Single-cell analysis of shared signatures and transcriptional diversity during zebrafish development

Summary

During development, animals generate distinct cell populations with specific identities, functions, and morphologies. We mapped transcriptionally distinct populations across 489,686 cells from 62 stages during wild-type zebrafish embryogenesis and early larval development (3–120 hours post-fertilization). Using these data, we identified the limited catalog of gene expression programs reused across multiple tissues and their cell-type-specific adaptations. We also determined the duration each transcriptional state is present during development and identify unexpected long-term cycling populations. Focused clustering and transcriptional trajectory analyses of non-skeletal muscle and endoderm identified transcriptional profiles and candidate transcriptional regulators of understudied cell types and subpopulations, including the pneumatic duct, individual intestinal smooth muscle layers, spatially distinct pericyte subpopulations, and recently discovered best4+ cells. To enable additional discoveries, we make this comprehensive transcriptional atlas of early zebrafish development available through our website, Daniocell.

Blurb:

Sur et al. maps transcriptional populations with high temporal resolution during zebrafish development. They identify putative long-term progenitor states, reused developmental gene expression programs, rare perivascular cell subtypes, transcriptional profiles of molecularly uncharacterized cell types (pneumatic duct and intestinal smooth muscle), and infer the developmental program of best4+ intestinal cells.

Graphical Abstract

Introduction

A central quest of developmental biology is to understand how morphological and functional diversity observed in distinct cell types relates to transcriptional diversity—which transcriptional differences are present, and how do they confer distinct morphologies and functions? In a crucial step toward addressing this question, single-cell RNAseq (scRNAseq) approaches provide an opportunity to map the transcriptionally distinct populations of cells during development in an unbiased way^1–13. scRNAseq-derived molecular cell type catalogs from diverse animals have identified previously invisible cell types or states^2,14–17 and enabled detailed comparisons of cell type transcriptional similarity across species^18–20. Additionally, high-resolution scRNAseq timecourses analyzed with pseudotime trajectory and related approaches^2,21–24 infer the gene expression cascades that accompany cell specification and differentiation to identify candidate regulatory factors or functional genes for downstream testing. Altogether, single-cell transcriptomic approaches complement classical genetic and embryological study of developmental biology by identifying unexpected complexity, characterizing cells at a whole-transcriptome level instead of with selected marker genes, and helping generate and sharpen hypotheses.

Zebrafish are a powerful model for studying vertebrate embryogenesis, since their genetic tractability, optical clarity, external fertilization, high fecundity, and the high degree of conservation among vertebrates have greatly facilitated developmental screens^25–28, rigorous lineage tracing^29–32, detailed mechanistic studies, and human disease modeling^33,34. Here, we present a global view of cell type heterogeneity and transcriptional progression during zebrafish development, encompassing the first transcriptional events after zygotic genome activation to a freely-swimming, feeding animal. We generated a 489,686 single-cell transcriptome atlas from 62 closely spaced developmental stages. We searched for gene expression programs reused across multiple tissues and identified the transcriptional states present for the longest durations during early development. Since global analysis of such complex data rarely exposes its full cellular diversity, we also performed focused analyses within the endoderm and non-skeletal muscle. This identified uncharacterized cell subtypes, resolved the expression profiles of some enigmatic cell types, and enabled us to propose putative progenitors, candidate regulator genes, and cell type-specific gene expression cascades within them. To make these data accessible to the zebrafish community and other researchers, we developed the web portal, Daniocell.

Results

Temporally dense sampling of zebrafish development using single-cell RNAseq

Developmental processes are incredibly dynamic, and understanding them requires identifying the timing, ordering, and coordination of gene expression within and between cell types. To profile wild-type molecular cell states during zebrafish development, we generated single-cell transcriptomes from TL/AB zebrafish embryos and larvae across 50 developmental stages, covering 14–120 hours post-fertilization (hpf). To capture underrepresented populations and allow robust cluster identification, we sampled heavily every 12 hours; to provide high temporal continuity for investigating gene expression dynamics, we sampled smaller cell numbers every 2 hours (Figure 1A, Table S1). To reduce cost, we employed MULTI-seq cell hashing³⁵, which barcodes cells before sample collection and enables higher input cell concentrations followed by computational identification and removal of resulting increased cell doublets. We mapped reads to the GRCz11 genome, annotated using the Lawson Lab Zebrafish Transcriptome Annotation (v4.3.2) that harmonizes Ensembl and Refseq annotations, includes improved 3’ UTR models, and proposes additional gene models³⁶. We combined these with published 3.3–12 hpf wild-type zebrafish single-cell data² to generate a continuous single-cell time-course of 489,686 cells from 62 developmental stages spanning 3.3 to 120 hpf (Figure 1A–B, Figure S1A–C). We detected an average of 8,621 transcripts/cell and 1,745 genes/cell across these transcriptomes (Figure S1D–G, Table S1). Our dataset complements recent wild-type zebrafish single-cell atlases that use different techniques (nuclei vs. cells), profile shorter developmental durations, or have lower frequency of collection timepoints^2,9,37–40.

Figure 1: A high temporal resolution single-cell RNAseq timecourse encompassing embryogenesis and early larval development.

(A) Developmental stages (colored dots) from which single-cell transcriptomes were collected. (B–C) UMAP projection of single-cell transcriptomes, colored by (B) developmental stage (as in Figure 1A) and (C) curated major tissues. (D) Expression pattern specificity for each gene. Genes were categorized by thresholding (dotted lines) based on their coefficient of variation (CV) of cluster mean expression (log-transformed) and number of cell types expressing each gene. (E) Example gene expression patterns for different expression specificity categories. (F) Comparison of temporal (X-axis) and cell-type (Y-axis) expression variation for each gene. See also Figure S1, Tables S1–S3.

We next identified transcriptional cell types and cell states during development. We iteratively clustered single-cell transcriptomes in a semi-supervised manner, first assigning cells to 19 broad tissue types, then to individual clusters by clustering within each tissue (Figure S1H). Clusters represent groups of transcriptionally similar cells — where transcriptional differences can indicate (a) different cell types, (b) the same cell type at different times (since gene expression often changes within a cell type during development) or in a different anatomical location, or (c) the same cell type in a different cell state (such as a cell transcriptionally responding to DNA damage). Clusters were annotated according to their highly and differentially expressed genes (see Materials and Methods), based on gene expression patterns described by prior publications, the ZFIN gene expression database⁴¹, and the Thisse in situ collection^42,43. Altogether, this identified 521 clusters that represent ~200–300 terminal and intermediate cell types or cell states (Table S2). For visualization, cells were projected onto a Uniform Manifold Approximation and Projection (UMAP) and colored by developmental stage or tissue of origin (Figure 1B–C, Figure S1A, H). This dataset captures the specification and differentiation of cells comprising most major organs, though a few tissues dissociated poorly and are therefore underrepresented (heart, chondrocytes, and swim bladder).

To enable public access to these data, we created the website, Daniocell (https://daniocell.nichd.nih.gov/) that contains pre-computed information about gene expression during zebrafish development. Each gene has a page demonstrating: (1) its expression pattern (as UMAP and dot plot) across each tissue and cell type, resolved by time, and (2) which genes have the most similar and dissimilar expression patterns across all cells and each tissue. Each cell cluster has a page demonstrating: (1) the developmental stages of cells in the cluster, (2) whether they are undergoing the cell cycle, and (3) the most specific and most highly expressed genes within each cluster. Daniocell is intended to facilitate rapid research within the zebrafish community and also to act as a companion to this manuscript.

Spatiotemporal variation of gene expression during development

To investigate the quality of these data and understand global gene expression patterns during development, we determined whether we observed expression of each gene and whether its expression pattern was constant or varied during development. We detected expression of 67% of genes in the reference transcriptome, with expression defined as ≥1 transcript detected in ≥22 cells (the size of our smallest cluster) and mean expression of ≥0.1 counts/cell in at least one cluster. To assay gene expression variation during development, we used as rough measures the coefficient of variation (CV) of cluster mean expression and the number of cell types or tissues that expressed a gene. Defining thresholds for categorization within these continuous metrics involves a degree of arbitrariness, but helps describe the overall trends. ~16.7% of genes are expressed ubiquitously, with ~3.6% having extremely low fluctuation between clusters, often colloquially referred to as ‘housekeeping genes’—far fewer than identified by studies in adult organs^44,45 (Figure 1D, Table S3). ~21.5% of genes had very restricted expression patterns, with ~11.2% restricted to ≤3 cell types and the others restricted to ≤4 tissues (Figure 1D, E, Table S3). Most genes (62%) are neither specific nor ubiquitous and have varied expression across several cell types and tissues. Very few genes varied temporally but not between cell types (Figure 1F), and all seemed to be maternally supplied transcripts with slow degradation. Finally, about half of annotated transcripts lack proper gene names—unnamed cDNA clones (si:ch211-, si:dkey-, zgc:, etc.), computationally predicted genes (LOC-, CABZ-, BX-, AL-, etc.), and putative genes identified through de novo transcriptome assembly by the Lawson lab (XLOC-). Despite their lower level of experimental support, unnamed cDNA clones were detected with similar frequency to ‘named’ genes and exhibited slightly more spatiotemporal variation (Figure S1I-K). Computationally predicted and ‘XLOC–’ genes were detected much less frequently, but exhibited higher spatiotemporal restriction when detected (Figure S1I-K). In conclusion, while scRNAseq does not capture all mRNA within each cell, these data describe quantitative, temporally-resolved, and cell-type specific gene expression patterns for about two-thirds of the genome during the first five days of zebrafish development. Moreover, it emphasizes that most genes vary in expression across cell types, emphasizing the breadth of developmental gene regulation that remains to be dissected. Lastly, it suggests that many understudied and ‘unnamed’ genes are expressed during development with restricted expression patterns and may warrant additional investigation.

Mapping the duration of transcriptional states using a genome-wide approach

The genes expressed in a developing cell (i.e. its transcriptional state) change significantly during specification and differentiation. Our dataset provides an opportunity to use cells’ full transcriptomes to identify the duration of each transcriptional state (i.e., how long cells with that particular gene expression profile can be found during development). In this analysis, we sought to (1) determine the duration of transcriptional states during development, (2) assess whether cells in each transcriptional state were undergoing continued cell division, and (3) use these combined analyses to identify cycling populations present for long developmental durations.

