NEUROSCIENCE

Building a vast digital simulation of the brain could transform and medicine and reveal new ways of making more powerful computers By

IT’S TIME TO CHANGE THE WAY WE STUDY THE BRAIN. Reductionist biology—examining to investigate the parts and then reas- individual brain parts, neural circuits semble them to re-create the whole. and molecules—has brought us a long Putting things together to devise a way, but it alone cannot explain the complete simulation of the human workings of the human brain, an in- brain is the goal of an undertaking formation processor within our skull that intends to construct a fantastic

that is perhaps unparalleled any- new scientific instrument. Nothing ) where in the universe. We must con- quite like it exists yet, but we have be- struct as well as reduce and build as gun building it. One way to think of well as dissect. To do that, we need a this instrument is as the most power-

new paradigm that combines both ful flight simulator ever built—only analysis and synthesis. The father of rather than simulating flight through AND www.humanconnectomeproject.org reductionism, French philosopher open air, it will simulate a voyage ( René Descartes, wrote about the need through the brain. This “virtual brain”

IN BRIEF Imaging Neuro of Laboratory

Computer simulation will intro­ By the year 2020 digital brains A sim brain can act as a stand-in duce ever greater verisimilitude in­ may be able to represent the inner for the genuine article, thus foster­ Imaging Biomedical for Center Martinos to digital depictions of the work­ workings of a single brain cell or ing a new understanding of autism ings of the human brain. even the whole brain. or per­mitting virtual drug trials. COURTESY OF ARTHUR W. TOGA TOGA W. ARTHUR OF COURTESY BUCKNER RANDY

