RESEARCH ARTICLE SUMMARY◥

SINGLE-CELL ANALYSIS

Single-cell reconstruction ofdevelopmental trajectories duringzebrafish embryogenesisJeffrey A. Farrell,* Yiqun Wang,* Samantha J. Riesenfeld, Karthik Shekhar,Aviv Regev,† Alexander F. Schier†

INTRODUCTION:During embryogenesis, plu-ripotent cells gradually become specialized andacquire distinct functions andmorphologies. Be-cause much of the specification process is con-trolled through changes in gene expression, theidentification of the transcriptional trajectoriesunderlying cell fate acquisition is paramount tounderstanding and manipulating development.

RATIONALE: Traditional approaches havestudied specific fate decisions by analyzing thetranscription of a few selectedmarker genes orby profiling isolated, predefined cell populations.The advent of large-scale single-cell RNA se-

quencing (scRNA-seq) provides the means tocomprehensively define the gene expressionstates of all embryonic cells as they acquiretheir fates. This technology raises the possibilityof identifying the molecular trajectories thatdescribe cell fate specification by samplingdensely during embryogenesis and connectingthe transcriptomesof cells thathave similar geneexpression profiles. However, the numeroustranscriptional states and branch points, as wellas the asynchrony in developmental processes,pose major challenges to the computationalreconstruction of developmental trajectoriesfrom scRNA-seq data.

RESULTS:We generated single-cell transcrip-tomes from 38,731 cells during early zebrafishembryogenesis at high temporal resolution,spanning 12 stages from the onset of zygotictranscription through early somitogenesis. Wetook two complementary approaches to identifythe transcriptional trajectories in the data. First,wedeveloped a simulateddiffusion-based compu-tational approach, URD, which identified the tra-jectories describing the specification of 25 celltypes in the form of a branching tree. Second,we identified modules of coexpressed genes andconnected themacross developmental time. Com-

bining the reconstructeddevelopmental trajectorieswithdifferentialgeneexpres-sionanalysisuncoveredgeneexpressioncascades leadingto each cell type, includingpreviouslyunidentifiedmark-

ers and candidate regulators. Combining thesetrajectories with Seurat, which infers the spatialpositions of cells on the basis of their transcrip-tomes, connected the earlier spatial position ofprogenitors to the later fate of their descendants.Inspection of the developmental tree led to

new insights about molecular specification inzebrafish. For example, the first branch point inthe tree indicated that the first molecular spec-ification event may not only separate the germlayers but also define the axial versus nonaxialmesendoderm.Additionally, somedevelopmen-tal branch points contained intermediate cellsthat expressed genes characteristic ofmultipledownstreamcell fates. Gene expression analysisat one such branch point (the axialmesoderm)suggested that the intermediate cells switch theirspecification from one fate (notochord) to another(prechordal plate). Last, analysis of single-celltranscriptomes from a Nodal-signaling mutantrevealed that even at the whole-transcriptomelevel, mutant cells were canalized into a subsetof wild-type states and did not adopt any tran-scriptional states not observed in wild type, de-spite abnormal developmental signaling.

CONCLUSION: These findings reconstruct thegene expression trajectories during the embryo-genesis of a vertebrate and highlight the con-current canalization and plasticity of cell typespecification. The scRNA-seq data and develop-mental tree provide a rich resource for futurestudies in zebrafish: The rawandprocessed dataand the URD software are available for down-load, and the data can be browsed interactivelyonline. Last, this approach provides a broadlyapplicable frameworkwith which to reconstructcomplex developmental trajectories from single-cell transcriptomes.▪

RESEARCH

Farrell et al., Science 360, 979 (2018) 1 June 2018 1 of 1

The list of author affiliations is available in the full article online.*These authors contributed equally to this work.†Corresponding author. Email: aregev@broadinstitute.org(A.R.); schier@fas.harvard.edu (A.F.S.)Cite this article as J. A. Farrell et al., Science 360, eaar3131(2018). DOI: 10.1126/science.aar3131

Developmental tree of early zebrafish embryogenesis. Single-cell transcriptomes were generatedfrom zebrafish embryos at 12 developmental stages (six of which are shown).The transcriptionaltrajectories that describe the fate specification of 25 cell types were reconstructed from the data.Molecular specification is visualized with a force-directed layout, in which each cell is represented bya point (colored by developmental stage), proceeding from pluripotent cells (at the bottom center)outward to 25 distinct cell types. A subset of the identified trajectories are labeled in groups.

ON OUR WEBSITE◥

Read the full articleat http://dx.doi.org/10.1126/science.aar3131..................................................

on Novem

ber 12, 2020

http://science.sciencemag.org/

Dow

nloaded from

RESEARCH ARTICLE◥

SINGLE-CELL ANALYSIS

Single-cell reconstruction ofdevelopmental trajectories duringzebrafish embryogenesisJeffrey A. Farrell,1* Yiqun Wang,1* Samantha J. Riesenfeld,2 Karthik Shekhar,2

Aviv Regev,2,3† Alexander F. Schier1,2,4,5,6,7†

During embryogenesis, cells acquire distinct fates by transitioning through transcriptionalstates.To uncover these transcriptional trajectories during zebrafish embryogenesis, wesequenced 38,731 cells and developed URD, a simulated diffusion-based computationalreconstruction method. URD identified the trajectories of 25 cell types through earlysomitogenesis, gene expression along them, and their spatial origin in the blastula. Analysisof Nodal signaling mutants revealed that their transcriptomes were canalized into a subsetof wild-type transcriptional trajectories. Some wild-type developmental branch pointscontained cells that express genes characteristic of multiple fates.These cells appeared totrans-specify from one fate to another.These findings reconstruct the transcriptionaltrajectories of a vertebrate embryo, highlight the concurrent canalization and plasticity ofembryonic specification, and provide a framework with which to reconstruct complexdevelopmental trees from single-cell transcriptomes.

