Gene-Expression Maps Fall Into Place

NEWS & VIEWS RESEARCH

unaffected. Remarkably, forced expression whether calsyntenin 3β functions as a general Maryland 21218, USA. of S100b in the brown adipocytes of Clstn3b- ER chaperone for the secretion of proteins that e-mail: [email protected] knockout mice is sufficient to correct the deficits lack signal peptides, or is specific for S100b. 1. Cannon, B. & Nedergaard, J. Physiol. Rev. 84, in sympathetic innervation and thermogenesis In humans, BAT had been thought to 277–359 (2004). in these animals. The emerging picture is of an exist only in infants, until imaging studies 2. Zeng, X. et al. Nature 569, 229–235 (2019). adipocyte-derived mechanism that mediates revealed deposits of thermogenic brown fat 3. Bartness, T. J., Vaughan, C. H. & Song, C. K. Int. J. crosstalk between brown adipocytes and in adults10. Clstn3b is expressed in human Obesity 34, S36–S42 (2010). 4. Wang, W. & Seale, P. Nature Rev. Mol. Cell Biol. 17, innervating sympathetic fibres (Fig. 1). adipose tissue, and might be involved in the 691–702 (2016). Zeng and colleagues’ findings raise several regulation of BAT innervation, as it is in mice. 5. Ludwin, S. K., Kosek, J. C. & Eng, L. F. questions. For example, how is calsyntenin 3β Zeng and colleagues’ study might therefore J. Comp. Neurol. 165, 197–207 (1976). 6. Duteil, D. et al. Cell Rep. 17, 1008–1021 (2016). expression regulated in brown adipocytes? The inform therapeutic strategies to enhance 7. Glebova, N. O. & Ginty, D. D. J. Neurosci. 24, authors found that one upstream regulator sympathetic innervation, and thereby harness 743–751 (2004). of this protein is the enzyme lysine-specific the thermogenic potential of BAT to combat 8. Hall, J. A. et al. Endocrinology 151, 4573–4582 (2010). demethylase 1 (LSD1), which mediates obesity and its metabolic consequences. ■ 9. Rabouille, C. Trends Cell Biol. 27, 230–240 (2017). 10. Nedergaard, J., Bengtsson, T. & Cannon, B. Am. J. the differentiation of brown and beige fat. Physiol. Endocrinol. Metab. 293, E444–E452 (2007). Adipose-specific loss of LSD1 in mice causes Rejji Kuruvilla is in the Department of downregulation of BAT-specific genes and Biology, Johns Hopkins University, Baltimore, This article was published online on 1 May 2019. aberrant induction of genes typically found in white adipose tissue6. The growth of sympathetic neurons mostly TRANSCRIPTION relies on a protein called nerve growth factor (NGF). NGF-deficient mice have pronounced disruption of sympathetic innervation of multiple peripheral tissues7, but BAT Gene-expression maps innervation has not been specifically examined in this context. It remains to be seen whether NGF acts together with S100b to coordinate fall into place BAT innervation, or is dispensable, as in the innervation of the trachea7. Could other Knowing the gene-expression pattern of individual cells can unlock their factors derived from peripheral tissue control identity. A refined method for generating cellular RNA profiles offers a way the sympathetic innervation of specific to obtain such data at a high level of spatial resolution in intact tissues. targets? If so, they could be useful tools for manipulating such innervation and studying its role in individual tissues. SAMANTHA A. MORRIS has brought huge efficiency gains and cost Sympathetic innervation controls the reductions3–5. However, single-cell RNA expression of genes involved in thermo onitoring messenger RNA in cells sequencing requires tissue disruption and cell genesis and the differentiation of brown and is a way to gather data that provides destruction, which causes loss of spatial infor beige fat4. However, Zeng et al. found that many biological insights. Improve mation about cellular location