figures 1e–h; 2b,d–g; and 4c and Supplementary Figures 2a–d, expressed Cre or Flp recombinases, including PV-Cre32, Islet-Cre33, 4a and 5, primary antibodies are chicken anti-GFP (1:2,000), CAGGS-CreER34, L7-Cre35, ChAT-Cre36, Thy1-Cre37, Wnt-Flp43 and rabbit anti-mCherry (1:1,000, for mOrange2) and guinea pig Actin-Flp42. Both male and female mice were used. All experiments anti-mKate2 (1:500). Secondary antibodies are Alexa dye– conformed to NIH guidelines and were carried out in accordance with conjugated goat anti-chicken 488, anti-rabbit 546, and anti–guinea protocols approved by the Harvard University Standing Committee pig 647. In Figure 3b–d and Supplementary Figure 2e, primary

© 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature on the Use of Animals in Research and Teaching. antibodies are chicken anti-GFP (for ECFP), rabbit anti-PhiYFP (1:1,000) and guinea pig anti-mKate2. Secondary antibodies are AAV injection. Two Brainbow AAVs were mixed to equal titer Alexa dye–conjugated goat anti-chicken 488, anti-rabbit 546 and 12

npg (7.5 × 10 genome copies per ml) before injection. For retina anti-guinea pig 647. In Figure 4b, rabbit anti-PhiYFP and Alexa injection, adult mice were anesthetized with ketamine-xylazine by dye–conjugated goat anti-rabbit 514 were used. In Figure 5d, intraperitoneal injection. A small hole was made in the temporal primary antibodies are rat anti-mTFP and rabbit anti-mCherry. eye by puncturing the sclera below the cornea with a 30 1/2–G Secondary antibodies are Alexa dye–conjugated goat anti-rat 488 needle. With a Hamilton syringe with a 33-G blunt-ended needle, and anti-rabbit 546. In Figure 5f–j and Supplementary Figure 9, 0.5–1 µl of AAV virus was injected intravitreally. After injections, primary antibodies are guinea pig anti-mKate2 (for TagBFP), rat animals were treated with Antisedan (Zoetis) and monitored for anti-mTFP (1:1,000), chicken anti-GFP (for EYFP) and rabbit full recovery. For cortex injection, adult mice were anesthetized anti-mCherry. Secondary antibodies are Alexa dye–conjugated with isoflurane via continuous delivery through a nose cone and goat anti-rat 488, anti-chicken 488, anti-rabbit5 46 and anti- fixed to a stereotaxic apparatus. Surgery took place under sterile guinea pig 647. In Supplementary Figure 1, primary antibodies conditions with the animal lying on a heating pad. One microliter are rat anti-PhiYFP (1:1,000) and rabbit anti-mCherry (1:1,000, of 1:5 saline-diluted AAV mix (1.5 × 1012 genome copies per ml) for mOrange2). Secondary antibodies are Alexa dye–conjugated was injected over 10 min. The head wound was sutured at the goat anti-rat 488 and anti-rabbit 546. In Supplementary Figure 7, end of the experiment. One injection of the nonsteroidal anti- primary antibodies are rabbit anti-EGFP (1:1,000, for ECFP) and inflammatory agent meloxicam was given at the end of the chicken anti-SUMOstar (1:1,000). Secondary antibodies are Alexa surgery, and mice were kept on a heating pad with accessible dye–conjugated goat anti-rabbi 514 and anti-chicken 594. moistened food pellets and/or HydroGel until they had fully recovered. The mice were given another dose of meloxicam 1 d Imaging. Fixed brain and muscle samples were imaged using a later and examined 4–6 weeks after infection. Zeiss LSM710 confocal microscope. Best separation of multiple fluorophores was obtained by using a 405-nm photodiode laser Antibodies. Expression vectors were constructed to produce His for TagBFP and DyLight 405, a 440-nm photodiode laser for tag fusions of XFPs in Escherichia coli. Proteins were produced mTFP, a 488-nm Argon line for EGFP and Alexa 488, a 514-nm