During embryogenesis, a single totipotentcell gives rise to numerous cell types withdistinct functions, morphologies, and spa-tial positions. Because this process is pri-marily controlled through transcriptional

regulation, the identification of the transcrip-tional states underlying cell fate acquisition isparamount to understanding and manipulatingdevelopment. Previous studies have presenteddifferent views of cell fate specification. For ex-ample, artificially altering transcription factorexpression (such as in reprogramming) has re-vealed remarkable plasticity of cellular fates (1–3).Conversely, classic embryological studies have in-dicated that cells are canalized to adopt perduringfates separated by epigenetic barriers. Techno-logical limitations necessitated that traditionalembryological studies focus on specific fate de-cisions with selected marker genes, but the ad-vent of single-cell RNA sequencing (scRNA-seq)raises the possibility of fully defining the tran-scriptomic states of embryonic cells as they ac-

quire their fates (4–8). However, the large numberof transcriptional states and branchpoints, aswellas the asynchrony in developmental processes,posemajor challenges to the comprehensive iden-tification of cell types and the computationalreconstruction of their developmental trajectories.Pioneering computational approaches to uncoverdevelopmental trajectories (5–7, 9–11) were eitherdesigned to address stationary or steady-stateprocesses or accommodate only small numbersof branch points and thus are insufficient foraddressing the complex branching structure oftime-series developmental data. We addressedthese challenges by combining large-scale single-cell transcriptomics during zebrafish embryo-genesis with the development of a new simulateddiffusion-based computational approach to re-construct developmental trajectories, called URD(named after the Norse mythological figure whonurtures the world tree and decides all fates).

High-throughput scRNA-seqfrom zebrafish embryos

We profiled 38,731 cells from 694 embryos across12 closely spaced stages of early zebrafish de-velopment using Drop-seq, a massively parallelscRNA-seq method (12). Samples spanned fromhigh blastula stage (3.3 hours postfertilization,just after transcription from the zygotic genomebegins), when most cells are pluripotent, to six-somite stage (12 hours postfertilization, shortlyafter the completion of gastrulation), whenmanycells have differentiated into specific cell types(Fig. 1A and table S1). In a t-distributed stochasticneighbor embedding (tSNE) plot (13) of the en-tire data set based on transcriptional similarity,

it is evident that developmental timewas a strongsource of variation in the data, but the underlyingdevelopmental trajectories were not readily ap-parent (Fig. 1B). Consistent with the understand-ing that cell types becomemore transcriptionallydivergent over time, cells fromearly stages formedlarge continuums in the tSNE plot, whereas morediscrete clusters emerged at later stages (Fig. 1C).

URD reconstructs complex branchingdevelopmental trajectories

Acquisition of many single-cell embryonic tran-scriptomeswith high temporal resolution createdthe possibility of reconstructing developmentaltrajectories through similarity in gene expressionprofiles. Such an approach would allow the in-vestigation of the gene expression dynamics andthe timing ofmolecular specification—when pro-genitor populations become transcriptionally dis-tinct from each other and begin to express theregulators that will drive their future fates. There-fore, we developed URD, an approach to uncovercomplex developmental trajectories as a branch-ing tree. URD extends diffusion maps [originallypresented for single-cell differentiation analysisin pioneering work by Haghverdi et al. and theirR package, destiny (9, 10)] through several ad-vances: It introduces a new way to order cells inpseudotime, finds developmental trajectories, dis-covers an underlying branching tree that abstractsspecification, and visualizes the data (Fig. 1D andsupplementarymaterials,materials andmethods).In general, URD operates by means of “sim-

ulating diffusion,” using discrete random walksandgraph searches to approximate the continuousprocess of diffusion (supplementarymaterials,ma-terials and methods). Briefly, URD constructs ak-nearest-neighbor graph between transcriptomesin gene expression space; graph edges are assignedtransition probabilities that are used as weightsin later simulations and describe the chancea random walk would move along each edge(Fig. 1D, 1) (9, 10). The user identifies the root(s)(starting points) and tips (end points) of the de-velopmental process. Cells are next assigned apseudotime—an ordering that should reflect theirdevelopmental progress rather than absolutetime—in order to compensate for developmentalasynchrony. URD calculates pseudotime by sim-ulating diffusion from the root to determineeach cell’s distance from the root (as the averagenumber of diffusive transitions needed to reachit across several simulations) (Fig. 1D, 2). Next,the developmental trajectory (the path in geneexpression) between each tip and the root is de-termined by identifying which cells are visitedby simulated biased random walks initiated inthat tip; the walks are biased to only transitionto cells of equal or earlier pseudotime so thatwhen they reach developmental branch points,they proceed toward the root and do not ex-plore other cell types (Fig. 1D, 3). Then, URD re-constructs a branching tree structure by joiningpairs of trajectories where they pass through thesame cells (Fig. 1D, 4; black and purple edges, forexample). Last, the data are visualized with aforce-directed layout based on cells’ visitation

RESEARCH

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 1 of 7

1Department of Molecular and Cellular Biology, HarvardUniversity, Cambridge, MA 02138, USA. 2Klarman CellObservatory, Broad Institute of MIT and Harvard, Cambridge,MA 02142, USA. 3Howard Hughes Medical Institute, KochInstitute for Integrative Cancer Research, Department ofBiology, Massachusetts Institute of Technology, Cambridge,MA 02140, USA. 4Center for Brain Science, HarvardUniversity, Cambridge, MA 02138, USA. 5FAS Center forSystems Biology, Harvard University, Cambridge, MA 02138,USA. 6Biozentrum, University of Basel, Switzerland. 7AllenDiscovery Center for Cell Lineage Tracing, University ofWashington, Seattle, WA 98195, USA.*These authors contributed equally to this work.†Corresponding author. Email: aregev@broadinstitute.org (A.R.);schier@fas.harvard.edu (A.F.S.)

on Novem

ber 12, 2020

http://science.sciencemag.org/

Dow

nloaded from

frequency by the random walks from each tip(Fig. 1D, 5) (14). The developmental trajectoriesidentified with URD are akin to cell lineages butdiffer from classical definitions of cell lineage be-cause they are reconstructed from observed geneexpression and do not measure mother-daughterrelationships between cells. URD does not re-quire any prior knowledge of the developmen-tal trajectories it seeks to find (such as thenumber of branch points or definition of inter-mediate states).