To identify developmental state durations, we used an ε-nearest neighbor approach: we (1) defined a distance in scaled gene expression space to represent transcriptionally similar cells (ε), (2) found each cell’s neighbors within that ε-neighborhood, and (3) computed the difference in developmental stage between each cell and its neighbors (Figure 2A). A longer “mean stage difference” would indicate a cell with transcriptionally similar counterparts from a longer duration in development. In this metric, a cell with “mean stage difference” of 36 hours indicates that its transcriptionally similar cells are on average between 36 hours younger and 36 hours older.

Figure 2: Duration of transcriptional states during development.

(A) Schematic of approach to identify transcriptionally similar cells by epsilon (ε) neighborhood, determine similarity of developmental stage to transcriptionally similar cells, and categorize into ‘short-term’ or ‘long-term’ states. (B) UMAP projection colored by categorization as cycling or non-cycling and ‘short-term’ or ‘long-term’. Populations of transcriptionally similar proliferating cells found for ≥24 hours are labeled. (C–E) Cells, colored by stage (left) or gene expression (right) indicate how different developmental stages and transcriptional states align with cycling, non-cycling, ‘short-term’, and ‘long-term’ categorizations in transcriptionally stable, terminally differentiating (C, fast muscle), transcriptionally changing but non-cycling (D, intestine), and classical long-term progenitor populations (E, radial glia). (F) Timeline bar plots showing the duration of ‘long-term’ cycling cell states. Each bar represents a cell population and the length of bar represents the minimum timespan that encompasses 80% of its ε-neighbors. (G–I) Cells represented as in panels C–E for uncharacterized, putative long-term cycling populations. In panels C–E, and G–I, dashed lines mark thresholds for cycling (blue) and ‘long-term’ (red). See also Figure S2, Table S4.

Most transcriptional states had short durations during development, consistent with development’s dynamic nature (Figure S2A–B). During the first 5 days, 85% of cells were in transcriptional states that last for ≤24 hours, and 96.8% last for ≤48 hours (Figure S2C). For downstream analyses, we classified cells into “short-term” and “long-term” states, where the 15% of transcriptional states present ≥24 hours were considered “long-term”. This threshold balanced focusing on states with unusually long durations, but spanned multiple landmark stages to ensure the analysis remained robust. Additionally, each cell was classified as “cycling” or “non-cycling” based on its expression of transcripts associated with different cell cycle phases (Figure 2B, Figure S2A–D).

First, we investigated cell types where development from early progenitors to terminal differentiation is well documented (e.g., fast muscle) through the lens of this analysis (Figure 2C). Early somitic myoblasts (tbx6+) were short-term and cycling; at 24hpf, most cells had exited the cell cycle but were still identified as short-term as they began myoblast fusion (mymk+); finally, starting at 60 hpf, cells were detected as long-term, non-cycling as they terminally differentiated into fast-twitch muscle fibers (myhz1.1⁺/tnni2b.2⁺) (Figure 2C). This pattern was observed across several cell types, including the slow-muscle (Figure S2E), lens fibers (Figure S2F), blood, posterior epidermis, and some peridermal cells. This suggests that (as one might expect) many cells first exit the cell cycle and then adopt stable transcriptomes as they differentiate, thereby becoming ‘long-term’.

Many other cell types, however, exited the cell cycle during later stages, but remained predominantly classified as “short-term” transcriptional states (Figure 2D, Figure S2G). For example, intestinal cells (cdx1b+) exit the cell cycle beginning around 60–72 hpf as they downregulate endodermal progenitor genes (e.g. foxa3) and upregulate enterocyte (apoa4a, abca1b) and secretory (scg3, fev) differentiation markers, but almost none are classified as “long-term” (Figure 2D). This aligns with previous observations that zebrafish digestive organs mature after 96 hpf, despite becoming physiologically functional as early as 76 hpf^46–48. A similar pattern was observed in axial mesoderm derivatives (Figure S2G), the pronephros, pigment cells, liver, and vasculature. Thus, many cell types stop dividing, but continue changing transcriptionally at 4 dpf (the end of this analysis), despite often being functionally required by larvae at that stage and thus presumably differentiated.

27% of cells in “long-term” transcriptional states were still undergoing cell division. Some represented fate-specified transit amplifying populations that proliferate over an extended period due to developmental asynchrony (e.g. erythroblasts and peridermal cells). This category also included known stem cell populations, such as muscle satellite cells (Figure 2E), radial glia (Figure S2H), Müller glia (Figure S2I), hematopoietic stem cells, mesenchymal progenitors, and some neural progenitors (e.g. her4.1⁺) (Figure 2F). As an example, dermomyotome-derived^49,50 satellite cells (pax7a+, meox1+) are observed both cycling and non-cycling from 11–120 hpf; they are initially short-term, but become ‘long-term’ at 48 hpf, while many are still cycling. This transition is marked by meox1 downregulation and upregulation of transcription factors including pitx3 and ECM genes like col18a1b (Figure 2E). This analysis suggests these long-term progenitor populations (satellite cells, radial glia, Müller glia, and others) remain transcriptionally consistent during development (Figure 2F, Table S4). Recent zebrafish neuronal development studies similarly identified 15 dpf retinal progenitors that are nearly transcriptionally identical to embryonic progenitors prior to 24 hpf⁵¹. We find that several eye progenitor populations usually studied with individual marker genes^52–56 are “long-term” even when considered at a whole-transcriptome level, including (i) vsx2⁺/vsx1^–/notch3⁺ optic progenitors, (ii) photoreceptor progenitors (otx5⁺/crx⁺), and (iii) progenitors of retinal ganglion cells, cone bipolar cells and oligodendrocytes (hes2.1⁺/atoh7⁺/vsx1⁺/olig2⁺) (Figure 2F). Our analysis additionally identified “long-term” cycling populations that are not well characterized, including gfra3⁺/ret⁺/msc⁺ cephalic muscle cells (Figure 2G), pou2f3⁺/sox8b⁺ taste epithelia (Figure 2H), and sp8a⁺/b⁺ olfactory epithelia (Figure 2I), among others (Figure 2F, Table S4). These may represent additional unappreciated stem cell states or transit amplifying states and merit future investigation.

Importantly, some classic stem cell populations will not be identified as ‘long-term’ because they mature transcriptionally during development — for instance, pax3+ spinal cord progenitors are present from 14–58 hpf, but are considered “short-term” because their transcriptomes differ markedly over that time. This was similarly shown for insm1a/her4.1+ hypothalamic progenitors⁵¹. Additionally, “long-term” transcriptional states that appear at or after 96 hpf are classified as “short-term” because we lack measurements after 120 hpf. Moreover, states would be identified as ‘long-term’ if cells are constantly found in that state, even if any individual cell only transiently occupies it — for instance if cells rapidly traverse through a state but are constantly initiating it (e.g., somitogenesis). Finally, the specific results of this analysis depend on the choice of parameters that define transcriptional similarity (ε) and ‘long-term,’ but the overall trends are robust to parameter choice (in that the ‘long-term’ states identified here are some of most persistent during development).

Identification of shared developmental gene expression programs and tissue-specific adaptations

Distinct cell types often share common cellular states, features, or elaborations, such as cilia, which are produced by olfactory sensory neurons and kidney tubular cells, among others. While ciliogenesis results from deployment of a shared genetic program across multiple cell types^57,58, for many shared features, it remains unclear whether the underlying genetic program is also shared. Whole-animal, time-course, single-cell RNAseq data can help address this question, as it enables gene expression correlation analysis across all cell types and many developmental stages. Here we catalogued gene expression programs (GEPs) that are shared between two or more tissues during zebrafish development (Figure 3A). To find these GEPs, we: (1) smoothed data using a 5 nearest-neighbor network to reduce technical noise⁵⁹, (2) used fuzzy c-means (FCM) clustering to group genes with similar expression over time and across tissues^60–63, (3) filtered out poor quality GEPs and GEPs expressed in a single tissue. This approach identified 87 shared GEPs with 8–735 member genes (average 95). We could clearly annotate 79 of these shared GEPs based on individual member genes (Figure 3A, Table S4); it is currently unclear whether the remaining 8 represent technical artifacts or developmental GEPs undescribed in the literature.

Figure 3: Identification of reused gene expression programs.

(A) Binary heatmap showing expression domains of gene expression programs (“GEPs”, x-axis with select annotations) shared by cell types in multiple tissues (y-axis). Table S5 contains full GEP annotations. (B) Gene expression dot plot of megalin-associated and SLC transporter genes shared between intestinal lysosome-rich enterocytes (“LREs”) and pronephros proximal convoluted and straight tubules (“PCT” and “PST”). SLC transporter genes are colored according to the family/category they belong to. (C) Dot plot of top-loaded genes from five epithelial GEPs: one shared across all epithelia, two comprising classical epithelial genes, and two tissue-specific. X-axis: cell types with epithelial characteristics and muscle as a non-epithelial comparison. (D) Dot plot of shared and tissue-specific members of module GEP-94, associated with mucin O-glycosylation. RPE: retinal pigmented epithelium; EVL: enveloping layer. See also Figure S3, Table S5.

13 GEPs were shared across most tissues and represented essential cell functions such as metabolism, cell cycle, or cytoskeletal organization (Figure 3A, Table S4). The remaining 74 annotated GEPs were restricted to particular cell types within 2 or more tissues and generally represented functionally related genes involved in conferring specific cellular features. For instance, as confirmation of our approach, we identified four GEPs associated with one of the most well-studied recurring cellular features—the motile cilium^64,65 (Figure S3A).