50 Scientific American, June 2012 © 2012 Scientific American June 2012, ScientificAmerican.com 51 © 2012 Scientific American will run on supercomputers and incorpo­ Henry Markram directs the at the Swiss Federal rate all the data that neuroscience has Institute of Technology in Lausanne. He has done extensive work on how neurons interconnect, communicate and learn. He also discovered generated to date. fundamental principles of brain plasticity and is co-discoverer of the A digital brain will be a resource for the intense world theory of autism and the theory of how the brain entire scientific community: researchers computes as a liquid that is constantly perturbed. will reserve time on it, as they do on the biggest telescopes, to conduct their experi­ ments. They will use it to test theories of at them simply as sets of symptoms. The ing to—at least, not all of it. We intend to how the human brain works in health and breadth of this perspective would allow reproduce the myriad connections among in disease. They will recruit it to help them us to move forward to develop a genera­ brain cells by different means. develop not only new diagnostic tests for tion of treatments that selectively target The key to our approach is to craft the autism or schizophrenia but also new ther­ the underlying abnormalities. basic blueprint according to which the apies for depression and Alzheimer’s dis­ The second reason is that computing is brain is built: the set of rules that has guid­ ease. The wiring plan for tens of trillions of fast approaching barriers to further devel­ ed its construction over evolution and neural circuits will inspire the design of opment. Computers cannot do many tasks does so anew in each developing fetus. In brainlike computers and intelligent robots. that animal brains do effortlessly, despite theory, those rules are all the information In short, it will transform neuroscience, the inexorable increase in processing pow­ we need to start building a brain. The medicine and information technology. er. For instance, although computer scien­ skeptics are right: the complexity they tists have made huge progress in visual generate is daunting—hence our need for BRAIN IN A BOX recognition, the machines still struggle to supercomputers to capture it. But unravel­ scientists could be running the first simu­ make use of context in a scene or to use ar­ ing the rules themselves is a far more trac­ lations of the human brain by the end of bitrary scraps of information to predict fu­ table problem. If we pull it off, there is no this decade, when supercomputers will be ture events in the way the brain can. logical reason why we cannot apply the powerful enough to support the massive Moreover, because more powerful com­ blueprint in the same way that biology number of calculations needed. The in­ puters require more energy, supplying their does and build an “in silico” brain. strument will not require that all myster­ needs will one day no longer be feasible. The kind of rules we are talking about ies of the brain be unraveled first. Instead The performance of today’s supercomput­ are ones that govern the genes that lead to it will furnish a framework to accommo­ ers is measured in petaflops—quadrillions the types of cells there are in the brain and date what we do know, while enabling us of logic operations per second. The next the underlying plan for the way those cells to make predictions about what we do not. generation, due around 2020, will be 1,000 are distributed and how they are connect­ Those predictions will show us where to times faster and will be measured in exa­ ed. We know that such rules exist because target our future experiments to prevent flops—quintillions of operations per sec­ we discovered some of them while laying wasted effort. The knowledge we generate ond. By itself, the first exa-scale machine the groundwork for the HBP. We started will be integrated with existing knowl­ will probably consume around 20 mega­ doing that almost 20 years ago by measur­ edge, and the “holes” in the framework watts, roughly the energy requirement of a ing the characteristics of individual neu­ will be filled in with increasingly realistic small town in winter. To create increasing­ rons. We collected vast amounts of data detail until, eventually, we will have a uni­ ly powerful computers that perform some about the geometric properties of different fied working model of the brain—one that of the simple but useful things that hu­ neuronal types and digitally reconstructed reproduces it accurately from the whole mans are capable of in an energy-efficient hundreds of them in three dimensions. brain down to the level of molecules. way, we need a radically new strategy. Using a painstaking method called patch Building this instrument is the goal of We could do worse than take inspira­ clamping, which involves placing the tip the (HBP), an initia­ tion from the human brain, which per­ of a microscopic glass pipette up against tive involving about 130 universities forms a range of intelligent functions on a a cell membrane to measure the voltage around the world. The HBP is one of six mere 20 or so watts—a million times fewer across its ion channels, we also recorded projects competing for a glittering prize, than an exa-scale machine and equivalent the neurons’ electrical properties. up to €1 billion to be provided over 10 to a weak lightbulb. For that, we need to In 2005 modeling a single neuron took years by the to each of understand the multilevel organization of a powerful computer and a three-year two winners, who will be announced in the brain, from genes to behavior. The Ph.D. project. It was clear that more ambi­ February 2013. knowledge is out there, but we need to tious goals would soon become achiev­ We need the simulator for at least two bring it together—and our instrument will able, however, and that we could model reasons. In Europe alone, brain diseases provide the platform on which to do that. larger elements of brain circuitry even if affect 180 million people, or roughly one Critics say that the goal of modeling our knowledge of those elements was in­ in three—a number that is set to grow as the human brain is unachievable. One of complete. At the Brain Mind Institute at the population ages. At the same time, their principal objections is that it is im­ the Swiss Federal Institute of Technology pharmaceutical­ companies are not invest­ possible to reproduce the connectivity in Lausanne, we launched one of the ing in new treatments for the ailing ner­ among the brain’s 100 trillion synapses be­ HBP’s predecessors, the Blue Brain Proj­ vous system. A holistic view of the brain cause we cannot measure it. They are cor­ ect. We would build what we call “unifying would enable us to reclassify such diseas­ rect that we cannot measure the web of computer models”—models that integrate es in biological terms rather than looking connections, which is why we are not go­ all existing data and hypotheses about a