Reconstructed developmental treerecapitulates molecular specificationduring zebrafish embryogenesis

Application of URD to the early zebrafish em-bryogenesis scRNA-seq data generated a treewhose branches reflected embryonic specifica-tion trajectories. To define the final cell popula-tions (the tips of the tree), we clustered cellsfrom the final stage of our time course and de-termined cluster identity through the expressionof known marker genes (fig. S1). The recoveredtree followed the specification of 25 final cellpopulations across 16 branch points (Fig. 1E,

fig. S2, and movie S1). The reconstructed treesubstantially recapitulated the developmentaltrajectories expected from classical embryologicalstudies (15–17 ). For example, the primordial germcells, the enveloping layer cells (EVLs), and thedeep-layer blastomeres already formed separatetrajectories by high stage (Fig. 1F). Unexpectedly,the first branch point within the blastoderm notonly separated the ectoderm from the mesendo-derm but also divided the axial mesoderm fromthe remainder of the mesendoderm (Fig. 1G).Later branching events were also recovered, suchas the separation of paraxial, lateral, and inter-mediate mesoderm (Fig. 1H); the separation ofthe non-neural and neural ectoderm (Fig. 1I); andthe eventual branching of the non-neural ecto-derm into epidermis, neural plate border, andmultiple preplacodal ectoderm trajectories (Figs.1E and 2, trajectory, and fig. S3). Displaying theexpression of classic marker genes on the de-velopmental tree highlighted expected trajectoriesand confirmed their annotation (Fig. 2, geneexpression, and fig. S3). For example, consistentwith its known expression, the notochord markergene noto was restricted primarily to two tra-

jectories (the entire notochord trajectory andthe later stages of the tailbud trajectory) andconfirmed that the branch point between thenotochord and prechordal plate was correctlyplaced (18). These results show that URD canreconstruct the highly complex branching tra-jectories of early zebrafish embryogenesis solelyon the basis of large-scale scRNA-seq data.

Connected gene modulessupport reconstructeddevelopmental trajectories

To complement the specification tree, and in orderto find groups of genes that are coexpressedwithin cell populations, we applied non-negativematrix factorization (NMF) to the single-cell tran-scriptomes (19). This approach producedmodulesof covarying genes and described cells in termsof module expression (which is more robust thanindividual gene expressionmeasurements) (4, 20).Modules were annotated post hoc accordingto their highly ranked classic marker genes,and similar modules from adjacent stages werelinked to each other according to the overlap oftheir most highly ranked genes. This approach

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 2 of 7

Fig. 1. Generation of a developmentalspecification tree for early zebrafishembryogenesis using URD. (A) Single-celltranscriptomes were collected from zebrafishembryos at 12 developmental stages (coloreddots) spanning 3.3 to 12 hours postfertilization(hpf). (B) tSNE plot of the entire data, coloredby stage [as in (A)]. Developmental time is astrong source of variation, and the underlyingdevelopmental trajectories are not immediatelyapparent. (C) tSNE plot of data from two stages(top, 50% epiboly; bottom, six-somite). Clustersare more discrete at the later stage. (D) URD’sapproach for finding developmental trajectories.1: Transition probabilities are computed fromthe distances between transcriptomes and usedto connect cells with similar gene expression.2: From a user-defined “root” (such as cells ofthe earliest time point), pseudotime is calculatedas the average number of transitions required toreach each cell from the root. 3: Trajectoriesfrom user-defined “tips” (such as cell clusters inthe final time point) back to the root areidentified by simulated random walks that arebiased toward transitioning to cells younger orequal in pseudotime. 4: To recover an underlyingbranching tree structure, trajectories are joinedagglomeratively at the point where they containcells that are reached from multiple tips.5: The data are visualized with a force-directedlayout based on cells’ visitation frequency bythe random walks from each tip. (E) Force-directedlayout of early zebrafish embryogenesis,optimized for 2D visualization (supplementarymaterials, materials and methods, fig. S2,and movie S1), colored by stage [as in (A)] withterminal populations labeled. EVL, envelopinglayer; P, placode; Adeno., adenohypophyseal;Trig., trigeminal; Epi., epibranchial; Panc.+Int.,pancreatic + intestinal; RBI, rostral blood island;ICM, intermediate cell mass. (F to I) Cell populations downstream of early and intermediate branch points recovered by URD.

RESEARCH | RESEARCH ARTICLEon N

ovember 12, 2020

http://science.sciencem

ag.org/D

ownloaded from

created chains of connected gene modules thatprovided an alternative way to track develop-mental trajectories (fig. S4 and table S2). Forexample, the prechordal plate chain of connectedgene modules extended from 50% epiboly tosix-somite stage, during which the top rankinggenes gradually changed fromearly to latemarkersfor the prechordal plate (table S3).The URD-generated developmental tree and

the chains of connected gene modules providedtwo differentways to analyze the scRNA-seq dataand define developmental trajectories. To deter-mine how congruent these approaches are, wehighlighted cells in the developmental tree ac-cording to their expression of connected genemodules. Cells that express connected genemodules occupied specific URD-recovered de-velopmental trajectories, further supportingthe structure of the developmental tree re-constructed by URD (Fig. 2, module expression,and fig. S3).

Gene cascades revealexpression dynamics alongdevelopmental trajectories

Gene expression and gene module analysis werecombined in order to find candidate regulatorsandmarkers along each trajectory uncoveredwithURD. Genes and connected gene modules wereassociated with developmental trajectories bytesting for their differential expression down-stream of each branch point of URD’s recoveredbranching structure (Fig. 3, A and B; and sup-

plementary materials, materials and methods).Gene expression dynamics were then fit with animpulse model (21) so as to determine the onsetand offset time of their expression,whichwas thenused to order genes. As an example, sequentialexpression was observed for 197 genes enrichedin the prechordal plate during its specification,including several well-known transcription fac-tors or signaling molecules that confirm the va-lidity of our approach (such as gsc, foxa3, klf17,and frzb) (Fig. 3C and figs. S5 to S7) (22–25).This cascade contained several regulatory factors(such as fzd8a, fzd8b, mllt1b, and inhbaa) with-out described roles in the prechordal plate thatwould be candidates for reverse genetic screens.This cascade also contained more than 40 genesthat were not previously annotated as associatedwith the prechordal plate (fig. S6), and thosetested by means of in situ hybridization were in-deed expressed in the prechordal plate (fig. S7).Thus, combining URD and genemodule analysisuncovered the transcriptional cascades that ac-companied the development of progenitors intodifferentiated cells and highlighted both previouslycharacterized and newly identified trajectory-enriched genes.