A recurring theme was that many processes identified as modules had a core, shared program that was then modified within individual cell types; we observed this trend across cells specialized for protein absorption, mucous-producing cells, epithelial cells, and others. For instance, GEP-193 was expressed in cells specialized for protein absorption (pronephros proximal tubules and intestinal lysosome-rich enterocytes, LREs) and contained Megalin-associated receptor-mediated endocytic machinery genes (lrp2a/b, cubn, amn, dab2), known to be shared between these cell types^66,67 (Figure 3B). This module identified several additional shared membrane-spanning transporters, suggesting that Megalin-associated endocytosis is accompanied by a broader nutrient absorption program (Figure 3B). GEP-193 also included several pronephros-specific transporters, demonstrating that the shared absorption program is further specialized in individual tissues (Figure 3B). LREs and proximal tubules additionally expressed GEP-121, a module of lysosome-associated catabolic enzymes (Figure S3B), suggesting that Megalin-associated endocytosis is accompanied by a corresponding increase in lysosomal degradation capacity in tissues other than the intestine, where it has been previously shown⁶⁸. GEP-121 was also shared with other lysosome-rich cell types, including melanophores, macrophages, microglia, and lymphatic endothelia, suggesting that lysosomal degradation capacity and protein absorption activity are separately regulated and that reused GEPs are sometimes combined to accomplish broader tasks (Figure S3B).

Similarly, mucous-secreting cells within the intestine, esophagus, and skin expressed GEP-94—a shared program of mucins and synthetic enzymes that catalyze characteristic mucin post-translational modifications, including O-glycosylation-catalyzing enzymes, GALNTs (which modify the mucin backbone), and B3GNTs and B4GNTs (which elongate glycan chains). However, intestinal goblet cells exhibited a cell type-specific adaptation—they expressed more sialtransferases (ST3Gal) than esophagus or skin mucous-secreting cells (Figure 3C). This suggests that zebrafish goblet-cell mucins may have longer sialic acid chains, similar to rats, where longer sialic acid chains are hypothesized to protect gut mucins from bacterial proteolytic enzymes⁶⁹.

Within epithelial cell types, we observed five GEPs (Figure 3D), four of which are previously known, including two tissue-specific epithelial GEPs (GEP-91 and GEP-16) and two GEPs (GEP-100 and GEP-37) comprised of traditional epithelial marker genes (e.g. epcam, annexins, occludins, claudins, and keratins). Interestingly, ‘traditional’ epithelial markers were excluded from some epithelial cell types, including the lens, liver, and retinal pigmented epithelium (RPE). However, we unexpectedly identified GEP-106 that was shared across all epithelial cell types (those that express ‘traditional’ markers and those that do not). It included adiponectin receptor 2 (adipor2), the acidic phosphorylated glycoprotein tuftelin (tuft1a), MHC-class I antigen (mhc1zba), beta-microglobulin (b2ml), enolase superfamily member 1 (enosf1), and a bZIP transcription factor (nfe2l2a) (Figure 3D). This identifies a set of genes expressed across all epithelial cells whose function is currently unclear and merits future investigation.

Altogether, these results highlight that while some cellular features are produced by re-using gene expression programs during development across multiple tissues, the catalog is actually relatively limited. In most cases, those shared programs are accompanied by cell-type-specific elaborations that customize them for each distinct cell type.

Focused analysis of zebrafish non-skeletal muscle identifies distinct pericyte subpopulations

Immense anatomical cell type diversity is present during zebrafish development, and iterative clustering within this whole-animal dataset identified a similar transcriptional diversity. However, there is not perfect correspondence between molecular and anatomical categorizations. So, we performed focused analyses within non-skeletal muscle cells and endodermal derivatives to further explore their transcriptional diversity and to better align it with functional and anatomical categorizations.

Non-skeletal muscle includes smooth muscle cells (SMCs) and pericytes, among other cell types. These cells provide structural and functional support to luminal organs, including the digestive system (visceral smooth muscle), large blood vessels (vascular smooth muscle), and capillaries (pericytes)^70–73. However, the transcriptional heterogeneity, spatial distribution, and developmental regulators of SMCs and pericytes remain areas of active investigation in multiple organisms^74,75. To further explore these enigmatic tissues in zebrafish, we iteratively re-clustered 3,866 non-skeletal muscle cells (Figure 4A, Figure S4A). Our clusters include 2 cardiac muscle populations (clusters C14 and C17), hepatic stellate cells (C18), and 2 pdgfra+ populations that we were unable to annotate (lyve1a+ or cxcl11+, C5 and C19). We identified smooth muscle based on expression of the traditional smooth muscle markers acta2 and tagln, which included five vascular SMC populations (C2, C11, C12, C15, and C21). Additionally, based on expression of the traditional visceral SMC markers desmin-b (desmb) and smoothelin-a and b (smtna and smtnb),^76,77 we identified three visceral SMC populations (C8, C10, C13) (Figure 4B) and three potentially visceral SMC populations (C16, C22, and C23) that expressed 1–2 traditional visceral SMC markers (Figure 4A, Figure S4B). Lastly, we identified three pericyte populations (C4, C9, and C20) (Figure 4A, C, Figures S4C) and a population of putative myofibroblasts (C3) (Figure S4D–G). Vascular SMCs are relatively well characterized; in this section, we focus on distinct pericyte populations and return to visceral SMCs in the next section.

Figure 4: Subclustering of non-skeletal muscle identifies distinct pericyte subtypes.

(A) UMAP projection of 3,866 non-skeletal muscle cells, numbered and color-coded by cluster. Populations further analyzed are highlighted with dotted circles (pericytes, Figure 4) and boxes (smooth muscle, Figure 5). (B) Selected differentially expressed pericyte-specific markers (x-axis) compared to vascular (vaSMCs) and visceral SMCs (viSMCs). (C) Selected differentially expressed genes (y-axis) between the three pericyte clusters (x-axis, P0–P2) and myofibroblasts compared to vascular SMCs (x-axis, vaSMCs). See Figure S4C for additional markers. (D) Expression of pericyte marker genes. (E–E’) Proportion of adma⁺ (E) and epas1a⁺ (E’) cells across the three pericyte clusters (C9, C20 and C4) (n = 172, 79, and 27 cells respectively). (F–H”) RNA in situ hybridization for ndufa4l2a (general pericyte marker) and epas1a (pericyte-2 specific marker) with immunofluorescent vascular co-stain (flk:mCherry-CAAX). Panels F–F’’: lateral view of the whole zebrafish head, G–H’’ higher magnification of hindbrain posterior cerebral vein (PCeV) with 3 (G–G”) or 1 (H–H”) epas1a⁺ pericyte(s). Arrows: ndufa4l2a⁺/epas1a⁺ pericytes near PCeV, arrowheads: ndufa4l2a⁺/epas1a^– pericytes near other hindbrain vessels, asterisks: autofluorescent red blood cells. (I) Number of ndufa4l2a⁺/epas1a⁺ pericytes per animal visible near the PCeV in a similar-sized field of view (n = 35). (J–L) RNA in situ hybridization does not identify ndufa4l2a+/epas1a+ cells near other blood vessels in the forebrain (J), eye (K), and pharyngeal arches (L). Arrowheads mark ndufa4l2a⁺ cells; no ndufa4l2a⁺/epas1a⁺ cells were observed in these regions. (M) Proportion of ndufa4l2a⁺ pericytes that were also epas1a⁺ in different regions of the zebrafish head (n = 17). Error bars indicate standard error of mean (S.E.M). Scale bar: 25 μm. See also Figures S4–S5.

Pericyte clusters were identified based on expression of previously described mouse and zebrafish marker genes—abcc9, pdgfrb and the recently described pericyte-specific (within zebrafish perivascular cells) marker, ndufa4l2a^78,79(Figure 4C–D, Figure S4C). Transcriptionally distinct pericyte subpopulations have been demonstrated in mouse⁷⁹ but not zebrafish. We observed three distinct transcriptional states or subpopulations of pericytes – pericyte-0 which does not have any unique markers, and two others with >10 additional specific markers (ure S4C). For instance, pericyte-1 expresses adrenomedullin (adma), and pericyte-2 expresses the proangiogenic factor epas1a⁸⁰ and leukocyte extravasation genes such as esama⁸¹ (Figure 4C–E). Traditional pericyte and SMC gene expression varied among these populations—pericyte-1 expressed pdgfra (sometimes considered a fibroblast marker), and the traditional SMC marker tagln was expressed in pericyte-0, but not pericyte-1 or pericyte-2 (Figure 4C). A few cells expressed pericyte-1 and pericyte-2 markers simultaneously, suggesting the two subpopulations are closely related or represent non-exclusive transcriptional states. To exclude the possibility that these subpopulations represent artefacts from incomplete dissociation, we computationally simulated pericyte-0 cell doublets with other cells, which never recapitulated the expression profiles of pericyte-1 or pericyte-2 (Figure S5A–C). Pericytes can arise from neural crest and mesodermal origins^82,83; both pericyte 1 and 2 clusters contained cells that expressed foxc1a/b and foxf2b, which are often considered markers of perivascular cells derived from cranial neural crest^84–89 (Figure S5D).

To further characterize the pericyte-2 population, we performed in situ hybridization for epas1a (pericyte-2 specific marker) and ndufa4l2a and abcc9 (general pericyte markers). We observed epas1a⁺/ndufa4l2a⁺ and epas1a⁺/abcc9⁺ cells surrounding the posterior cerebral vein in the hindbrain (Figure 4F–H, Figure S5E–G, arrows), but ndufa4l2a+ cells surrounding other hindbrain vessels did not express epas1a (Figure 4H, arrowheads). Interestingly, ndufa4l2a⁺/epas1a⁺ cells were not usually distributed along the vessel, but were observed in a similar location in many animals (Figure 4G, H, Figure S5E–F, arrows). When multiple spatially proximal ndufa4l2a⁺/epas1a⁺cells were observed, usually at least one cell was separated from the vessel (Figure 4G, I, Figure S5E, F). Other regions of the zebrafish head, including the forebrain, eye, and pharyngeal arches contained many ndufa4l2a+ cells, but almost none were epas1a+, suggesting that the pericyte-2 population is spatially restricted to the hindbrain (Figure 4J–M, Figure S5H–J‘). This mirrors other organisms, including mammals, where transcriptionally distinct pericyte subpopulations are associated with particular tissues⁷⁹. Altogether, these results demonstrate that zebrafish have multiple distinct subpopulations with pericyte transcriptional signatures that exhibit spatial segregation.