that would be loss of Clstn3b does not affect the expression Mments in the methods for detecting cellular valuable for cell-type identification. of genes involved in adipocyte development, mRNA are helping to provide more-detailed Efforts to retain spatial information are a thermogenesis or mitochondrial function, pictures of gene expression in the cells of a current focus in the development of methods including the gene that encodes UCP1. Only particular tissue. Writing in Science, Rodriques for single-cell analysis. For example, a the expression of the Dio2 gene, which encodes et al.1 describe a new approach to assessing method called multiplexed in situ hybridiza an enzyme that produces the active form of a cellular RNA that they term Slide-seq. This tion, along with certain sequencing techniques, thyroid hormone, is significantly altered by the versatile technique couples high-throughput has enabled RNA to be monitored on a sub loss or overexpression of Clstn3b. The Dio2 RNA sequencing with a way to capture spatial cellular scale in intact tissues6–8. However, protein is abundant in BAT, and, as is seen in information about the location of the analysed these approaches require substantial technical the Clstn3b-knockout mice, mice lacking Dio2 cells in tissues. expertise, which has limited their widespread have impaired thermogenesis and decreased Scientists often try to tackle the complexity adoption. cold tolerance, and are susceptible to diet- of a biological system by breaking it down to Rodriques and colleagues’ approach is induced obesity8. Altered thyroid signalling focus on cells, and cataloguing these individual conceptually rooted in an earlier3 method for might therefore contribute to BAT dysfunction units on the basis of their identity. Conven single-cell RNA analysis. In that technique, in the absence of calsyntenin 3β. tional microscopic imaging of cells in a tissue called Drop-seq, the tissue being analysed The final issue arising from Zeng and can provide much information, such as cell is disrupted to separate the cells, which are colleagues’ work is how calsyntenin 3β type and function, but imaging data of many then loaded onto a microfluidic device. Indi promotes the secretion of S100b. The synthesis cells can still lack a high level of detail about vidual cells are encapsulated in nanolitre- of proteins destined for cellular secretion cellular features — the characteristics known scale droplets together with a microbead is initiated in the cytoplasm. The nascent as the cellular phenotype. coated with copies of a DNA sequence that proteins typically contain an amino-acid The development of techniques for provides a unique ‘barcode’ for the identifi sequence known as a signal peptide, which rigorously analysing the molecular contents cation of material from each droplet (which directs them from the cytoplasm to the ER of individual cells offers an alternative way to is assumed to contain a single cell). The for the start of their secretory journey while capture thousands of cellular features and gen cells are broken open, which allows the cel they are still being synthesized. However, some erate an unbiased picture of the cells in a given lular RNA to bind to the DNA barcodes. A secreted proteins lack signal peptides and tissue. For example, single-cell RNA sequen step called reverse transcription generates are imported to the ER after they have been cing has been rapidly and widely adopted since DNA corresponding to the captured RNA synthesized9. S100b lacks a peptide signal, its emergence ten years ago2. The use of micro sequences, and this DNA is tagged with the and so the new findings raise the question of fluidic-based technologies in this approach barcode DNA. DNA sequencing then enables