Combining developmental trajectorieswith spatial analysis infersprogenitor locations

The URD-generated tree is a powerful way tovisualize developmental trajectories but lacksspatial information. We therefore asked whether

the trajectories could be traced to their spatialorigin at the late blastula stage (16, 17). First, aspatial map of the Drop-seq 50% epiboly tran-scriptomes was generated by using Seurat, amethod we previously developed to infer the spa-tial locations of single-cell transcriptomes by com-paring the genes expressed in each transcriptomewith the spatial expression patterns of a fewlandmark genes obtained from RNA in situ hy-bridization (20). Second, Seurat’s spatial mapwas combined with either URD or connectedgene module analysis (as parallel, independentapproaches) in order to associate cell popula-tions at six-somite stage with the location oftheir “pseudoprogenitors” at 50% epiboly. Inone approach, we used URD’s simulated randomwalks from cell populations at six-somite (Fig. 3A,pink bars) to infer their pseudoprogenitor cellsand then plotted the spatial location of the 50%epiboly pseudoprogenitors using Seurat (Fig.4A). In the other approach, we plotted the spa-tial expression of each 50% epiboly gene moduleand identified its connected gene modules atlater stages (Fig. 3B, blue bars on right); the cellsthat express these connected gene modules arehighlighted on the developmental tree (Fig. 4B)and are the “pseudodescendents” that arise fromthe spatial domain identified by the 50% epibolygene module.The results from the two approaches were

highly concordant and agreed with classic fate-mapping experiments (16, 17). For example, bothapproaches associated the dorsal margin of the50% epiboly embryo with axial mesodermal fates(prechordal plate and notochord). Likewise, inboth cases the animal pole was associated withectodermal fates, with the ventral animal sidebiased to non-neural fate and the dorsal animalside to neural fate. Overall, these results showthat scRNA-seq data can be used not only toreconstruct specification trajectories and theirassociated gene cascades but also to connectthe earlier spatial position of progenitors to thelater fate of their descendants.

Nodal signaling mutant cells arecanalized into a subset of wild-typetranscriptional states

Previous studies of mutant embryos have pro-vided important insights into embryonic fatespecification, but scRNA-seq raises the possibilityof rapidly phenotyping mutants both genome-wide and at single-cell resolution. To test thisidea,weprofiledmaternal-zygotic one-eyedpinheadmutants (MZoep), which lack the coreceptor forthe mesendoderm inducer Nodal (26, 27). Wefirst askedwhether previousMZoep embryologicalresults could be reconstructed simply from scRNA-seq data. Transcriptomes were generated withSmart-seq2 (28), with deeper mRNA coveragethan that of Drop-seq (fig. S8); 325 MZoep and1047 wild-type transcriptomes were collected at50% epiboly, when Nodal signaling is normallyactive at the blastodermmargin. Using Seurat, weinferred the spatial origin of both wild-type andMZoep transcriptomes based on a wild-type land-mark map (Fig. 5A). As predicted, no MZoep cells

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 3 of 7

Fig. 2. Developmental trajectories, genes, and connected gene modules overlaid on theforce-directed layout. From top to bottom, (i) the trajectories identified by URD from the root to agiven population (or group of populations), (ii) gene expression of a classicalmarkerof that population, and(iii) expression of a six-somite gene module active in the population and its connected modules fromearlier stages.The remainder are presented in fig. S3.

RESEARCH | RESEARCH ARTICLEon N

ovember 12, 2020

http://science.sciencem

ag.org/D

ownloaded from

mapped to the margin of a wild-type embryo,where mesendodermal progenitors arise. Apply-ing NMF revealed that in MZoep mutants,expression of the marginal dorsal, dorsal, andmarginal gene modules is greatly reduced or ab-sent, but nomutant-specificmodules were found(Fig. 5B). NMF recovered the same 50% epibolygenemodules from Smart-seq andDrop-seq data(Fig. 5B and table S4), which allowed us to usethe chains of connected Drop-seq gene modulesto predict the MZoep mutant phenotype at laterstages. Namely, our approach successfully pre-dicted the loss of paraxialmesoderm, ventrolateralmesoderm, axial mesoderm, and endoderm (Fig.5C), as well as the continued presence of thetailbud, as found in previous embryologicalstudies (26). Analysis of additional single-celltranscriptomes from wild-type and MZoep mu-tants at six-somite stage largely verified our pre-dictions (fig. S9 and table S5). Together, theseresults show that combiningmodest-scale scRNA-seq inmutantswith the large-scale developmentalreference tree constructed for wild type can rap-idly provide phenotypic insights: In the absenceof Nodal signaling, marginal blastomeres thatwould normally become mesendodermal pro-genitors instead become ectodermal and tailprogenitors, resulting in the absence of mesen-dodermal cell types and an altered fate map, asshown in previous cell-tracing experiments (27).We next askedwhether themutant scRNA-seq

dataset could provide novel insights into cellidentity in the absence of Nodal signaling. Mor-phological analysis and fate mapping of MZoepmutants has found that all cells appear to adoptwild-type fates (26, 27). However, intricate inter-actions exist between the several signaling path-ways active during early development, and theelimination of Nodal signaling changes levelsand domains of other developmental signals inthe embryo. Thus, MZoepmutant cells could po-tentially perceive signaling input combinationsthat do not occur in wild-type embryos, whichmay result in gene expression states not observedin wild type. We therefore wondered whethermutant cells expressed previously unknown com-binations of gene modules under this alteredsignaling landscape or were transcriptionallyequivalent to a subset of wild-type states. Co-clustering of wild-type and MZoep transcriptomesaccording to their gene module expression re-vealed thatwhile somemesendodermal cell typeswere absent inMZoep at 50%epiboly, the remain-ing cells clustered with wild-type states (Fig. 5Dand fig. S10). This result indicates that even on thewhole-transcriptomic and single-cell level,mutantcells are canalized into a subset of wild-type fatesafter the loss of an essential signaling pathway.