Molecular characterization of zebrafish intestinal smooth muscle cell types

The intestine is surrounded by smooth muscle that regulates its morphogenesis⁹⁰, stem cell maintenance⁹¹, enteric nervous system patterning⁹², and gastrointestinal motility^93,94. Intestinal smooth muscle (part of the visceral SMCs) is arranged in two layers, one oriented longitudinally and the other circularly⁷⁷. In humans, mice, and zebrafish, intestinal SMC markers that label both layers have been reported^90,91,95,96; however, to our knowledge, markers distinguishing between circular and longitudinal SMCs have remained undescribed. We found two desmb⁺/smtnb⁺ intestinal SMC populations in zebrafish with distinct markers. C8 expressed kcnk18, fsta, foxf2a, gucy1a1, npnt, and C10 expressed il13ra2, tesca, rgs2, fhl3b (Figure 4A, Figure 5A–B). In situ hybridization for the C8/C10 markers il13ra2 and fsta/kcnk18 demonstrated that these markers overlap with acta2 expression (which marks smooth muscle), but not with each other, confirming these represent distinct populations (Figure 5C, Figure S6A– E”‘). il13ra2⁺ cells appeared to be oriented along the length of the intestine and fsta⁺/kcnk18⁺ cells appeared oriented perpendicularly to it (Figure 5C, Figure S6A– E”‘). When viewed in cross-section, acta2:mCherry signal formed a ring around the intestine; fsta+ nuclei were oriented along the ring, either within it or contacting its inner surface (Figure 5D, arrows), while il13ra2+ nuclei primarily contacted the ring’s outer surface (Figure 5D, arrowheads), suggesting these nuclei belong to different layers. Furthermore, GFP-positive cells within F0 mosaic Tg(kcnk18 1.8kb:sfGFP) embryos confirmed that C8 cells (fsta+/kcnk18+) are oriented perpendicularly to the intestinal tract, and when viewed in cross-section wrap around the intestine circularly (Figure 5E–F, Figure S6F–G). Based on these results, we propose that C8 (fsta/kcnk18+) constitutes cells of the inner, circular layer of intestinal smooth muscle and that C10 (il13ra2+) putatively represents cells from the outer, longitudinal layer (Figure 5G), whose transcriptional differences are generally unclear.

Figure 5: Distinct gene expression within intestinal smooth muscle cell (SMC) subtypes.

(A) Top differentially expressed genes (y-axis) between intestinal SMC clusters (x-axis). (B–D) Expression of common (acta2, cald1b) and differentially expressed (il13ra2, tesca, fsta, kcnk18) intestinal SMC markers. (C–D) RNA in situ hybridization for intestinal SMC cluster-specific markers (il13ra2 and fsta) with general smooth muscle co-stain (C: acta2 in situ, D: acta2:mCherry immunofluorescence) in lateral (C) or transverse (D) view. Arrows indicate fsta⁺ cells (pink) and arrowheads indicate il13ra2⁺ cells (cyan). (E–F) Mosaic F0 transgenic labeling of C8 iSMCs with injected Tg(kcnk18 1.8kb:sfGFP) construct and SiR700 actin co-stain in lateral (E) and transverse (F) orientations (images representative of 15 fluorescence-positive larvae chosen at random). (G) Diagram of proposed longitudinal and circular intestinal SMC identities for C8 and C10. (H–I) Force-directed layout of URD-inferred transcriptional trajectory calculated on foxc1a/b^– and prrx1a/b^– (putatively non-neural crest derived) SMCs and myofibroblasts, colored by stage (H, as in Figure 1A) or gene expression (I). C: circular SMCs; L: putative longitudinal SMCs. Scale bar: 25 μm. See also Figures S6–S7.

To identify the transcriptional events accompanying the acquisition of these two SMC fates, we reconstructed developmental trajectories using URD². We focused on smooth muscle (acta2+) populations that were putatively non-neural crest derived (foxc1a/b– and prrx1a/b–), which included three visceral SMC populations, two vascular SMC populations, and putative myofibroblasts. Trajectories spanned from 3 dpf progenitors to distinct 5 dpf clusters and predicted a close relationship between the two types of intestinal SMCs (Figure 5H). We examined gene expression dynamics along the two intestinal SMC trajectory branches, identifying early onset transcription factors (TFs) that may be important for the specification of these two populations (Figure 5I, Figure S6I–K, Figure S7, Table S6). Several TFs were shared between the two intestinal SMC populations, including foxf1, foxp4, meis2a, and pbx3b (Figure S6I), though these factors were each expressed broadly in the animal. However, we found that expression of the TFs foxq1a, foxq1b, foxf2a and tcf21 was restricted to the circular fsta/kcnk18a+ SMC trajectory, while the il13ra2+ SMC cells (putatively longitudinal) instead expressed cremb, itpr1a, and tead3a (Figure 5I, Figure S6J–K). Interestingly, despite their widespread expression, previous studies in mouse⁹⁷ and Xenopus⁹⁸ have demonstrated a requirement for foxf1 and foxf2 for general intestinal SMC differentiation. Altogether, these results identified transcriptional differences between these two intestinal smooth muscle layers that may underlie or be responses to their described morphological and biophysical differences⁹⁰.

Molecular characteristics and candidate regulators of pneumatic duct and best4+ cells

Like non-skeletal muscle, we characterized the cellular heterogeneity of endodermal derivatives and inferred the transcriptional events that underlie the specification of endodermal cell types during zebrafish development. We iteratively subclustered and annotated 12,592 endodermal cells (Figure 6A, Figure S8A) and identified the transcription factors expressed by each cluster (Figure S8B). The primary source of heterogeneity varied across distinct endodermal tissues: primarily distinct cell types within the pancreas, metabolic specialization in the liver (e.g. cholesterol-biosynthesis specialized msmo1+ hepatocytes³⁷, and both anterior-posterior position and distinct cell types within the intestine.

Figure 6: Subclustering of endodermal derivatives enables molecular characterization of pneumatic duct and best4+ cells.

(A) UMAP projection of 12,592 endodermal cells, color coded and numbered by cluster. (B) Expression of specific (sftpba, sim1b) and strongly expressed (mnx1, ihha) pneumatic duct markers. (C) Top differentially expressed pneumatic duct markers (y-axis), compared to other endodermal derivatives (x-axis). (D–E”) RNA in situ hybridization for two specific pneumatic duct (pd) markers (sftpba and sim1b, D) and a swim bladder (sb) marker (slc16a3b, E). Yellow arrowhead: pneumatic duct, white arrow: anterior swim bladder bud primordium that inflates at 21 dpf. (F) Top differentially expressed markers (y-axis) in best4⁺ cells compared to other intestinal cell types. (G) Expression of general intestinal marker (cdx1b) and two best4⁺ cell markers (best4 and otop2). White arrowhead: otop2 expression in LREs, yellow arrowhead: expression within the best4⁺ cells. (H) RNA in situ hybridization for best4 and otop2. Arrowheads as in G. Scale bar: 50 μm. EC: enterocyte; prog: progenitors; LREs: lysosome-rich enterocytes; EECs: enteroendocrine cells. See also Figure S8.

Among endodermal derivatives, the least transcriptionally characterized are those that allow fish to regulate their buoyancy: the swim bladder and the pneumatic duct, which connects the swim bladder to the esophagus. We identified a cluster (C32) spanning 2–5 dpf which expressed genes previously reported in both the pneumatic duct and swim bladder (anxa5b, hb9/mnx1, ihha, shha, sox2) (Figure 6B–C), but not genes previously reported exclusively in the swim bladder (acta2, elovl1a, fgf10, has2, hprt1l, ptch1, ptch2, slc16a3b)^42,99,100 suggesting that this cluster represents the pneumatic duct. Differential gene expression testing demonstrated that surfactant protein ba (sftpba) and the transcription factor, sim1b are distinctly expressed in cluster C32 (Figure 6B–C, yellow arrowheads), compared to other endodermal derivatives. While somewhat less specific, pneumatic duct cells also expressed several additional transcription factors (arnt2, sim2, sim2.1) that may be important for their specification (Figure 6C). RNA in situ hybridization confirmed that sftpba and sim1b are expressed exclusively in the pneumatic duct and uninflated anterior swim bladder bud (Figure 6D). Co-staining with a specific marker of the swim bladder (slc16a3b) confirmed that sftpba expression is confined to the pneumatic duct at 5 and 7 dpf (Figure 6E, Figure S8C). Thus, we present specific molecular markers for the pneumatic duct, whose transcriptional profile has remained undefined despite its morphological appreciation for at least a century¹⁰¹.

Within the intestine, in addition to three general enterocyte populations, we also detected two non-canonical populations that express enterocyte markers (Figure 6A), including the lysosome-rich enterocytes that are specialized for protein absorption^67,102. Another cluster (C16) resembled human best4⁺ cells (Figure 6F-G”), a recently described cell type that is potentially reduced in inflamed intestines^14,103. These cells were also recently identified in 6 dpf and adult zebrafish by two other groups^68,104,105, but we observe them beginning at 3 dpf (Figure S8A), during their initial specification. RNA in situ hybridization identified that best4⁺ cells are located throughout the zebrafish intestine (Figure 6H), evidenced by overlap with the general intestinal marker cdx1b (Figure S8D–D”). Anterior best4⁺ cells co-expressed otop2 (Figure 6G–H, yellow arrowhead) which is also expressed in a posterior patch that is likely the LREs (Figure 6G’, H, white arrowhead). This mirrors recent human single-cell studies that decribe otop2 expression within best4⁺ cells only within particular parts of the intestine, though that location is posterior in humans (the colon) and anterior in zebrafish^106,107.

We then considered the gene expression of zebrafish best4⁺ cells more broadly. They express genes characteristic of both absorptive cells (318 genes) and secretory cells (288 genes), but during early development, secretory-associated genes are expressed more strongly (Figure S9A–D). Overall, best4+ cells are quite transcriptionally distinct — they express 36 genes that are found only within this cell type, among intestinal cells. When considering all genes that mark any intestinal population or class (secretory/absorptive), zebrafish best4⁺ cells most strongly resemble human best4⁺ cells, compared to all other cell types in the human colon (Figure 7A). Their resemblance is as strong as other zebrafish-human counterparts, including goblet cells, enteroendocrine cells, tuft-like cells, mature enterocytes, and absorptive progenitors (Figure 7A). Specific shared markers between zebrafish and human best4⁺ cells included best4, otop2, cftr, carbonic anhydrases (ca2 and ca4b instead of ca4/7), notch2, her9, and gucy2c (Figure 6F, Figure 7B–C; Figure S9E, F)^103,108. Additionally, we find that zebrafish best4⁺ cells strongly express hormones and hormone receptors, similar to human best4⁺ cells, though the cast of signaling molecules differs; zebrafish best4⁺ cells express the hormone cholecystokinin (cckb) and hormone receptors such as secretin receptor (sctr), prostaglandin E receptor 4 (ptger4c), adrenergic receptor (adra2a), and tachykinin receptor 2 (tacr2) (Figure 6F, Figure 7B, Figure S9E–F).