RESEARCH NEWS & VIEWS

this region by analysing 66 sequential tissue a Glass Microbead b c sections. The authors generated RNA profiles slide layer Frozen tissue of 1.5 million microbeads, demonstrating that Slide-seq can be scaled up as necessary. Their method identified gene-expression patterns that could be used to assign distinct cerebel lar cell types together into spatially defined subpopulations in the mouse brain. Such Location of RNA expression of many genes is spatially depend Microbead sequences assigned ent and is independent of cell type, but this valuable information is usually lost in conven tional single-cell profiling approaches during the cell-dissociation step9. The authors also Unique DNA characterized cellular transcription dynamics ‘barcode’ RNA sequence in a mouse model of traumatic brain injury, DNA Reverse revealing that an initial wave of cell prolif sequencing transcription eration after injury was followed by cellular Spatially resolved RNA pro le of tissue differentiation. One limitation of Slide-seq is that infor mation about the size and shape of each cell Barcode-tagged and other key physical properties is not deter DNA mined, unlike the case for RNA-profiling methods that use a microscopy technique 10 Barcode location called histology. A Nature paper published recorded Microbeads dissociated in March offers a way to capture such cellular and DNA sequenced features while also tracking the subcellular localization of mRNA — this technique to Figure 1 | Recording the spatial location of RNA sequences in tissues. Rodriques et al.1 report a image RNA transcripts in cells in tissues is technique called Slide-seq, which enables RNA in frozen tissue samples to be observed in a way that called seqFISH+. Previous approaches similar provides spatial information about gene-expression patterns. a, The first step is the preparation of to seqFISH+ were restricted by the optical- glass slides that contain a layer of microbeads. Each microbead has copies of a unique DNA sequence resolution limits of microscopes and the attached that provides an identification ‘barcode’ (only one barcode shown per bead). Sequencing high density of mRNA transcripts in a cell. these barcodes directly on the slide allowed the authors to record the slide position associated with each barcode. b, Rodriques and colleagues placed a frozen sample of tissue, such as a sample of mouse SeqFISH+ overcame these problems through brain, on the slide. The cellular RNA bound to the underlying barcode DNA. An enzyme-mediated step a method that sequentially uses probes that called reverse transcription generated a DNA sequence corresponding to the RNA, and this DNA was provide an imaging palette of 60 colours as a attached to the barcode. The microbeads were dissociated from the slide and the DNA was sequenced. tool with which to monitor individual mRNA c, DNA-sequence analysis enabled the authors to generate a map of where specific RNA sequences were molecules. This considerable palette breadth expressed in the tissue. allows each mRNA to be localized as a dot in an individual cell using a confocal microscope, all of the RNA sequences associated with a and kidney, and for specimens of the human enabling the expression of about 10,000 genes particular barcode sequence to be deduced. brain region called the cerebellum. to be visualized. Drop-seq provided a huge increase in the Slide-seq might not capture the full profile of Together, Slide-seq and seqFISH+ repre numbers of individual cells that could have RNA in single cells. For example, the RNA mol sent valuable and complementary tools to their RNA sequences determined in a stand ecules bound to a microbead are not necessarily aid major advances in efforts to retain spatial ard experiment. representative of an entire cell if only part of the information during transcriptional analyses With Slide-seq, Rodriques and colleagues cell is captured by a microbead. Alternatively, of cells. A key feature of these tools lies in developed an innovative twist on the Drop-seq if a microbead is positioned at the boundary their relative ease of use. Before the develop approach by assembling the DNA-barcoded between two cells, RNA from both cells might ment of SeqFISH+, similar types of approach microbeads in a layer on a glass slide (Fig. 1). be recorded. Therefore, to infer the single-cell required a complex imaging system called DNA sequencing of this material on the slide identity corresponding to each microbead pro super-resolution microscopy. Slide-seq is per allowed the authors to determine the position file, the authors turned to more-comprehensive haps even more accessible, with experimental and sequence of the barcodes corresponding single-cell transcriptional data collected in pre costs of just a few hundred US dollars and only to each microbead. A frozen tissue sample viously published single-cell RNA sequencing about 40 hours of work required to prepare the was then placed on the microbead layer, and work. Using these data sets, the authors con materials needed for the sequencing step. It is RNA from the cells was captured by the under structed a cell-type ‘dictionary’ to infer cell an exciting time in this field. A commitment to lying beads. As with Drop-seq, in subsequent identity from the Slide-seq data. This strategy create access to these technologies will facili DNA sequencing steps, the unique barcodes required the development of a computational tate their adoption and continued develop allowed each RNA transcript to be traced back approach to reconstruct each Slide-seq bead ment, placing them at the heart of current to its bead of origin, and hence allowed assem gene-expression profile, based on a weighted efforts to construct high-resolution cellular bly of a bead-specific transcriptional profile. combination of the cell-type transcriptional sig atlases. ■ The ability to assign each RNA transcript to a natures detected in a corresponding single-cell specific microbead location enabled the RNA RNA-sequencing experiment. Samantha A. Morris is in the Departments profiling information to be presented at a spa Rodriques et al. conducted a range of experi of Developmental Biology and of Genetics, tial resolution on the scale of individual cells ments that showcased Slide-seq’s power. They and in the Center of Regenerative Medicine, in a tissue sample. Rodriques et al. confirmed investigated the RNA profiles of cells in the Washington University School of Medicine in that Slide-seq worked effectively for differ hippocampal region of the mouse brain, St. Louis, St Louis, Missouri 63110, USA. ent tissue types, including mouse brain, liver using a strategy to capture the 3D structure of e-mail: [email protected]

NEWS & VIEWS RESEARCH

1. Rodriques, S. G. et al. Science 363, 1463–1467 5. Zheng, G. X. et al. Nature Commun. 8, 14049 Zhuang, X. Science 348, aaa6090 (2015). (2019). (2017). 8. Lee, J. H. et al. Science 343, 1360–1363 (2014). 2. Tang, F. et al. Nature Methods 6, 377–382 (2009). 6. Shah, S., Lubeck, E., Zhou, W. & Cai, L. Neuron 92, 9. Zhu, Q., Shah, S., Dries, R., Cai, L. & Yuan, G. C. 3. Macosko, E. Z. et al. Cell 161, 1202–1214 (2015). 342–357 (2016). Nature Biotechnol. 36, 1183–1190 (2018). 4. Klein, A. M. et al. Cell 161, 1187–1201 (2015). 7. Chen, K. H., Boettiger, A. N., Moffitt, J. R., Wang, S. & 10. Eng, C.-H. L. et al. Nature 568, 235–239 (2019).