Hybrid gene expression states revealdevelopmental plasticity

Inspection of the developmental tree revealedthat most cells fell along tight trajectories, butsome cells were located in intermediate zonesbetween branches. This observation seemed atodds with the view that embryonic cells traversea developmental trajectory until a branch point

funnels them cleanly and irreversibly into one ofmultiple downstream branches. For example, atthe axial mesoderm branch point, most cells fellalong the classic bifurcation fromprogenitor intonotochord or prechordal plate fates (15, 29), withwaves of gene expression corresponding to theirspecification and differentiation status (Fig. 6A,highlighted). However, ~5.4% were intermediatecells that expressed both notochord and pre-chordal plate markers (Fig. 6, A and B). Theintermediate cells and the completely bifurcatedaxial mesoderm cells with similar pseudotimescame from embryos at the same developmental

stage (Fig. 6C); moreover, they no longer ex-pressed genes characteristic of the progenitors(such as nanog and mex3b). These observationseliminated models in which intermediate cellsretained their progenitor state and delayed spec-ification. Instead, we noticed that these cellsexpressed early markers of both programs (gsc,frzb, ta, and noto) but later markers of only thenotochord program (ntd5 and shha). This observa-tion raised two possible models: (i) Cells initiallyexpress both programs, then shut off the pre-chordal plate program, and produce only noto-chordmarkers later (“dual-specification”model);

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 4 of 7

Fig. 3. Association of developmental trajectories with temporal gene expression patterns. (A) Theunderlying branching structure found by URD. Pink bars demarcate collections of cell types used inFig. 4A. (B) The structure of connected gene modules. Each circular node represents a module and iscolored by the developmental stage from which the module was computed (as in Fig. 1A). Blue barsdemarcate collection of modules downstream of each 50% epiboly (5.3 hpf) gene module used inFig. 4B. (C) Gene expression cascades during specification of the prechordal plate and notochord.Expression is displayed as a moving-window average in pseudotime (along the x axis), scaled to themaximum observed expression. Selected genes are labeled along the y axis. Genes are annotated withwhether they were identified as a differentially expressed gene, as a top ranking member of adifferentially expressed connected gene module, or both. Cascades for all trajectories (with all geneslabeled) are presented in fig. S5.

RESEARCH | RESEARCH ARTICLEon N

ovember 12, 2020

http://science.sciencem

ag.org/D

ownloaded from

or (ii) cells specified as notochord first expressearly and late notochordmarkers and then changetheir specification by shutting off notochord ex-pression and initiating early prechordal platemarker expression (“trans-specification”model).To distinguish between these two models, we

investigated their properties and location usingfluorescent RNA in situ hybridization. At 75%epiboly, some cells located in the border regionbetween the two tissues coexpressed an earlyprechordal plate marker (gsc) with either anearly (ta/ntl) or late (ntd5) notochord marker(Fig. 6, D to F, and fig. S11). The existence of theseintermediate cells in situ confirms that they arenot an artifact of scRNA-seq (such as cell doublets),and their defined spatial localization near theborder of the two tissues suggests that they arenot merely biological or technical noise. Instead,the location of coexpressing cells raises the pos-sibility that the boundary between the notochordand prechordal plate territories is refined duringgastrulation, after the two populations becometranscriptionally distinct. Many of the coexpress-ing cells exhibited bright nuclear foci that denotedsites of active transcription of the probed genes.Most of the cells with nuclear transcription focifor a single gene exhibited transcription of theearly prechordal plate marker, gsc. By contrast,

cells with active transcription of the notochordmarkers ta/ntl or ntd5 were rare (Fig. 6, E andF). The active transcription of the prechordalplate program supports the trans-specificationmodel—the intriguing possibility that some axialmesoderm cells move down the notochord spec-ification path but then trans-specify into pre-chordal plate cells.

Discussion

We describe a molecular specification tree of anearly vertebrate embryo by exploring large-scalesingle-cell transcriptomic data with URD, a com-putational approach to reveal developmentaltrajectories in transcriptional space. Our studylays the foundation for multiple areas of futureexploration.First, URD is a powerful tool to reveal the

transcriptional trajectories of complex devel-opmental processes such as zebrafish embryo-genesis. Combining URD with connected genemodule analysis uncovered the transcriptionalcascades underlying fate specification. More-over, combining URD with Seurat enabled theanchoring of developmental trajectories to theirspatial origins. We anticipate that similar aug-mentation of URDwith information about lineagerelationships (30, 31), chromatin dynamics (32),

and signaling will further deepen insights intodevelopmental processes. Additionally, some ofURD’s limitations could be addressed by futureenhancements. For instance, branching eventsdriven by single genes (such as sox32 in the en-doderm) (33, 34) are not capturedwith URDuntiladditional transcriptional differences arise, whichcould potentially be improved with more aggres-sive, iterative branch calling. Also, improvementsto the throughput and quality of scRNA-seq maydrive improved performance from URD: Moreclosely spaced time points could enable detectionof rapidly changing fate decisions and could re-veal sequential bifurcations at branch points cur-rently thought to enter multiple trajectories;larger numbers of cells could enable reconstruc-tion of rare, transcriptionally indistinct popula-tions (such as the floor plate and hypochord);and improved sequencing depth could enabledetection of important but lowly expressed regu-lators that are not found in the current data (suchas npas4l/cloche). In the future, URD could beused to analyze additional systems, such as invitro differentiation, tissue regeneration, cancer,and disease progression.Second, the scRNA-seq data and developmen-

tal tree provide a rich resource for future studiesof zebrafish embryogenesis. The presented datadescribe embryonic gene expression with unparal-leled temporal and cellular resolution, and thereconstructed tree thus provides an atlas of theexpression pattern and dynamics for nearly allgenes. This allows inspection of gene coexpres-sion withmuch greater ease thanwithmulticolorin situ hybridization, revealsmarkers for cell typesof interest, and associates uncharacterized geneswith particular cell types. Moreover, the data re-veal the progressive nature of cell fate specifica-tion and suggest potentially redundant regulatorsof developmental decisions with overlapping spa-tial and temporal expression, which would bemissed in forward genetic screens.Third, this work begins to illustrate the tra-

jectories, canalization, and potential plasticity offate specification.With respect to trajectories, ouranalysis suggests that not all cell fate decisionsare binary; multiple trajectories can arise simul-taneously from a pool of equipotent progenitors.For example, at the molecular level, the earliestspecification of blastomeres separates the axialmesoderm, nonaxialmesendoderm, and ectoderm,rather than just separating the germ layers. Thisconclusion is compatible with the observationthat the axial mesoderm progenitors reside atthe dorsal blastula margin, where maternal fac-tors that specify the Mangold-Spemann organizersimultaneously affect the fate of those progenitors(15). Concerning canalization, the analysis of aNodal signaling mutant reveals that even on thewhole-transcriptomic level, mutant cells adopt asubset of wild-type transcriptomic states, but notnew ones, demonstrating the extensive canaliza-tion during embryogenesis even after abnormaldevelopmental signaling.With respect to plasticity,the identification of intermediate cells at the axialmesoderm branch point demonstrates hybrid de-velopmental states. In situ hybridization experiments

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 5 of 7

Fig. 4. Molecular specification maps relate cell position at 50% epiboly to cell fate atsix-somite. (A) Visitation by random walks from given tip(s) (as proportion of visitation from alltips) and the spatial location of visited 50% epiboly cells (ventral side to the left). The six tip groupsare marked in Fig. 3A. (B) Spatial expression of 50% epiboly gene modules. Expression of connectedgene modules is plotted on the force-directed layout to highlight populations that will emergefrom the 50% epiboly module’s expression domain. The six groups of connected gene modules aremarked in Fig. 3B.