nature methods doi:10.1038/nmeth.2450 Argon line for EYFP and Alexa 514, a 561-nm photodiode for collected at 545–600 nm in channel 2. In the antibody-amplified mOrange2 and Alexa 546, a 594-nm photodiode for mCherry, samples, conjugated Alexa dyes normally produced much stronger mKate2 and Alexa 594 or a 633-nm photodiode for Alexa 647. fluorescence signal than XFPs. The Zeiss microscope we used was Images were obtained with 16× (0.8 NA), and 63× (1.45 NA) oil optimized for imaging the Alexa 488/546/647 combination. The objectives. Confocal image stacks for all channels were acquired fixed dichroic mirror was DM488/561/633. Alexa 488 and Alexa sequentially, and maximally or 3D-view projected using ImageJ 647 were excited by 488-nm and 633-nm lasers simultaneously, (NIH). Intensity levels were uniformly adjusted in ImageJ. and fluorescence was collected at 495–590 nm in channel 1 and Optimal imaging for Brainbow 3 tissue used a Zeiss LSM710 638–780 nm in channel 2, respectively. In the subsequent scan, with fixed dichroic mirror combinations of DM455+514/594 a 561-nm laser was used to excite mOrange2, and fluorescence to reduce lag time between the two sequential scans. EGFP and was collected at 566–626 nm in channel 2. mKate2 were excited by 458-nm and 594-nm lasers simultane- ously, and fluorescence was collected at 465–580 nm in channel 1 51. Hancock, J.F., Cadwallader, K., Paterson, H. & Marshall, C.J.A. CAAX or a and 605–780 nm in channel 2, respectively. In a subsequent scan, a CAAL motif and a second signal are sufficient for plasma membrane 514-nm laser was used to excite mOrange2, and fluorescence was targeting of ras proteins. EMBO J. 10, 4033–4039 (1991). © 2013 Nature America, Inc. All rights reserved. America, Inc. © 2013 Nature npg

doi:10.1038/nmeth.2450 nature methods

NEW TOOLS FOR THE BRAINBOW TOOLBOX Dawen Cai, Kimberly B. Cohen, Tuanlian Luo, Jeff W. Lichtman and Joshua R. Sanes

Supplementary Figures and Tables

Contents: • Supplementary Figures 1 – 11 • Supplementary Tables 1 – 2

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 1. Use of the non-fluorescent default protein to analyze expression potential of a Brainbow3.1 mouse line.

Sagittal section of brain from Brainbow 3.1 (line 3), stained with rat anti-PhiYFP. (a) Low power image. C, cerebellum; H, Hippocampus. (b, c) Higher magnification view of cerebellar (C) and hippocampal (H) regions boxed in a. Bars are 1mm in a, 200μm in b, 20μm in c.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 2. Neuronal labeling in Brainbow3;cre double transgenic and Autobow mice.

(a) Whole mount of retina from Islet1-Cre;Brainbow 3.1 (line 10) showing retinal ganglion cells. (b,c) Cerebellum from Brainbow 3.1 (line 3); parvalbumin-cre mouse. Low power image in b shows Purkinje (yellow arrows) and basket cells (cyan arrow). Axon terminals of basket cells arborize on initial segments of Purkinje cell axons. Higher power view in c shows axons from multiple basket cells, labeled in distinct colors, wrapping a Purkinje cell body and axonal initial segment (yellow arrow). (d) Brainstem from PV-Cre;Brainbow 3.1 (line 4) showing axonal terminals of inhibitory neurons. In a-d, antibody amplified EGFP, mOrange2 and mKate2 are in blue, green and red, respectively. (e) Labeling of hippocampal neurons by Autobow (line 2). Antibody amplified Cerulean, PhiYFP and mKate2 are in blue, green and red, respectively. Bars are 30μm in a, 20μm in b, 10μm in c and d, 100μm in e.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 3. Farnesylated XFPs label neurons more evenly than cytoplasmic XFPs.

Retinal bipolar cells are labeled by (a) cytoplasmic GFP and (b) farnesylated membrane GFP. To image processes of neurons labeled with cytoplasmic GFP, it is necessary to saturate the somata. Bars are 5µm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 4. Immunostaining enhances Brainbow signals without compromising color diversity.