NANOTECHNOLOGY phase-change materials; these are located on two types of micro-ring resonator, which have a synaptic or neuronal function. Unmodulated light that is injected at the A role for optics in various operating wavelengths picks up the neuronal excitations that have accumu lated in the phase-change material, and then AI hardware passes them to the next layer of the network. Even without on-chip optical gain (a process Experiments show how an all-optical version of an artificial neural network — a in which a medium transfers energy to the type of artificial-intelligence system — could potentially deliver better energy light that is transmitted through it), this set- efficiency can conventional computing approaches. See Article p.208 up could potentially be scaled up to larger networks. The authors demonstrate, on a small scale, both supervised and unsupervised GEOFFREY W. BURR opportunity for unconventional computing learning — that is, training is achieved using techniques. For example, researchers are labelled data, which is how DNNs learn, and ptical fibres transmit data across the exploring DNN accelerators that are based using data without such labels, which is how world in the form of light and are on emerging non-volatile memory devices6,7. humans tend to learn. the backbone of modern telecom Such devices retain information even when Because the weights are implemented by Omunications1. However, when such data need their power source is switched off, and light absorption, negative weights require to be analysed, they get converted from light can offer improved speed and energy effi a large bias signal, which must not activate into electrons and are then processed using ciency for DNNs through analog electronic the phase-change material. An alternative electronics. There was a time when optics was computation. approach13 that can readily offer negative considered as the basis for a potential com Why not, therefore, also consider optics? weights uses devices called Mach–Zehnder puting technology2, but it became difficult for Structures that direct light — whether they interferometers. In these devices, a single optical computing to compete with the rapid be an optical fibre for use in telecommunica waveguide is split into two arms and then improvements made by its electronic counter tions or a waveguide patterned onto a photonic recombined; this causes the amount of trans part. In the past few years, however, concern chip — can be packed with vast amounts of mitted light to depend on the difference in has been growing about the energy costs of data. Inside such a waveguide, many wave optical phase between the two paths. How computation. Therefore, optics is receiving lengths of light can propagate together, using ever, it might be challenging to combine this attention again, both as a way to decrease a technique known as wavelength division approach with wavelength division multiplex energy requirements3, and as a special-purpose multiplexing. Each wavelength can then be ing, because the arms of each interferometer hardware for accelerating artificial-intelligence modulated (altered in such a way that it can would need to introduce the appropriate phase algorithms such as deep neural networks carry information) at a rate that is limited difference for each wavelength. (DNNs). On page 208, Feldmann et al.4 report by the available bandwidths associated Photonic DNNs still present substantial an intriguing advance towards all-optical with electronic-to-optical modulation and challenges. Their total power usage can be low implementations of such networks. optical-to-electronic detection. in ideal situations, but thermo-optic power A DNN comprises many layers of artificial Structures called resonators enable indi is frequently required to adjust and maintain neurons and artificial synapses, which are con vidual wavelengths to be added to or removed the differences in optical phase in the arms of nections between the neurons. The strengths from the waveguide, like wagons on a freight each Mach–Zehnder interferometer. More of these connections are called weights and train. For example, micrometre-scale, ring- over, the total optical power that is injected can be either positive, indicating neuronal shaped (micro-ring) resonators can implement into a system containing phase-change excitation, or negative, implying inhibition. A arrays of synaptic weights8. Such resonators materials must be calibrated carefully, so that DNN learns to perform tasks such as image can be modulated thermally9, electro-opti the materials respond to incoming signals recognition by varying its synaptic weights in cally10,11 or, as in Feldmann and colleagues’ exactly as intended. Although phase-change a way that minimizes the difference between its work, through phase-change materials12. These materials can also be used to adjust Mach– actual output and the desired output. materials can switch between an amorphous Zehnder phases, unavoidable cross-coupling Central processing units and other digital- phase and a crystalline phase, which differ between how strongly the materials absorb based hardware accelerators5 are typically used greatly in their ability to absorb light. Under light and how much they slow it down poses a for DNN computations. A DNN can be trained ideal conditions, the resulting multiply– considerable complication. using a known set of data, whereas an already accumulate operations would require only a Phase-change materials seem to be well trained DNN can be applied to unknown small amount of power. suited for the non-volatile long-term storage data in a task called inference. In either case, Feldmann et al. present an all-optical neural of synaptic weights that are based on micro- although the amount of computation is vast, network on a millimetre-scale photonic chip, ring resonators needing only infrequent the variety of operations is modest, because in which there are no optical-to-electronic adjustment. However, when used in the role ‘multiply–accumulate’ operations dominate conversions within the network. Inputted data of neuron, the speed of crystallization of across the many synaptic weights and neuronal are electronically modulated onto different such materials will limit the maximum rate at excitations. wavelengths for injection into the network, but which neurons can be excited. Furthermore, DNNs are known to still work well when after that has been performed, all the data stay the need to melt the materials to induce a full computational precision is low5. As a result, on the chip. Both weight modulation and neu neuronal reset after every potential excitation these networks represent an intriguing ron integration are achieved using integrated event will rapidly consume the large, but finite,