Figure 7: Transcriptional trajectory analysis of zebrafish best4⁺ cells and comparison to human counterparts.

(A) Transcriptome correlation between adult human colon (x-axis) and larval zebrafish intestinal cell types, based on markers of all intestinal cell types. (B) Average log-fold enrichment of genes in best4⁺ cells compared to other intestinal cells in adult human colon and small intestine (x-axis, highest enrichment in either tissue) and larval zebrafish intestines (y-axis). Genes colored black are shared, blue are human-specific, and red are zebrafish-specific best4⁺ cell markers. Figures S9 contains comparisons to colon and small intestine individually. (C) Number of genes shared between the whole transcriptome of best4+ cells in human colon¹⁰³, human small intestine¹⁰⁸, and zebrafish larval intestine. (D–E) Force-directed layout of an URD-inferred transcriptional trajectory of zebrafish intestinal cells colored by developmental stage (D) and gene expression (E). (F) Temporal dynamics of selected genes along the best4⁺ cell cascade. (G–H’’) RNA in situ hybridization of best4 and candidate transcriptional regulator pbx3a in a 5 dpf zebrafish intestine. Scale bar – 100 μm. (I) Temporal dynamics of selected genes along the posterior LRE cascade. In panels F and I, X-axis: pseudotime, Y-axis: scaled expression. See also Figures S9–S10, Table S6.

While best4⁺ cells and lysosome-rich enterocytes have now been found across multiple organisms, the transcriptional events underlying their specification is unknown. To establish candidate regulators, we reconstructed developmental trajectories among intestinal cells using URD and identified transcription factors with dynamic expression along trajectories toward these two populations² (Figure 7D). We observe a split into two progenitor populations (36–60 hpf), one associated with absorptive enterocytes (EC-1, −2, −3) and one associated with best4⁺ cells and secretory cells (EECs and goblet cells) (Figure 7D, arrowheads). In the progenitors associated with best4⁺ cells, we observed expression of the RANKL receptor tnfrsf11a and the transcription factors atoh1b, ascl1a, sox4a, and sox4b (Figure 7E–F). As best4⁺ cells become distinct, these factors decline in expression, alongside apparent Notch signaling (upregulation of notch2 and the Notch-responsive genes her2 and her15.1)(Figure 7E, Figure S9G). best4⁺ cells then express several cell type-specific TFs (including dacha, meis1b, and pbx3a), followed by several cell type-specific differentiation markers such as best4, cftr, and otop2 (Figure 7E, Figure S9G, S10A–B, Table S6). In situ hybridization confirmed that some best4⁺ cells express pbx3a (Figure 7G–H’’). In other contexts, Pbx3 and Meis1 cooperatively activate transcription^109,110, and we propose that pbx3a and meis1b may act combinatorially in best4+ cells as well.

LREs were connected to the progenitor population associated with other enterocytes. These progenitors were less distinct in their transcriptional state, but we observed an initial decline in expression of the early intestinal TFs foxd2 and satb2 (Figure 7E, I, Figure S9H). This was followed by upregulation of several TFs in multiple waves (first re-upregulation of foxd2 and satb2, followed by cdx1a and skilb, then mafa and tfeb, then mafbb and atoh1b) (Figure 7E, I, Figure S9H). Genes functionally characteristic of LREs^67,111 (cubn, dab2, amn, and lgmn) were upregulated starting along with mafa and tfeb (Figure 7I, Figure S9H). Not many genes were transiently expressed during the specification of this cell type. Expression of mafbb and tfeb was unique to LREs among intestinal cells. tfeb regulates lysosome biogenesis^112,113—the most characteristic feature of these cells—suggesting that these other TFs may also be involved in either specification of LREs or their functional specialization.

In summary, we catalogued endodermal derivatives during zebrafish development, describing gene expression within the molecularly uncharacterized pneumatic duct and multiple non-canonical intestinal populations in zebrafish, including best4+ cells. We established the molecular similarity of zebrafish and human best4+ cells and used trajectory approaches to identify putative progenitor populations and candidate developmental regulators that may be important for the specification or differentiation of these intestinal populations.

Discussion

Understanding specification and differentiation of distinct cell types benefits massively from a complete understanding of which molecular cell types/states exist, and which genes are expressed by each. In this study, we mapped gene expression landscapes with high temporal resolution during zebrafish development. Comparison across different cell types, tissues, and developmental times at the whole-transcriptome level enabled building a catalog of developmentally reused gene expression programs and identified dividing cell populations that are present for unusually long developmental times. Moreover, these data identified undescribed or rare cell types/subtypes, and enabled molecular characterization of tissues with few known marker genes, allowing us to generate testable hypotheses about crucial regulators of cellular function and cell specification. We anticipate that additional re-analyses of these data by other investigators will lead to further discoveries.

Focused analyses identified gene expression profiles of molecularly uncharacterized cellular subtypes (including the pneumatic duct and intestinal smooth muscle layers), which provide the molecular handles required for dissecting the development and function of these tissues. For instance, in the pneumatic duct, we identify cell type-specific transcription factors that may regulate its development (sim1b, sim2) and Sftpb surfactant protein, which prevents terminal airway collapse in human lungs^114,115. Other genes associated with pulmonary surfactant metabolism dysfunction (e.g. abca3)¹¹⁶ or associated with lung disease (including ceacam1, cd151, and abca12) are also expressed, suggesting that the pneumatic duct may serve as a valuable model to study these disorders and develop treatments. Similarly, this study identified transcriptional differences between intestinal SMCs, including kcnk18/fsta in circular smooth muscle and il13ra2 in putatively longitudinal smooth muscle. Interestingly, many TFs restricted to circular smooth muscle (foxq1a, foxq1b, tcf21) inhibit mammalian SMC differentiation by opposing the activities of the Foxf and Myocardin pathways⁹⁷ which we observe in both intestinal SMC types. This presents the intriguing possibility that circular smooth muscle differentiation may involve inhibition of alternative SMC fates, including longitudinal muscle, though the downstream targets and functional contribution of these TFs remains to be tested. The developmental signals that specify the intestinal smooth muscle layers are known⁹⁰, but our identification of distinct downstream gene expression will enable their genetic manipulation to test their functional contributions, developmental origins, and the regulation underlying their specification.

For some cell types, focused analysis identified molecular heterogeneity with potential functional implications. For instance, this study identifies multiple transcriptionally distinct zebrafish pericyte subpopulations. We observe pericyte-2 cells associated with a specific hindbrain vessel, suggesting that zebrafish pericytes may also exhibit tissue-dependent transcriptional differences, similar to observations that mammalian pericytes express different transmembrane transporters in the brain and lung⁷⁹. Notably, some subtype-specific genes are expressed in rodent pericytes and imply potential functional differences. For instance, the peptide hormone adrenomedullin (adma) marks pericyte-1 and is also produced by rat cerebral pericytes, where it triggers vasodilation of associated vessels¹¹⁷. Similarly, pericyte-2 markers (epas1a and esama) are expressed in mouse pericytes¹¹⁸. esama is associated with leukocyte extravasation^81,119; since these pericytes were not always in direct contact with a vessel, perhaps pericyte-2 expression contributes to particular morphological or migratory characteristics. Future long-term imaging assays will be required to determine whether these distinct transcriptional profiles represent persistent pericyte identities or transient cell states that pericytes in certain anatomical locations enter and exit, and reverse genetic approaches will be needed to test whether they perform different functions.

This study also reconstructed trajectories that describe the development of best4⁺ cells, which were recently discovered in humans^{14,103,107,108} and also noticed in zebrafish at different developmental stages^68,104,105. Decreased numbers of best4+ cells in ulcerative colitis patients and disruption of cGMP signaling in colorectal cancer^14,103,120 suggest a potential importance for best4+ cells in human disease. best4+ cells are not found in mice^121–123, so zebrafish represent an excellent opportunity to study the function and development of these cells. Human and zebrafish best4+ cells are transcriptionally similar, suggesting potential functional homology. Human and zebrafish best4+ cells share expression of the best4 and otop2 ion channels that may enable these cells to respond to luminal pH¹⁴, expression of cftr and carbonic anhydrases which may regulate ion or fluid homeostasis in the gut, and adra2a which may indicate a role in intestinal motility in both organisms¹²⁴. Human best4+ cells are the source of the critical hormone Uroguanylin/Guca2b (a guanylate-cyclase agonist peptide that increases cGMP levels to regulate satiety and intestinal tone)¹²⁵. Although zebrafish guts do not express a Uroguanylin ortholog, zebrafish best4+ cells instead produce the intestinal hormone cholecystokinin which also regulates satiety, intestinal pH, and intestinal motility in part by activating cGMP production^126–128. This suggests that cGMP production in response to pH changes or other signals may be a conserved feature, accomplished differently in humans and zebrafish. Lastly, human best4+ cells in different regions of the intestine exhibit distinct gene expression profiles^14,103,108, and in situ hybridization for otop2 demonstrated that zebrafish similarly exhibit best4+ cell regionalization. Understanding the development and function of best4+ cells would be crucial to enabling therapeutic approaches that manipulate or target them. Thus, in this study, we identify the molecular characteristics of a putative progenitor state that gives rise to best4+ cells (atoh1b/ascl1a/sox4a/b/tnfrsf11a+), identify candidate developmental signals (Notch2), identify putative transcriptional regulators (e.g. dacha, pbx3a, meis1b), and identify cell-type specific markers of best4+ cells. These lay the groundwork for experiments and genetic tools to manipulate these cells and dissect their gene regulatory network to understand their development and function in intestinal homeostasis and cell-cell communication with neighboring cell types.