RESEARCH | RESEARCH ARTICLEon N

ovember 12, 2020

http://science.sciencem

ag.org/D

ownloaded from

suggest that these cells trans-specify from onecell type (notochord) to another (prechordal plate),and their border zone spatial localization suggestsrefinement of the boundary between these tworegions. These results support an alternative view

of developmental fate choice to the commoninterpretation of Waddington’s model: Down-stream of some branch points, some cells stilltransition across the “ridges” between differentlineages even after they are canalized into well-

defined transcriptional states. This suggests adevelopmental plasticity that has been observedafter perturbation (1–3, 35) but has not been wellestablished in normal developmental processes.We propose that in a continuous morphogen

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 6 of 7

Fig. 5. Characterization of Nodal signalingmutant with scRNA-seq and the developmentalspecification tree. (A) Spatial assignment ofwild-type and MZoep transcriptomes according toa wild-type landmark map indicates an absenceof wild-type marginal fates in MZoep (ventral, left).Shown at random are 311 wild-type transcriptomes(to match MZoep cell number). (B) (Top) wild-typeexpression domain of spatially restricted gene mod-ules identified in Smart-seq data (ventral to left).(Bottom) Violin plot of the maximum-scaled genemodule levels in wild-type and MZoepmutant cells.The marginal dorsal, dorsal, and marginal genemodules are absent or strongly reduced in MZoep.(C) Expression of gene modules connected to thosemissing in MZoep (marginal dorsal, dorsal, andmarginal; red) and connected to those remaining inMZoep (blue). (D) Hierarchical clustering of wild-typeand MZoepmutant transcriptomes, based on thescaled expression of gene modules. Number ofclusters is determined by the Davies-Bouldin index.Genotype is indicated beneath the heatmap (wildtype, green; MZoep, red). Clusters 3 and 8 containonly wild-type cells. All other clusters contain amixture of wild-type and MZoep cells.This clusteringanalysis was sufficiently sensitive to detect computa-tionally simulated altered states (fig. S10).

Fig. 6. Hybrid state of cells in the axial meso-derm. (A) Branch point plot, showing pseudotime(y axis) and random walk visitation preferencefrom the notochord (N, left) and prechordal plate(P, right) tips (x axis), defined as the differencein visitation from the two tips divided by the sum ofvisitation from the two tips. Direct trajectories tonotochord (green) and prechordal plate (pink) arehighlighted, and intermediate cells are circled.(B) Gene expression of notochord markers (top row)andprechordal platemarkers (bottomrow)at theaxialmesoderm branch point. Intermediate cells expressearly (ta/ntl and noto) and late (ntd5 and shha)notochord markers but only early prechordal platemarkers (gsc and frzb). (C) Cells at the branchpoint, colored by developmental stage. Intermediatecells have the same developmental stage as thatof fully bifurcated cells with similar pseudotimes.(D) Illustration of the prechordal plate (P) andnotochord (N) in the 75% epiboly embryo.(E and F) Double fluorescent in situ expression of theearly prechordal plate marker gsc (red) and eitherthe early notochord marker ta/ntl [(E), green]or the late notochord marker ntd5 [(F), green] at75% epiboly (8 hpf). Most cells contain onlyprechordal plate marker mRNA (for example, “1”) ornotochord marker mRNA (for example, “5”). Cellswith both prechordal plate and notochord markermRNA are observed at the boundary of the twotissues, with red nuclear transcription foci, indicatingactive transcription of gsc (for example, “2” to “4”) (fig. S11). Cells that contained both ntd5 and gscmRNAwere identified and scored for their nuclear transcriptionfoci, which indicate active transcription (supplementary materials, materials and methods); 56% had one or more observable nuclear transcription dots, of which80% showed only active gsc transcription, 7% showed both active gsc and active ntd5 transcription, and 13% showed active ntd5 transcription alone.

RESEARCH | RESEARCH ARTICLEon N

ovember 12, 2020

http://science.sciencem

ag.org/D

ownloaded from

gradient that resolves into discrete states, somecellswill receive an amount of signal that is on thecusp between the two states; these cells wouldend up at the boundary between the two tissues.It is conceivable that continued signaling couldprovide a “push” over a shallow ridge, such ascontinued Nodal signaling in the axial meso-derm (22) or continued retinoic acid signalingin the hindbrain (36, 37). Future studies com-bining lineage tracing and in vivo imaging ofgene expression will be needed to directly ob-serve such transitions.Collectively, our approach provides a frame-

work with which to reconstruct the specificationtrajectories ofmany developmental systemswith-out the need for prior knowledge of gene ex-pression patterns, fatemaps, or lineage trajectories.The generation of developmental trees for dif-ferent species will enable comparative studies,anddevelopmental statistics generated from suchcomparisons will help reveal conserved pathwaysand species-specific idiosyncracies.

REFERENCES AND NOTES

1. R. L. Davis, H. Weintraub, A. B. Lassar, Expression of asingle transfected cDNA converts fibroblasts to myoblasts. Cell51, 987–1000 (1987). doi: 10.1016/0092-8674(87)90585-X;pmid: 3690668

2. S. J. Tapscott et al., MyoD1: A nuclear phosphoproteinrequiring a Myc homology region to convert fibroblasts tomyoblasts. Science 242, 405–411 (1988). doi: 10.1126/science.3175662; pmid: 3175662

3. K. Takahashi, S. Yamanaka, Induction of pluripotent stem cellsfrom mouse embryonic and adult fibroblast cultures bydefined factors. Cell 126, 663–676 (2006). doi: 10.1016/j.cell.2006.07.024; pmid: 16904174

4. A. Wagner, A. Regev, N. Yosef, Revealing the vectors ofcellular identity with single-cell genomics. Nat. Biotechnol.34, 1145–1160 (2016). doi: 10.1038/nbt.3711;pmid: 27824854