Motor axons innervating extraocular muscles of a Brainbow 3.0 mouse (line D). a shows native XFP fluorescence. b shows a nearby section immunostained with non-crossreactive chicken-anti-GFP, rabbit-anti-mOrange2 and guinea pig-anti-mKate2. Bars are 50μm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 5. XFP expression balance in Brainbow 3.

Immunostained cerebellum section from Brainbow 3.1 (line 3);PV-cre mouse. Merged image is shown in a and separate channels in b (EGFP), c (mOrange2) and d (mKate2). (e) Percent of Purkinje cells expressing each XFP (208 neurons from 12 regions). Bars are 20μm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 6. Tests for incompatible mutant Frt sites.

Canonical Frt site and multiple Frt variants have been reported to allow Flp-dependent recombination between the same variants but not between each other17, 18. To verify incompatibility, three constructs were generated and tested: (a) Construct testing incompatibility between Frt5T2 and Frt545 sites. When co-transfected with this construct and FLP recombinase, HEK293 cells showed high level of Kusabira-Orange but no detectable mKate2. This indicated that Frt5T2 and Frt545 are incompatible. (b) Adding FrtF3 site and TagBFP to the construct in a allowed a three-way test of FrtF3, Frt5T2 and Frt545. Co- transfection of this construct with Flp recombinase showed TagBFP expression only, indicating that the FrtF3 site is incompatible with both Frt5T2 and Frt545. (c) Adding the canonical Frt site and TagBFP to the construct in a allowed a three-way test of Frt, Frt5T2 and Frt545. Co-transfection of this construct with Flp recombinase showed high level of TagBFP expression as well as detectable levels of Kusabira-Orange and mkate2. This indicates that canonical Frt site undergoes Flp dependent excision with Frt5T2 and Frt545 sites. Bars are 100μm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 7. SUMOstar as epitope tag in transgenic mouse.

Section from the hippocampus from a transgenic mouse that expressed SUMOstar-ECFP fusion protein from the Thy1 promoter. The section was stained with rabbit anti-GFP and chicken anti-SUMOstar, plus appropriate secondary antibodies. (a) Native fluorescence of ECFP. (b) Rabbit anti-GFP (c) Chicken anti-SUMOstar. Bar is 50μm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 8. Tests of Brainbow AAV constructs.

In order to accommodate the 5kB packaging size of AAV yet achieve high expression level, several promoter elements, polyadenylation sequences and stop mechanisms were tested in constructs based on Brainbow1. (a) This series tested the performance of a 1kB hybrid promoter CB6 (CMV enhancer, chicken β-actin promoter and hybrid intron, Penn Vector Core) as well as a small stop cassette relying on an open reading frame shift mechanism (“ORF shift” round corner empty box). The ORF shift encodes a Kozak sequence and translational start (red solid arrow) to initiate an “out of frame” translation of the following mTFP. Upon Cre recombination, this stop cassette is removed and mTFP, PhiYFP or TagRFPt is expressed. AAV made from this construct displayed cre-independent expression of XFPs in wildtype mouse brains. Moreover, expression from the CB6 promoter was weaker than that from a EF1α promoter, tested in parallel. (b) This series used a pair LoxP left and right element mutants (LE-RE) to construct a Cre dependent “unidirectional spin” cassette. A Brainbow1 construct expressing Kusabira-Orange, EGFP and mCherry was placed in inverted orientation between the LE-RE loxP sites. AAV generated from this construct exhibited no Cre- independent expression in mouse brain. However, the proportions of the three XFPs expressed following infection of Cre transgenic mice was greatly influenced by the level of Cre expression. When Cre levels were low, Kusabira-Orange was expressed in most cells. In contrast, when Cre levels were high, very little expression of Kusabira-Orange was detected, with most cells expressing mCherry and a minority expressing EGFP. This imbalance led us to use the LE-RE elements exclusively (Fig. 2). (c) This series incorporated a STOP cassette generated from a tandem Myc epitope tag and a bovine growth hormone polyadenylation sequence, totaling 0.6kB. To accommodate the packaging size of AAV, only two XFPs were incorporated into each construct. Infection with pairs of AAVs led to multicolor labeling (see d). As in constructs shown in b, however, relative expression of the two XFPs in each vector varied among Cre lines. Based on these tests, we designed new AAVs (Fig. 5a) incorporating EF1α promoter, LE-RE mutant sites, farnesylated XFPs, a single WPRE sequence, and a 0.5kB human growth hormone polyadenylation sequence. (d) Brain from a Thy1-cre mice infected with the two AAV shown in c. Native fluorescence of TagBFP in blue, Kusabira-Orange and EGFP in green, mCherry in red. Bar is 50μm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 9. Labeling color pallet changes at different distances to Brainbow AAV injection site.