Single-cell genomics has seen wide adoption among developmental biologists, and these data join complementary zebrafish scRNAseq datasets. All profile whole animals, but each study focused on different tissues or developmental processes, such as the early embryo and axial mesoderm², early embryo and pharyngeal arch⁹, the intestine and non-skeletal muscle (this work), liver and notochord³⁷, parachordal cartilage and cranial ganglia³⁸, or neuromuscular progenitors¹²⁹. This highlights the potential for the reuse and reanalysis of these data to make additional discoveries, and the need to make these data accessible and easy to work with. To this end, we created Daniocell to enable researchers to browse our data rapidly to address simple questions.

We anticipate these data can augment large-scale efforts to build models or understand mechanisms of development. For instance, we envision that label transfer approaches^130,131 used with our annotations will accelerate the annotation of future zebrafish single-cell RNAseq data. Thus, we encourage others to submit improvements to the annotations through Daniocell. Integration with data labeled via transgenes or CRISPR barcoding techniques^132,133 can ascribe lineage information to each cell population and help understand their developmental origins. While we predict candidate regulators in this work based on the timing of their expression in different cell types, we do not analyze their potential target genes or cis-regulatory elements. In the future, these data also provide a framework for unraveling the gene regulatory network underlying vertebrate development, especially when combined with single-cell chromatin accessibility assays and computational gene regulatory network (GRN) inference approaches^134,135. Thankfully, increasingly high-throughput CRISPR screening techniques mean functionally testing GRN predictions is also becoming easier¹³⁶. The greatest gains from these approaches will be realized as cooperation increases within the zebrafish community to collectively integrate the results generated across many labs using different techniques, stages, and transgenic and mutant lines to make them more broadly useable.

Limitations of the study

One limitation of this study is that while clustering cells into transcriptionally similar populations is unbiased, annotations of those populations are researcher interpretations of their identities based on their gene expression and may change based on additional information or community input through Daniocell. STAR methods describes the process we followed during annotations. Additionally, like all single-cell genomic analyses, the presented results follow from the choice of several parameters during analysis. Exact results of these analyses would change with altered parameters (e.g. exact long-term states, cell-type-specific genes, etc.), though the observed and reported trends should remain.

STAR METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact Jeffrey A. Farrell (jeffrey.farrell@nih.gov).

Materials availability

Plasmids generated in this study will be shared by the lead contact upon request. Transgenic zebrafish lines generated in this study will be shared by the lead contact upon reasonable request; in cases where international transport of the lines is too difficult due to import restrictions or other considerations, the Tol2 construct needed to recreate the line by injection may be shared instead.

Data and code availability

Daniocell is accessible at http://daniocell.nichd.nih.gov/. Sequencing data is available as FASTQs and UMI count tables under NCBI GEO accession GSE223922. Processed sequencing data in the form of a Seurat object is available on the Daniocell website. Code is available at Zenodo (https://doi.org/10.5281/zenodo.10048114) and Github (https://github.com/farrelllab/2023_Sur). Raw microscopy images are available from Mendeley Data (doi:10.17632/3378fwwm8j).

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Zebrafish (Danio rerio) used in this study include: wild-type TL/AB (Tupfel long fin/AB hybrids, generated from separately maintained TL and AB breeding stocks), casper (mpv17/roy^a9/a9; mitfa/nacre^w2/w2), and the following transgenic lines: Tg(flk:mCherry-CAAX) ¹³⁷, Tg(flk:mCherry-CAAX; pdgfrb:GFP), Tg(acta2:mCherry; TP1:GFP), and F0 mosaics of Tg(kcnk18 1.8kb:sfGFP) (generated this study). This study includes the use of live zebrafish vertebrate embryos. Animals were handled according to National Institutes of Health (NIH) guidelines. Some zebrafish work was performed at the facilities of Harvard University, Faculty of Arts & Sciences (HU/FAS) under protocol 25–08. The HU/FAS Institutional Animal Care and Use program maintains full AAALAC accreditation, is assured with OLAW (A3593–01) and is currently registered with the US Department of Agriculture (USDA). Additional zebrafish work was performed at the Eunice Kennedy Shriver National Institute of Child Health and Human Development (NICHD) Shared Zebrafish Facility, under animal protocol 20–001. The NICHD Animal Care and Use program also maintains full AAALAC accreditation, is assured with OLAW (D16–00602) and is currently registered with the US Department of Agriculture (USDA). At the developmental stages profiled in this study, zebrafish sex is not yet determined, so sex was not considered a biological variable in this study.

Breeding adults were kept at standard (28°C) temperature on a 14-hour light / 10-hour dark cycle. Most adults were on a standard light cycle (lights on fully at 8:00 or 9:00 AM), but to facilitate collection of timepoints, some adults were kept in programmable light cabinets on a shifted light cycle (lights on fully at 3:00 PM for some single-cell RNAseq samples). Those adults were given a minimum of four weeks to adapt to the shifted light cycle prior to use for breeding. Adults were fed 1–2 times daily with artemia (single-cell RNAseq) or once daily Skretting Gemma Micro 300 (all other experiments). Larvae used for collection of single-cell transcriptomes or staining were never fed.

METHOD DETAILS

Collection of embryos and larvae

For each desired collection timepoint, 3–5 breeder tanks were set up late in the afternoon with usually three male and three female TL/AB fish in each. The males and females were kept separated overnight and combined 30 minutes prior to the desired time of embryo collection after a visual inspection that fish had not been mis-sexed and laid overnight. The fish were allowed to mate and lay embryos for 30 minutes, embryos were collected, and all clutches laid for a particular collection timepoint were combined.

For daytime ‘landmark’ timepoints (e.g., 24, 48, 72, 96, 120 hpf), embryos were generally collected at 11 AM and processed at 11 AM on a following day. For nighttime ‘landmark’ timepoints (e.g., 36, 60, 84, 108 hpf), embryos were generally collected at 7 PM and processed at 7 AM on a following day. For ‘intervening’ 2-hour timepoints (e.g. 26, 28, 30, 32, 34, 36 hpf, or 38, 40, 42, 44, 46, 48 hpf), several collections were performed (9:30 AM, 11:30 AM, 1:30 PM, 3:30 PM, 5:30 PM, 7:30 PM) and all processed at once at 9:30 AM or 9:30 PM. These multi-time collections were sampled on several days for different timepoint collections. Details of breeding and collection times for each sample are listed in Table S1. For each intervening timepoint, there are a minimum of 2 technical replicates; for each ‘landmark’ timepoint, there are a minimum of 2 technical and 2 biological replicates.

Dissociation of animals into cell suspensions

Fertilized eggs were collected from wild-type (Tupfel long fin/AB) in-crosses and then cultured in blue water (5 mM NaCl, 0.17 mM KCl, 0.33 mM CaCl2, 0.33 mM MgSO4, 0.1% methylene blue) at 28°C until they reached the desired stage. Per sample, 10–12 embryos or larvae were dechorionated using forceps in calcium-free Ringer’s solution (116 mM NaCl, 2.6 mM KCl, 5 mM HEPES, pH 7.0, 500 mL, sterile) with MESAB (400 μg/mL) and left to sit 5–10 minutes at 28°C to become anesthetized. Protease solution was mixed fresh each day (0.25% trypsin, 1 mM EDTA, pH 8.0, in 1xPBS, sterile; Sigma-Aldrich T4549). Per sample, 1.2 mL of protease solution was added to a well in a 24-well plate. This plate was placed at 28°C to equilibrate temperature. Meanwhile, collagenase P solution (100 mg/ml Collagenase P in Hank’s Balanced Salt Solution; Sigma-Aldrich H9269 and 11213865001) was thawed on ice. Each sample of embryos/larvae were transferred to a 1.5mL Eppendorf tube, and the volume of Ringer’s solution was reduced to 100 μL. The animals were de-yolked and abraded by pipetting up and down 5 times with a P200 set to 80 μL. Animals were rinsed by adding 1 mL of fresh Ringer’s solution, allowing the animals to settle to the bottom, and removing all but 100 μL of the Ringer’s solution. The animals (including the 100 μL of Ringer’s) were then transferred into the 24-well plate with protease solution, 30 μL of Collagenase P solution was added per well, and the plate was swirled to mix (final volume = 1.33 mL per well). The plate was incubated at 28°C, monitoring dissociation with a microscope, pipetting up and down with a P1000 20 times slowly every 5 minutes. After 10 minutes, each sample was passed once through a P200 by pipetting from one well into another. Digestion was stopped after 20–25 minutes by adding 270 μL of 6X STOP solution (6X, 30% calf serum, 6 mM CaCl2, in 1 x PBS, sterile) to each well and swirling gently to mix (final volume per well now 1.5 mL). In many cases, some small chunks of tissue remained, but attempting to achieve complete digestion usually reduced cell viability significantly. Suspensions were filtered through a 40 μM cell filter into 1.5 mL Eppendorf tube to remove any remaining tissue chunks. The tube was flicked vigorously 20 times.

MULTI-seq barcoding of cell suspensions

Cells were barcoded using the MULTI-seq cell hashing technique ³⁵. This approach uses lipidated oligonucleotides to anchor barcoded oligonucleotides that contain a poly-A tail to cell membranes, which are captured along with cellular mRNA during reverse transcription. In this study, we used a collection of 12 such barcodes for two purposes. First, these allowed us to load more cells than standard per 10X channel. Normally low cell numbers are used to ensure that the frequency of capturing multiple cells in one droplet is low. Since such ‘multiplet’ events would be identified by the presence of multiple MULTI-seq barcodes, higher cell numbers can be loaded and multiplets later computationally removed. Second, when processing intervening 2-hour timepoints (e.g. 26, 28, 30, 32, 34, 36 hpf), larvae of those ages were dissociated in parallel, cell suspensions were barcoded in parallel (e.g. barcodes 1 and 2 for 26 hpf, barcodes 3 and 4 for 28 hpf, and so on), and then the suspensions were combined and processed into single-cell transcriptomes. Later, the MULTI-seq hashes were used to assign cells to their developmental stage. For ‘landmark’ timepoints (e.g. 24 or 36 hpf), the same cell suspension was divided across 10–12 wells, barcoded separately, and then combined—in this context, the barcode does not encode temporal information and is only for multiplet removal.