5. R. Cannoodt, W. Saelens, Y. Saeys, Computational methodsfor trajectory inference from single-cell transcriptomics.Eur. J. Immunol. 46, 2496–2506 (2016). doi: 10.1002/eji.201646347; pmid: 27682842

6. S. C. Bendall et al., Single-cell trajectory detection uncoversprogression and regulatory coordination in human B celldevelopment. Cell 157, 714–725 (2014). doi: 10.1016/j.cell.2014.04.005; pmid: 24766814

7. C. Trapnell et al., The dynamics and regulators of cell fatedecisions are revealed by pseudotemporal ordering of singlecells. Nat. Biotechnol. 32, 381–386 (2014). doi: 10.1038/nbt.2859; pmid: 24658644

8. S. Jang et al., Dynamics of embryonic stem cell differentiationinferred from single-cell transcriptomics show a series oftransitions through discrete cell states. eLife 6, e20487(2017). doi: 10.7554/eLife.20487; pmid: 28296635

9. L. Haghverdi, F. Buettner, F. J. Theis, Diffusion maps for high-dimensional single-cell analysis of differentiation data.Bioinformatics 31, 2989–2998 (2015). doi: 10.1093/bioinformatics/btv325; pmid: 26002886

10. L. Haghverdi, M. Büttner, F. A. Wolf, F. Buettner, F. J. Theis,Diffusion pseudotime robustly reconstructs lineage branching.Nat. Methods 13, 845–848 (2016). doi: 10.1038/nmeth.3971;pmid: 27571553

11. M. Setty et al., Wishbone identifies bifurcating developmentaltrajectories from single-cell data. Nat. Biotechnol. 34, 637–645(2016). doi: 10.1038/nbt.3569; pmid: 27136076

12. E. Z. Macosko et al., Highly parallel genome-wide expressionprofiling of individual cells using nanoliter droplets. Cell 161,1202–1214 (2015). doi: 10.1016/j.cell.2015.05.002;pmid: 26000488

13. L. van der Maaten, G. Hinton, Visualizing data using t-SNE.J. Mach. Learn. Res. 9, 2579–2605 (2008).

14. T. M. J. Fruchterman, E. M. Reingold, Graph drawing by force-directed placement. Softw. Pract. Exper. 21, 1129–1164 (1991).doi: 10.1002/spe.4380211102

15. A. F. Schier, W. S. Talbot, Molecular genetics of axis formationin zebrafish. Annu. Rev. Genet. 39, 561–613 (2005).doi: 10.1146/annurev.genet.37.110801.143752; pmid: 16285872

16. C. B. Kimmel, R. M. Warga, T. F. Schilling, Origin andorganization of the zebrafish fate map. Development 108,581–594 (1990). pmid: 2387237

17. K. Woo, J. Shih, S. E. Fraser, Fate maps of the zebrafishembryo. Curr. Opin. Genet. Dev. 5, 439–443 (1995).doi: 10.1016/0959-437X(95)90046-J; pmid: 7580134

18. W. S. Talbot et al., A homeobox gene essential for zebrafishnotochord development. Nature 378, 150–157 (1995).doi: 10.1038/378150a0; pmid: 7477317

19. J.-P. Brunet, P. Tamayo, T. R. Golub, J. P. Mesirov, Metagenesand molecular pattern discovery using matrix factorization.Proc. Natl. Acad. Sci. U.S.A. 101, 4164–4169 (2004).doi: 10.1073/pnas.0308531101; pmid: 15016911

20. R. Satija, J. A. Farrell, D. Gennert, A. F. Schier, A. Regev,Spatial reconstruction of single-cell gene expression data.Nat. Biotechnol. 33, 495–502 (2015). doi: 10.1038/nbt.3192;pmid: 25867923

21. G. Chechik et al., Activity motifs reveal principles of timing intranscriptional control of the yeast metabolic network.Nat. Biotechnol. 26, 1251–1259 (2008). doi: 10.1038/nbt.1499;pmid: 18953355

22. C. Thisse, B. Thisse, M. E. Halpern, J. H. Postlethwait,Goosecoid expression in neurectoderm and mesendoderm isdisrupted in zebrafish cyclops gastrulas. Dev. Biol. 164,420–429 (1994). doi: 10.1006/dbio.1994.1212; pmid: 8045345

23. I. Seiliez, B. Thisse, C. Thisse, FoxA3 and goosecoid promoteanterior neural fate through inhibition of Wnt8a activitybefore the onset of gastrulation. Dev. Biol. 290, 152–163(2006). doi: 10.1016/j.ydbio.2005.11.021; pmid: 16364286

24. M. R. Gardiner, M. M. Gongora, S. M. Grimmond, A. C. Perkins,A global role for zebrafish klf4 in embryonic erythropoiesis.Mech. Dev. 124, 762–774 (2007). doi: 10.1016/j.mod.2007.06.005; pmid: 17709232

25. E. C. Swindell et al., Regulation and function of foxe3 duringearly zebrafish development. Genesis 46, 177–183 (2008).doi: 10.1002/dvg.20380; pmid: 18327772

26. K. Gritsman et al., The EGF-CFC protein one-eyed pinhead isessential for nodal signaling. Cell 97, 121–132 (1999).doi: 10.1016/S0092-8674(00)80720-5; pmid: 10199408

27. A. Carmany-Rampey, A. F. Schier, Single-cell internalizationduring zebrafish gastrulation. Curr. Biol. 11, 1261–1265 (2001).doi: 10.1016/S0960-9822(01)00353-0; pmid: 11525740

28. S. Picelli et al., Smart-seq2 for sensitive full-lengthtranscriptome profiling in single cells. Nat. Methods 10,1096–1098 (2013). doi: 10.1038/nmeth.2639; pmid: 24056875

29. K. Gritsman, W. S. Talbot, A. F. Schier, Nodal signalingpatterns the organizer. Development 127, 921–932 (2000).pmid: 10662632

30. A. McKenna et al., Whole-organism lineage tracing bycombinatorial and cumulative genome editing. Science 353,aaf7907 (2016). doi: 10.1126/science.aaf7907;pmid: 27229144

31. B. Raj et al., Simultaneous single-cell profiling of lineages andcell types in the vertebrate brain. Nat. Biotechnol. 40, 181(2018). pmid: 29608178

32. D. A. Cusanovich et al., The cis-regulatory dynamicsof embryonic development at single-cell resolution.Nature 555, 538–542 (2018). doi: 10.1038/nature25981;pmid: 29539636