(a) Sagittal section of PV-Cre mouse brain shows fluorescent labeling by Brainbow AAV infection. White arrow indicates injection site. (b) Magnified view of blue boxed regions in a. (c, d) Magnified views of regions boxed in b. Bars are 300μm in a, 50μm in b and 20μm in c, and d.

Nature Methods: doi:10.1038/nmeth.2450

Supplementary Figure 10. Chromatic aberration in Brainbow images.

(a,b) X-Y plane (a) and cross sections (b) show axial and lateral chromatic aberration, as well as laser misalignment cause color offset in X-Y-Z directions. (c,d) Color-shift correction of images in a,b align the channels and restore correct colors. Arrowheads indicate corresponding objects in the original and corrected images. Bars are 3µm.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Figure 11. Noise reduction methods comparison.

(a) Original image. (b) Enlarged view of region boxed in a. (c) Guassian filter with radius of 1.0 pixel. (d) Median filter with radius of 1.0 pixel. (e) Deconvolution by Huygens software. Bars are 10µm in a, 2µm in b-e.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Table 1. Fluorescent proteins screened in transgenic mice.

Brightness Fluorescent Protein Ex (nm) Em (nm) in vivo property species Antibody (EGFP=1) EBFP2 383 448 0.55 Dim; Scattering Aequorea victoria Ck : GFP TagBFP 399 456 1.00 Bright; Scattering Entacmaea quadricolor GP : mKate2 ECFP 433 475 0.33 Dim; High Expression Aequorea victoria Ck : GFP Cerulean 433 475 0.82 Bright; Photobleaching Aequorea victoria Ck : GFP mTFP1 462 492 1.63 Bright Clavularia sp. Rt : TFP EGFP 488 507 1.00 Bright; High Expression Aequorea victoria Ck : GFP EYFP 516 529 1.48 Bright; High Expression Aequorea victoria Ck : GFP PhiYFP * 525 537 1.58 Bright; PFA Quenching Phialidium sp. Rt : PhiYFP Kusabira Orange 548 559 0.94 Bright; Aggregation Fugia concinna Rb : KuO mOrange2 549 565 1.06 Bright; Aggregation Discosoma sp. Rb : DsRed tdTomato 554 581 2.88 Bright Discosoma sp. Rb : DsRed tagRFP-T 555 584 1.00 Bright Entacmaea quadricolor GP : mKate2 mCherry 587 619 0.48 Dim; Aggregation Discosoma sp. Rb : DsRed mKate2 588 635 0.76 Bright Entacmaea quadricolor GP : mKate2 eqFP650 592 650 0.48 Dim; Far-red Entacmaea quadricolor GP : mKate2

Columns are as follows: (1) Proteins tested; see refs. 2-12. (2) Ex, excitation peak. (3) Em, emission peak. (4) Brightness calculated as the product of extinction coefficient and quantum yield, with EGFP=1. (5) Global result from observation of brain tissue in transgenic mice. Main drawbacks encountered were low fluorescence in vivo despite high intrinsic brightness, presumable because of low expression; high scattering of deep blue proteins; rapid photobleaching; aggregation in neurons; and sensitivity to paraformaldehyde fixation. (6) Species from which XFP was derived. (7) Antibodies used to amplify signal: Ck:GFP, chicken-anti-GFP; GP:mKate2, Guinea pig-anti-mKate2; Rt:TFP, rat-anti-mTFP1; Rt:PhiYFP, rat-anti-PhiYFP; Rb:KuO, rabbit-anti-Kusabira Orange; Rb:DsRed, rabbit-anti-DsRed.

* The nonfluorescent mutant PhiYFP(Y65A) was used as stainable “STOP”.