Cells were spun down at 300 x g for 3 minutes at 4°C. Supernatant was removed (leaving ~ 100 μL) and cells were resuspended in 1 mL chilled DMEM/F12 (0% BSA; Gibco 12500062). Cells were washed by spinning at 300 x g for 3 minutes at 4°C, removing supernatant, and resuspending cells in 1 mL chilled DMEM/F12 (0% BSA). During the spin, 11 μL of anchor solution was added to each well of a round-bottom plate (200 nM anchor oligo, 200 nM barcode oligo in DMEM/F12 with 0% BSA; anchor oligo was a gift from Chris McGinnis and the Zev Gartner lab, now available from Sigma-Aldrich; barcodes ordered from IDT). Cell suspensions were distributed into 2–12 wells (100 μL cells per well), mixed with 5–10 gentle pipettings, then incubated on ice 5 minutes. To lock MULTI-seq barcodes in place, 11 μL of co-anchor oligonucleotide solution (200 nM co-anchor oligo in DMEM/F12 with 0% BSA; co-anchor oligo also a gift from Chris McGinnis and the Zev Gartner lab, now available from SigmaAldrich) was added to each well, pipetted gently to mix, and incubated on ice for 5 minutes. Labeling was halted with BSA by adding 50 μL of DMEM/F12 (4% BSA) and mixing gently. To remove excess labeling reagents, cells were washed twice by spinning cells down at 300 x g for 3 minutes at 4°C, removing supernatant, and resuspending in 200 μL of DMEM/F12 (1% BSA). After the final wash, samples were pooled in a 2 mL Eppendorf, washed by spinning at 250 x g for 4 minutes, removing the supernatant, and resuspending in 1 mL of DMEM/F12 with 1% BSA, spinning again at 250 x g for 4 minutes, and then resuspending in 300 μL of DMEM/F12 with 1% BSA. Cells were filtered through a 40 μm FlowMi cell filter (Bel-Art H13680–0040) into a clean tube.

Collection of single-cell transcriptomes

Droplet emulsions of single cells were generated using the 10X Genomics Chromium controller with Single Cell 3’ v3.1 consumable reagents, according to the manufacturer’s instructions. In brief, single-cell suspensions were stained with acridine orange / propidium iodide (11 μL of cells + 1 μL of Logos F23001 Acridine Orange/Propidium Iodide Stain) and then examined and quantified on a Logos Luna FL automated cell counter. Samples with sufficient concentration, low multiplet rate (<5% to proceed, <3.6% on average), and high viability (>85% to proceed, >95% on average) were then diluted with Ringer’s to load into the instrument. Cells had been barcoded with MULTI-seq hash oligos to enable overloading of the instrument, which would then return of a larger number of transcriptomes with an increased rate of doublets that could then be removed computationally based on identification of multiple hash barcodes associated with that cell barcode. Samples were loaded into two channels at normal (9,600 cells loaded, targeting 6,000 cells recovered) and high (34,000 cells loaded, targeting 21,250 cells recovered) concentrations. Downstream reactions were performed in Biorad C1000 Touch Thermal cyclers. 10–12 cycles were used for cDNA amplification, and the result was inspected using Agilent High Sensitivity DNA Kits on the Agilent Bioanalyzer 2100. For samples barcoded with MULTI-seq, MULTI-seq hashes were isolated after cDNA amplification by performing an additional 3.2X SPRI clean-up with 1.8X isopropanol of the supernatant left after performing the standard 0.6X SPRI clean-up to recover amplified cDNA recommended by 10X Genomics. The remainder of steps were performed according to the 10X Genomics protocoo, with 8–13 cycles used for sample index PCR, based on the concentration of amplified cDNA in the previous step. MULTI-seq libraries were built according to the “MULTI-Seq Barcoding and Library Preparation Protocol (July 16, 2019)” ³⁵. Final libraries were evaluated using Agilent High Sensitivity DNA Kits on the Agilent Bioanalyzer 2100 and quantitated using a ThermoFisher Qubit 4 with dsDNA High Sensitivity (HS) reagents. Average fragment length from the Bioanalyzer and concentration from the Qubit were used to pool libraries in equimolar concentrations. Libraries were sequenced across several sequencing runs. Some samples (including samples TC1–24, TC1–48, TC1–72, TC1–96, TC2–36, TC2–60, TC2–84, and TC3–48) were checked in three separate runs on a Nextseq 500 System (Illumina) with High Output 75 cycle kits, with 28 bases for Read 1, 8 bases for Index 1, and 56 bases for Read 2. Most sequencing was performed in three separate runs on a NovaSeq 6000 Sequencing System (Illumina), using S4 full flowcells, with 28 bases for Read 1, 8 bases for Index 1, and 91 bases for Read 2. PhiX control library was spiked in at 1%. Libraries were briefly analyzed and re-pooled after the second sequencing run to try to achieve similar reads/cell across the entire data after performing the third sequencing run. Reads from all runs above were used in downstream analysis. Most sequencing runs contained mixtures of libraries to measure gene expression and MULTI-seq barcodes.

Alignment of sequencing data

Alignment of sequencing reads and processing into a digital gene expression matrix was performed using Cell Ranger version 4.0.0, including the aligner STAR version 2.5.1b, with standard parameters. The --expect-cells parameter was set to 6,000–21,250 based on the number of cells loaded per sample. The data was aligned against GRCz11 release 99 (January 2020) using the Lawson Lab Zebrafish Transcriptome Annotation version 4.3.2, published in Lawson et al. eLife 2020, available from https://www.umassmed.edu/lawson-lab/reagents/zebrafish-transcriptome/. 320 entries annotated as pseudogenes by Ensembl were removed from the reference.

Removal of MULTI-seq doublets

UMI count tables of MULTI-seq cell hashing barcodes were generated using the deMULTIplex package (available https://github.com/chris-mcginnis-ucsf/MULTI-seq/). Reads were input from FASTQ and preprocessed (deMULTIplex::MULTIseq.preProcess, cell=c(1,16), umi=c(17,28), tag=c(1,8)), aligned against the barcode sequences used in each experiment and then deduplicated into a UMI counts table (deMULTIplex::MULTIseq.align). Visual inspection on a tSNE projection (deMULTIplex::barTSNE) was used to confirm that the run had been successful and cells fell into clearly defined barcode classes.

In order to remove doublets that resulted from overloading 10X channels in barcoded samples, MULTI-seq encoded cell hashes were used to remove resultant doublets. Briefly, two calculations were used — classification based on Seurat’s hash tag oligo demultiplexing functions, and a classification based on signal-to-noise ratio. For classification by Seurat, a Seurat object was created using the MULTI-seq UMI counts matrix, normalized (Seurat::NormalizeData, assay=“MS”, normalization.method = “CLR”), and classified (Seurat::HTODemux, assay=“MS”, positive.quantile=0.9999). For classification based on signal-to-noise, cells were called as ‘negative’ if they had <20 UMIs aligned to a single barcode. To be considered a ‘singlet,’ required that the signal-to-noise ratio was ≥5, where ‘signal’ was the number of UMIs assigned to the barcode with the most UMIs and ‘noise’ was the number of UMIs assigned to the barcode with the second most UMIs. Cells with signal-to-noise ratio <5 were classified as ‘doublets.’ Cells were removed for lacking MULTI-seq cell hash information if they were called as ‘negative’ in both approaches. Cells were removed for being doublets if they were scored as a ‘doublet’ by either approach.

Normalization and quality control

Cells that were scored as singlets based on MULTI-seq cell hashing were then processed and analyzed using Seurat version 3.1.5 ¹⁵ and R version 3.6.3. First, cells were scored (Seurat::PercentageFeatureSet) for their mitochondrial gene expression (using all genes beginning mt-) and ribosomal gene expression (using the genes: rpl18a, rps16, rplp2l, rps13, rps17, rpl34, rpl13, rplp0, rpl36a, rpl12, rpl7a, rpl19, rps2, rps15a, rpl3, rpl27, rpl23, rps11, rps27a, rpl5b, rplp2, rps26l, rps10, rpl5a, rps28, rps8a, rpl7, rpl37, rpl24, rpl9, rps3a, rps6, rpl8, rpl31, rpl18, rps27.2, rps19, rps9, rpl28, rps7, rpl7l1, rps29, rpl6, rps8b, rpl10a, rpl13a, rpl39, rpl26, rps24, rps3, rpl4, rpl35a, rpl38, rplp1, rps27.1, rpl15, rps18, rpl30, rpl11, rpl14, rps5, rps21, rpl10, rps26, rps12, rpl35, rpl17, rpl23a, rps14, rpl29, rps15, rpl22, rps23, rps25, rpl21, rpl22l1, rpl36, rpl32, rps27l).

Cells were removed with either a low number of detected features (≤200 genes detected) or abnormally high number of detected features (top 0.5%), or with abnormally high mitochondrial content (≥10%). They were then log-normalized (Seurat::NormalizeData, normalization.method = “LogNormalize”, scale.factor = 10000) and scaled, regressing against mitochondrial and ribosomal gene expression (Seurat::ScaleData, vars.to.regress = c(“percent.mt”, “percent.ribo”)).

Remapping and merging of 2018 Drop-seq data

For earlier stage cells in this analysis, previously published single-cell transcriptomes from wild-type TL/AB zebrafish covering 3.3–12 hours post-fertilization from Farrell et. al 2018 were merged with the data generated in this study. First, the previous data was re-aligned to the reference used in this study and processed using Drop-seq Tools version 1.12 ¹³⁸ and its included copy of Picard Tools. Since the final step of the Drop-seq processing pipeline corrects cell barcodes to account for oligonucleotide synthesis errors that occur during the manufacture of the Dropseq beads, we started from the BAM files generated in the 2018 study, where cell barcodes had already been corrected, in order to maintain consistency with the original study. Picard Tools SamToFastq was used to create a FASTQ file from the previous BAM file, which was then used as input to STAR version 2.5.4a with the same reference that had been used with CellRanger. Picard Tools RevertSam was used to remove the previous alignment from the 2018 BAM, then both the 2018 BAM and output of STAR were sorted into queryname order using Picard Tools SortSam (SO=queryname). The remaining steps were standard application of Dropseq Tools. The two BAM files were merged with PicardTools MergeBamAlignment, tagged with gene exon information using Dropseq Tools TagReadWithGeneExon and a digital gene expression matrix was produced using Dropseq Tools DigitalExpression with NUM_CORE_BARCODES = 12000. The digital gene expression matrices were then combined and trimmed to match exactly the cells included in the original 2018 study.