52 Scientific American, June 2012 © 2012 Scientific American given brain circuit, while reconciling con­ watched, the neurons began to speak to equations on supercomputers. We also flicts in that information and highlighting one another. “Spikes,” or action poten­ need to create software that will construct where knowledge is lacking. tials—the language of the brain—spread a brain that conforms to the inherent biol­ through the column as it began to work as ogy. We call it the “brain builder.” SYNTHESIS BIOLOGY an integrated circuit. The spikes flowed be­ The predictions of how the brain oper­ as a test case, we set out to build a unify­ tween the layers and oscillated back and ates offered up by —and ing model of a brain structure called the forth, just as they do in living brain slices. refined by new data—will accelerate our cortical column. The column is the equiva­ This was behavior we had not programmed understanding of brain function without lent of a processor in your laptop. To use a into the model; it emerged spontaneously measuring every aspect of it. We can make crude metaphor, if you were to put a min­ because of the way the circuit was built. predictions based on the rules we are un­ iature apple corer through the cortex and And the circuit stayed active even after the covering and then test those predictions pull out a cylinder of tissue about half a stimulation had stopped and briefly devel­ against reality. One of our current goals is millimeter in diameter and 1.5 mm in oped its own internal dynamics, its own to use knowledge of genes that give rise to height, that would be a column. Within way of representing information. the proteins for certain types of neurons to that tissue core, you would find a very Since then, we have been gradually in­ predict the structure and behavior of dense network consisting of a few tens of tegrating more of the information generat­ those cells. The link between genes and ac­ thousands of cells. The column is such an ed by laboratories around the world into tual neurons constitutes what we call an efficient design for an information-pro­ this unifying model of the column. The “informatics bridge,” the kind of shortcut cessing element that once evolution had software we have developed is also being that synthesis biology offers us. hit on the formula, it kept applying this refined continuously so that each week we Another kind of informatics bridge recipe again and again until no more rebuild the column, we do so with more that scientists have made use of for years space was left in the skull and the cortex data, more rules and more accuracy. The has to do with genetic mutations and had to fold in on itself to create more next step is to integrate data for an entire their link to disease: specifically, how room—hence, your convoluted brain. brain region and then for an entire brain— mutation changes the proteins that cells The column penetrates the six vertical to begin with, a rodent brain. manufacture, which in turn affect the ge­ layers of the neocortex, the cortex’s outer Our effort will depend heavily on a dis­ ometry and electrical characteristics of layer, and the neural connections between cipline called neuroinformatics. Vast quan­ neurons, the synapses they form and the it and the rest of the brain are organized tities of brain-related data from all over the electrical activity that emerges locally, differently in each layer. The organization world need to be brought together in a co­ in microcircuits, before spreading in a of these connections resembles the way herent way, then mined for patterns or wide swath across whole brain regions. telephone calls are assigned a numerical rules that describe how the brain is orga­ In theory, for example, we could pro­ address and routed through an exchange. nized. We need to capture the biological gram a certain mutation into the model A few hundred neuron types reside in a processes those rules describe in sets of and then observe how that mutation af­ column, and using our IBM Blue Gene su­ mathematical equations, while developing fects it at each step along the biological percomputer, we integrated all the avail­ the software that will enable us to solve the chain. If the resulting symptom, or con­ able information about how those types mix in each layer until we had a “recipe” for a column in a newborn rat. We also in­ POWER OF THE EXAFLOP structed the computer to allow the virtual neurons to connect in all the ways that More Computer = More Brain real neurons do—but only in those ways. It took us three years to build the software The ability to simulate the brain in enough detail to carry out vital scientific research will grow facility that, in turn, allowed us to con­ with computer power. A digital facsimile of a cylindrical piece of tissue in the rat cortex became struct this first unifying model of a col­ a reality in 2008, when speed was clocked in teraflops. As computers climb to the peta and exa umn. And with it we had our proof of con­ scales, the Human Brain Project envisages full-brain simulations of a mouse and of the same cept of what we call synthesis biology—a species that conceived Hamlet and Einstein’s general theory of relativity. simulation of the brain from the full diver­

VOL. 482; FEBRUARY 23, 2012 23, 482; FEBRUARY VOL. Computing Memory sity of biological knowledge—and how it 2011 Exabyte (1018 bytes) 2008 Cortical NATURE, can serve as both a feasible and an inven­ Column in mesocircuit tive new way of doing research. neocortex (100 neocortical 15 At that point, we had a static model— Petabyte (10 bytes) (10,000 columns) 2023 the equivalent of a column in a comatose 2005 neurons) Complete 12 Terabyte (10 bytes) Single- human brain brain. We wanted to know whether it 2014 (1,000 times would start to behave like a real column, neuron Complete Gigabyte (109 bytes) model rodent brain) albeit one isolated from the rest of the rodent brain (100 mesocircuits) brain in a slice of living brain tissue, so Megabyte (106 bytes) we gave it a jolt—some external stimula­ Gigaflop (109 flops) Teraflop (1012 flops) Petaflop (1015 flops) Exaflop (1018 flops) tion. In 2008 we applied a simulated elec­ Computing Speed (flops)

SOURCE: “BRAIN IN A BOX,” BY M. MITCHELL WALDROP, IN WALDROP, MITCHELL M. BY BOX,” A “BRAIN IN SOURCE: trical pulse to our virtual column. As we