Nature Methods: doi:10.1038/nmeth.2450 Supplementary Table 2. Brainbow 3 transgenic mouse lines. Expression Cell type labeled (Brightness*) st nd rd th # line subset size 1 /2 /3 /4 Spinal cord / Peripheral Central Nervous System A medium not tested RGC(+), CTL(++), HPL(++), CBR(+) B large not labeled whole brain (+) Brainbow-3.0 MN (+++), SIN (++), Of/Gf/Kf D large RGC (++), RBP (++), CTL(+++), HPL(+++) DRG (++), SSN (++) F medium MN(++), DRG(+++) CTL(++), HPL(++) 2 large not tested whole brain (+) 3 medium MN (++) CTL (++), HPL (+++), CBR (++), CPC (+++) 4 large MN (+), DRG (++) CTL(+), HPL(+++), CBR (++), HB (+) 10 small not labeled RGC (++), RAC (++), CTL (+), HPL (++) Brainbow-3.1 11 small MN (+), DRG (++) RGC (++), RAC (++), CTL(++), HPL(+++) nØ/Of/Gf/Kf 16 small not labeled CP (++), TH (++) 17 very small not labeled CTL(+++), HPL(+++), CP (+++) RGC (++), RAC (++), RBC (++), RHC (++), 18 medium not tested CTL(+++), CBR (+++), MB (+++), HB (+++) 21 small not labeled CTL (+), HPL (+) 23 medium MN (+), DRG (+) HPL (++), MF (+) 2 small NM (+++), DRG (+++) CTL (+++),HPL (+++) RGC (+++), RAC (+++), 4 small not labeled CTL (++++),HPL (++++) Brainbow-3.2 5 large not tested whole brain (+) nØ/Ofw/Gfw/Kfw 6 medium NM (+++), DRG (++) CTL (+++),HPL (+++) RGC(+++), RAC(+++), RBP (+++), 7 medium SIN (++) CTL (++++), HPL (++++), CGC (++++) 1 Large NM (++), DRG (+) CTL (++), HPL (++) Autobow 2 very small not labeled CTL (++), HPL (++) CreINT/Cer/P/K 3 medium NM(+) CTL (+), HPL (+) 1 medium NM (+), DRG (+) CTL (+), HPL (+) 2 large not tested whole brain (+) 3 large DRG (+) RGC (++), whole brain (+) Flpbow-3.0 5 very small not labeled RGC (+) T/P/C RGC (++), CTL (++), HPL (++), 6 medium NM (++), SIN (++) CP (++), MB (++), HB (++) 7 large NM (++) CTL (++), HPL (++), MB (++), HB (++) 4 large not tested MF (+), HB (+) Flpbow-3.1 5 small not labeled RGC (+), CTL (+), HPL (+), MF (+) nØ/sOf/sGf/sKf 6 large MN (+), RDG (+) whole brain (+) A total of 46 Brainbow 3 transgenic mouse lines were generated. 15 lines, not described here died, had, no detectable XFP expression, or did not transmit the Brainbow transgene to their progeny. Lines available from Jackson Laboratory are indicated in bold. f, farnesylated XFP; n, nucleic localization sequence; w, WPRE; s, SUMO tag; CreINT, intron inserted Cre; O, monomeric Orange2; G, EGFP; K, monomeric Kate2; P, PhiYFP; Ø, “dark” PhiYFP; T, tandem-dimer Tomato; Cer, Cerulean; C, ECFP; MN, motor neuron; SIN, spinal cord interneuron; SSN, skin sensory neuron; DRG, dorsal root ganglion; RGC, retina ganglion cell; RAC, retina amacrine cell; RBP, retina bipolar cell; RHC, retina horizontal cell; CBR, cerebellar neuron; CPC, cerebellar Purkinje cell; CGC, cerebellar granule cell; CTL, cortical neuron; HPL, hippocampal neuron; CP, caudate putamen; TH, thalamus; MB, mid brain; HB, hindbrain; MF, Mossy fiber. * Specific antibodies effectively enhanced the brightness of all XFPs in all lines. § Autobow line 1 and 2, founders were screened; line 3, founder as well as the first and the 6th generations were screened.

Nature Methods: doi:10.1038/nmeth.2450