33. T. Dickmeis et al., A crucial component of the endodermformation pathway, CASANOVA, is encoded by a novel sox-related gene. Genes Dev. 15, 1487–1492 (2001). doi: 10.1101/gad.196901; pmid: 11410529

34. Y. Kikuchi et al., casanova encodes a novel Sox-related proteinnecessary and sufficient for early endoderm formation inzebrafish. Genes Dev. 15, 1493–1505 (2001). doi: 10.1101/gad.892301; pmid: 11410530

35. M. E. Halpern et al., Cell-autonomous shift from axialto paraxial mesodermal development in zebrafish floatinghead mutants. Development 121, 4257–4264 (1995).pmid: 8575325

36. C. B. Moens, V. E. Prince, Constructing the hindbrain: Insightsfrom the zebrafish. Dev. Dyn. 224, 1–17 (2002). doi: 10.1002/dvdy.10086; pmid: 11984869

37. L. Zhang et al., Noise drives sharpening of gene expressionboundaries in the zebrafish hindbrain. Mol. Syst. Biol. 8, 613(2012). doi: 10.1038/msb.2012.45; pmid: 23010996

ACKNOWLEDGMENTS

We thank B. Raj and J. Gagnon for assistance collectingsamples when two hands were insufficient, M. Rabani forassistance with the impulse model, M. Haesemeyer for thecritical suggestions of directly simulating diffusion and of usingvisitation frequency as a method of dimensionality reduction,the Harvard Center for Biological Imaging and Bauer CoreFacility for support, and members of the Schier and Regevlaboratories for helpful discussions. We thank B. Raj, J. Gagnon,S. Pandey, N. Lord, M. Rabani, A. Carte, M. Haesemeyer,I. Whitney, and T. Montague for comments on the manuscript.Funding: This research was supported by the NIH (A.F.S.,J.A.F., and A.R.), Allen Discovery Center for Cell Lineage Tracing(A.F.S.), Jane Coffin Childs Memorial Fund (J.A.F.), CharlesA. King Trust (J.A.F.), Howard Hughes Medical Institute (A.R.),and Klarman Cell Observatory (A.R.). Author contributions:J.A.F., Y.W., A.R., and A.F.S, conceived the study; J.A.F., Y.W.,and A.F.S. wrote the paper with revisions by A.R. and S.J.R.;J.A.F. and Y.W. collected the data; S.J.R. performed preliminaryanalyses; J.A.F., Y.W., and A.F.S. performed the data analysispresented in the manuscript; J.A.F. developed URD with inputfrom A.R, S.J.R., and Y.W.; K.S. developed the variable geneidentification method. S.J.R. and J.F. developed the 5′ UMISmart-seq2 mapping pipeline; Y.W. performed the connectedgene modules and spatial analyses. Competing interests:A.R. is a SAB member of ThermoFisher Scientific, SyrosPharmaceuticals, and Driver Group. Data and materialsavailability: The raw data reported in this paper are archived atNCBI GEO (accession no. GSE106587) and in processed andinteractively browseable forms in the Broad Single-Cell Portal(https://portals.broadinstitute.org/single_cell/study/single-cell-reconstruction-of-developmental-trajectories-during-zebrafish-embryogenesis). URD is available from GitHub(https://github.com/farrellja/URD).

SUPPLEMENTARY MATERIALS

www.sciencemag.org/content/360/6392/eaar3131/suppl/DC1Materials and MethodsFigs. S1 to S11Tables S1 to S5References (38–51)Movie S1Supplementary Analysis

26 October 2017; accepted 5 April 2018Published online 26 April 201810.1126/science.aar3131

Farrell et al., Science 360, eaar3131 (2018) 1 June 2018 7 of 7

RESEARCH | RESEARCH ARTICLEon N

ovember 12, 2020

http://science.sciencem

ag.org/D

ownloaded from

Single-cell reconstruction of developmental trajectories during zebrafish embryogenesisJeffrey A. Farrell, Yiqun Wang, Samantha J. Riesenfeld, Karthik Shekhar, Aviv Regev and Alexander F. Schier

originally published online April 26, 2018DOI: 10.1126/science.aar3131 (6392), eaar3131.360Science 

, this issue p. 981, p. eaar3131, p. eaar5780; see also p. 967Sciencepave the way for the comprehensive reconstruction of transcriptional trajectories during development.early organogenesis, to map cell states and differentiation across all cell lineages over time. These data and approaches

examined whole frog embryos, spanning zygotic genome activation throughet al.more and more specialized. Briggs 25 distinct zebrafish cell types. The branching tree revealed how cells change their gene expression as they becomeand applied a computational approach to construct a branching tree describing the transcriptional trajectories that lead to

profiled the transcriptomes of tens of thousands of embryonic cellset al.layer formation, and early organogenesis. Farrell more than 90,000 cells throughout zebrafish development to reveal how cells differentiate during axis patterning, germ

sequenced the transcriptomes ofet al.development of vertebrate embryos (see the Perspective by Harland). Wagner Three research groups have used single-cell RNA sequencing to analyze the transcriptional changes accompanying

As embryos develop, numerous cell types with distinct functions and morphologies arise from pluripotent cells.Mapping the vertebrate developmental landscape

ARTICLE TOOLS http://science.sciencemag.org/content/360/6392/eaar3131

MATERIALSSUPPLEMENTARY http://science.sciencemag.org/content/suppl/2018/04/25/science.aar3131.DC1

CONTENTRELATED

http://science.sciencemag.org/content/sci/360/6392/967.fullhttp://science.sciencemag.org/content/sci/360/6387/367.fullhttp://science.sciencemag.org/content/sci/360/6392/981.fullhttp://science.sciencemag.org/content/sci/360/6392/eaar5780.full

REFERENCES

http://science.sciencemag.org/content/360/6392/eaar3131#BIBLThis article cites 49 articles, 11 of which you can access for free

PERMISSIONS http://www.sciencemag.org/help/reprints-and-permissions

Terms of ServiceUse of this article is subject to the

is a registered trademark of AAAS.ScienceScience, 1200 New York Avenue NW, Washington, DC 20005. The title (print ISSN 0036-8075; online ISSN 1095-9203) is published by the American Association for the Advancement ofScience

Science. No claim to original U.S. Government WorksCopyright © 2018 The Authors, some rights reserved; exclusive licensee American Association for the Advancement of

on Novem

ber 12, 2020

http://science.sciencemag.org/

Dow

nloaded from