June 2012, ScientificAmerican.com 53 © 2012 Scientific American © 2012 Scientific American LAYER BY LAYER stellation of symptoms, matches what we see in real life, that virtual chain of events Deconstructing the Brain becomes a candidate for a disease mecha­ nism, and we can even begin to look for The Human Brain Project intends to create a computer simulation of the 89 billion neurons potential therapeutic targets along it. inside our skull and the 100 trillion connections that wire those cells together. A meticulous This process is intensely iterative. We virtual copy of the human brain would potentially enable basic research on brain cells and integrate all the data we can find and pro­ circuits or computer-based drug trials. The project, which is seeking ¤1 billion in funding gram the model to obey certain biological from the European Union, would model each level of brain function, from chemical and rules, then run a simulation and compare electrical signaling up to the cognitive traits that underlie intelligent behaviors. the “output,” or resulting behavior of pro­ teins, cells and circuits, with relevant ex­ perimental data. If they do not match, we Molecular go back and check the accuracy of the data A century of research, beginning and refine the biological rules. If they do with the first inspection of a brain cell under a microscope, would translate match, we bring in more data, adding ever into a digital facsimile that combines more detail while expanding our model to component molecular parts to as­ a larger portion of the brain. As the soft­ sem­ble a cell that demonstrates the ware improves, data integration becomes essential properties of a neuron— faster and automatic, and the model be­ the transmission of electrical and haves more like the actual biology. Model­ chemical signals. ing the whole brain, when our knowledge of cells and synapses is still incomplete, no Cellular longer seems an impossible dream. A brain-in-a-box simulation will To feed this enterprise, we need data have to capture every detail of and lots of them. Ethical concerns restrict neurons and nonneuronal glial the experiments that can cells, including the exact geometric shapes of the dendrites and axons perform on the human brain, but fortu­ that receive and send information. nately the brains of all mammals are built according to common rules, with species- specific variations. Most of what we know about the genetics of the mammalian brain Circuits comes from mice, while monkeys have giv­ A model of the neural connections en us valuable insights into cognition. We between different brain areas and can therefore begin by building a unifying among neighboring cells may model of a rodent brain and then using it furnish clues to the origins of complex brain diseases such as as a starting template from which to de­ autism and schizophrenia. velop our human brain model—gradually integrating detail after detail. Thus, the models of mouse, rat and human brains will develop in parallel. The data that neuroscientists generate Regions Major neural substructures— will help us identify the rules that govern the amygdala (emotions), the brain organization and verify experimen­ hippocampus (memory), the tally that our extrapolations—those pre­ frontal lobes (executive control)— dicted chains of causation—match the bi­ can be inspected alone or as they ological truth. At the level of cognition, we interact with one another. know that very young babies have some grasp of the numerical concepts 1, 2 and 3 but not of higher numbers. When we are Whole Organ finally able to model the brain of a new­ An in silico brain might substitute for the actual organ. By removing born, that model must recapitulate both the computer code for a “gene,” what the baby can do and what it cannot. the virtual system can, for instance, A great deal of the data we need al­ mimic the effects of a mutation, ready exist, but they are not easily accessi­ as scientists do today by “knocking ble. One major challenge for the HBP will out” a gene in mice. The tool would avoid the lengthy breeding process be to pool and organize them. Take the and could simulate a multitude medical arena: those data are going to be of experimental conditions. immensely valuable to us not only be­ cause dysfunction tells us about normal

54 Scientific American, June 2012 Illustration by Emily Cooper © 2012 Scientific American function but also because any model we Once our brain simulator has been experiences. We will need to build the ma­ produce must behave like a healthy brain built, researchers will be able to set up in chinery to allow the model to change in re­ and later get sick in the same way that a silico experiments using the software sponse to input from the environment. real brain does. Patients’ brain scans will specimen much as they would a biological The litmus test of the virtual brain will therefore be a rich source of information. specimen, with certain key differences. To come when we connect it up to a virtual Currently every time a patient has a give you an idea of what these might be, software representation of a body and scan, that scan resides in a digital archive. think about how scientists currently place it in a realistic virtual environment. The world’s hospitals stock millions of search for the roots of disease by using Then the in silico brain will be capable of scans, and although they are already used mice in which a gene has been “knocked receiving information from the environ­ for research purposes, that research hap­ out.” They have to breed the mice, which ment and acting on it. Only after this pens in such a piecemeal way they remain takes time, is expensive and is not always achievement will we be able to teach it a largely untapped resource. If we could possible—for example, if the knockout is skills and judge if it is truly intelligent. Be­ bring together those scans on Internet- lethal to the embryo—even if one lays cause we know there is redundancy in the accessible “clouds,” collecting them with aside ethical concerns surrounding ani­ human brain—that is, one neural system patients’ records and biochemical and ge­ mal experimentation. can compensate for another—we can be­ netic information, doctors could look With the in silico brain, they will be gin to find which aspects of brain function across vast populations of patients for pat­ able to knock out a virtual gene and see are essential to intelligent behavior. terns that define disease. The power of the results in “human” brains that are dif­ The HBP raises important ethical is­ this strategy will come from being able to ferent ages and that function in distinctive sues. Even if a tool that simulates the hu­ mathematically pinpoint the differences ways. They will be able to repeat the exper­ man brain is a long way off, it is legitimate and similarities among all diseases. A mul­ iment under as many different conditions to ask whether it would be responsible to tiuniversity endeavor called the Alzheim­ as they like, using the same model, thus build a virtual brain that possessed more er’s Disease Neuroimaging Initiative is try­ ensuring a thoroughness that is not ob­ cortical columns than a human brain or ing to do just that by collecting neuroim­ tainable in animals. Not only could this ac­ that combined humanlike intelligence aging, cerebrospinal fluid and blood rec­ celerate the process by which pharmaceu­ with a capacity for number crunching a ords from large numbers of dementia tical researchers identify potential drug million times greater than that of IBM’s patients and healthy control subjects. targets, it will also change the way clinical Deep Blue, its chess-playing computer. trials are conducted. It will be much easier We are not the only ones setting the THE FUTURE OF COMPUTING to select a target population, and drugs bar high in attempting to reverse the frag­ last but not least, there is the computing that do not work or that have unaccept­ mentation of brain research. In May 2010 issue. The latest generation of Blue Gene is able side effects will be filtered out more the Seattle-based Allen Institute for Brain a peta-scale beast consisting of close to quickly, with the result that the entire R&D Science launched its Allen Human Brain 300,000 processors packed into the space pipeline will be accelerated and made more Atlas, with the goal of mapping all the of 72 fridges. Petaflops are sufficient to efficient. genes that are active in the human brain. model a rat brain of 200 million neurons What we learn from such simulations Funding is likely to be the main limit­ at a cellular level of detail but not a human will also feed back into the design of com­ ing factor for any group making an at­ brain of 89 billion neurons. For that puters by revealing how evolution pro­ tempt of this kind. In our case, the goal achievement, we need an exa-scale super­ duced a brain that is resilient, that performs will be achievable only if we obtain the computer, and even then a molecular-level multiple tasks rapidly and simultaneously support we need. Supercomputers are ex­ simulation of the human brain will be be­ on a massive scale—while consuming the pensive, and the final cost of the HBP is yond our reach. same amount of energy as a lightbulb—and likely to match or exceed that of the Hu­ Teams worldwide are racing to build that has a huge memory capacity. man Genome Project. In February 2013 we such computers. When they arrive, like Brainlike computer chips will be used will know if we have the green light. Mean­ previous generations of supercomputers, to build so-called neuromorphic comput­ while we press ahead with an enterprise they are likely to be adapted to simulating ers. The HBP will print brain circuits on we believe will give us unparalleled in­ physical processes, such as those used in silicon chips, building on technology de­ sight into our own identities as creatures nuclear physics. Biological simulations veloped in the European Union projects capable of contemplating the chiaroscuro have different requirements, and in col­ BrainScaleS and SpiNNaker. of a Caravaggio painting or the paradoxes laboration with large computer manufac­ The first whole-brain simulations we of quantum physics. turers and other industrial partners, our run on our instrument will lack a funda­ consortium of high-performance-comput­ mental feature of the human brain: they MORE TO EXPLORE ing experts will configure one such ma­ will not develop as a child does. From birth Links to a few Human Brain Project sites: chine for the task of simulating a brain. onward, the cortex forms as a result of the Human Brain Project: www.humanbrainproject.eu They will also develop the software that proliferation, migration and pruning of BrainScaleS: http://brainscales.kip.uni-heidelberg.de will allow us to build unifying models neurons and of a process we call plasticity SpiNNaker:http://apt.cs.man.ac.uk/projects/ SpiNNaker from the lowest to the highest resolution that is highly dependent on experience. SCIENTIFIC AMERICAN ONLINE so that it will be possible, within our sim­ Our models will instead begin at any arbi­ Watch a video of a brain network in operation at ulator, to zoom in and out among mole­ trary age, leapfrogging years of develop­ ScientificAmerican.com/jun2012/brain-project cules, cells and the entire brain. ment, and continue from there to capture

June 2012, ScientificAmerican.com 55 © 2012 Scientific American © 2012 Scientific American