Using Seurat version 4.1.0, separate Seurat objects were created for the remapped 2018 Dropseq dataset ² and the 10X dataset generated in this study. These objects were combined using the “merge” command. For visualization, we identified the top 2000 variable genes (Seurat::FindVariableFeatures, selection.method = “vst”, nfeatures = 2000), performed PCA (Seurat::RunPCA), and identified significant PCs (Seurat::JackStraw, dims=100). A Uniform Manifold Approximation and Projection (UMAP) was calculated using 50 nearest neighbors and the 30 most significant PCs (Seurat::RunUMAP, n.neighbors = 50). For clustering, we used an iterative approach. First a broad clustering (“top clusters”) was generated by identifying the top 1500 variable genes (Seurat::FindVariableFeatures, selection.method = “vst”), performing PCA, and clustering using the Leiden approach on the top 30 PCs (Seurat::FindClusters, algorithm = 4, resolution = 0.1, n.start = 50, random.seed = 17), resulting in 25 clusters. Within each top cluster, sub-clusters were determined using a similar approach, by calculating the top 2000 variable genes, performing PCA, identifying the significant PCs, and performing Leiden clustering at multiple resolutions (Seurat::FindClusters, algorithm=4, resolution = c(0.75, 1, 2, and 3), n.start = 50, random.seed = 17). Markers were calculated for each subclustering and roughly annotated subclusters. The different resolutions were assessed manually, and a resolution was chosen based on which clustering seemed the most biologically relevant. Most often, clusters obtained at resolution 2 and 3 were found to best capture cell type differences within each subclustering, but within some less-complex tissues (e.g. the primordial germ cells), lower resolutions were more appropriate. In order to make the “top clusters” biologically relevant and represent individual tissues within the fish (for instance, “endoderm”), some subclusters were manually reassigned to different top clusters based on their cell type annotations, and some top clusters were manually split. The subclustering procedure was repeated, and final subcluster resolutions were chosen again based on biological relevance. Additional manual curation was performed by manually splitting some subclusters that represented more than one cell type based on their expressed genes and/or prior knowledge from the literature. Additionally, subclusters without sufficient differentially expressed genes were combined in order to ensure that each subcluster was sufficiently distinct: marker genes for individual clusters were identified using ROC and Wilcoxon Rank Sum Tests using the command Seurat::FindMarkers(test.use = “roc”/”wilcox”, min.pct = 0.25, logfc.threshold = 0.25) and subclusters without at least 3 differentially expressed genes were merged. While heavily manually curated, our overall goal was to represent the molecular heterogeneity of cell types recovered in our data while also representing the known biology of zebrafish development to the best of our ability. This final clustering comprised 19 tissue subsets (top clusters), which contained a total of 521 subclusters.

Cluster annotations

The final clusters were annotated after consulting several sources of information such as published RNA in situ hybridizations, published single-cell RNAseq studies, public repositories such as ZFIN, and elsewhere. Top level clusters / tissue-level clusters were annotated based on broadly previously described markers—for example, muscle cells were categorized based on expression of myod1, blood cells based on expression of hemoglobins (hbbe1.1, hbbe3 etc.), fin populations were identified by expression of and1, and2, and and3, and so on. Within these tissue specific categories, for each cluster, differential expression was performed at different levels – i.e., between other clusters within a specific tissue (e.g., intestinal clusters against each other within the intestine), within the same tissue subset (intestinal cell types against all other endodermal clusters), and across the whole dataset (i.e., intestinal clusters against all other non-endodermal cell types). Using this approach, specific markers were identified for each cluster among related cell types within a tissue as well as across the whole dataset. These markers were then aligned with published RNA in situ hybridizations available in public repositories such as ZFIN or more targeted scRNAseq datasets that have been annotated by experts to denote the identity of our clusters. For some clusters where zebrafish data was unavailable, we consulted published literature and scRNAseq datasets on human, mouse and rat. In many cases, in addition to using top differentially expressed markers, we also investigated the expression of genes that are accepted markers of particular cell types in the literature (or that have been described as distinguishing between cell types), as an orthogonal approach. Through this process, we were able to assign fairly definitive cell type identities for many clusters. However, in many other cases, the differences between clusters or the identity of clusters were less clear, which has resulted in less definitive annotations. For instance, a few clusters were detected that exhibited a mixed signature of two or more major cell identities, which we annotated as “likely doublets” that may indicate incomplete dissociation of cells during the scRNAseq cell capture step or two cell types adherent to each other. We did not remove them at this time, in case they represent real populations that we did not recognize. In other cases, clusters seemed to be similar cell types, but primarily composed of cells from different developmental stages, which we annotated as the same identity with different stage labels (such as “epidermis 24–34 hpf, posterior” and “epidermis, 36–60 hpf, posterior”). Alternatively, in other cases, clusters had fairly similar expression, but differed in expression of markers that are clearly spatially restricted. Prominent examples of this are particular keratins within the epidermis and periderm (such as krt97 or krtt1c19e, useful for differentiating head versus trunk, as in annotations “periderm – mucous producing, non-head” versus “periderm – mucous-producing, head”) or often hox genes that were helpful for assigning location (such as anterior/posterior). Some clusters were annotated as “technical” if they appeared to differ markedly in ribosomal or mitochondrial RNA content or have low UMIs from clusters with very similar gene express markers; this primarily only occurs in cell types which are very strongly over-represented, such as the periderm or epidermis. In many cases, clusters had clearly distinct gene expression profiles, but the significance of the differences were unclear to us — in these cases, we have often annotated them according to some highly expressed genes (e.g. “taste epithelium – sox8b+/id2b+” or “pharyngeal ectoderm – def6b+/wif1+/fgf24+” or “vasculature – lymphatic” versus “vasculature – lymphatic, oit3+”). Additional markers are generally available in Daniocell. Finally, several clusters have been simply labeled as “unknown” or by the group of specific markers expressed by that cluster when their identity remained elusive. Additionally, clusters that resembled each other in gene expression and stage representation were often merged and re-annotated following manual curation. Annotations are all compiled in Table S2 with a nested hierarchy containing information about germ layer/tissue/organs, broader tissue category, and finally the cell type. Some of the genes used in the annotation of each cluster are also listed in Table S2.

These annotations are made publicly available through this manuscript (Table S2) as well as Daniocell that can be easily assessed by everyone. For each cluster, the most specific markers and the most highly expressed markers are reported. As these annotations solely rely on our judgement and expertise, we have also created a pipeline for users of Daniocell to provide feedback on our annotations or submit improvements and corrections. We plan to periodically update the Daniocell website and continue to make updated versions of the annotations available there, based on such submissions.

Identifying short-term and long-term transcriptional states during development

To identify groups of transcriptionally similar cells, we used an ε-neighborhood approach. Euclidean distances in normalized and scaled gene expression space were calculated between cells in each of the 19 tissue subsets using the stats::dist function. Genes used were the union of the highly variable genes calculated on each of the 19 tissue subsets but excluding cell cycle associated genes (listed below). In order to find an optimal ε, we assessed neighbors identified for each cell using a range of different ε neighborhood sizes in normalized gene expression space (ε = 20, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 50, 55, 60) and scaled gene expression space (ε = 65, 66, 67, 68, 69, 70, 71, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 83, 84, 85, 86, 87, 88, 89, 90, 91, 92, 93, 94, 95, 96, 97, 98, 99, 100) and chose the smallest ε that allowed most cells in the data to have ε-neighbors (56%), but whose neighbors were restricted to biologically meaningful and related cell types, based on inspection of several example tissues: muscle, eye, glia, axial mesoderm, periderm, and epidermis. Scaled gene expression was used because it applies an equal weight to each gene’s expression without allowing highly expressed genes to dominate, and an optimum ε of 81 was chosen.

Each cell was assigned a cell cycle score based on expression of genes representative of G1/S and G2/M phases of the cell cycle using the Seurat::CellCycleScoring function. Genes used to assign cell cycle scores to cells in our dataset were as follows: G1/S – mcm5, pcna, tyms, mcm7, mcm4, rrm1, ung1, gins2, mcm6, cdca7, dtl, prim1, uhrf1, cenpu, gmnn, hells, ccne2, cdc6, rfc2, polr1b, nasp, rad51ap1, wdr76, slbp, ubr7, pold3, msh2, atad2, rad51, rrm2, cdc45, exo1, tipin, dscc1, blm, casbap2, usp1, clspn, pola1, chaf1b, mrpl36, e2f8; G2M – cdk1, ube2c, birc5, top2a, tpx2, cks2, nuf2, mki67, tacc3, cenpf, smc4, ckap4, kif11, cdca3, hmgb2, ndc80, cks1b, tmpo, pimreg, ccnb2, ckap2l, ckap2, aurkb, bub1, anp32e, tubb4b, gtse1, kif20b, hjurp, jpt1, cdc20, ttk, cdc25c, kif2c, rangap1, ncapd2, dlgap5, cdca8, cdca2, ect2, kif23, hmmr, aurka, psrc1, anln, lbr, ckap5, cenpe, ctcf, nek2, g2e3, gas2l3, cbx5, cenpa. Cells were then categorized as ‘cycling’ (based on either G1/S or G2/M scores above 0) or ‘non-cycling’ (based on G1/S and G2/M scores both below 0).

Using an ε neighborhood size of 81 in scaled gene expression space, for each cell, we identified the ε-neighbors, computed the absolute value of the difference in developmental stage between the cell and each of its neighbors, and then took the mean as that cell’s “stage difference.” The “stage difference” identifies whether for a given cell, the transcriptionally similar cells are generally very similar in stage (in which case this value will be small), or transcriptionally similar cells span a wide range of developmental cells (in which case this value will be large). Cells were categorized based on their cell cycle scores and stage difference values and visualized by their developmental stages on a scatter plot as shown in Figures 2C–E, G–I, and Figure S2E–I. We categorized cells based on “stage difference” into different groups (“<24hr”, “24–36hr”, “36–48hr”, and “≥48hr”), and considered cells with a “stage difference” of ≥24 hours as “long-term” states.

For plots of long-term cycling states in Figure 2F, clusters with at least 10% of cells in a long-term cycling state were considered. For these states, we identified their ε-neighbors and the clusters to which the neighbors belonged. Long-term cycling states that exhibited ≥99% overlap (i.e. that identified 99% of another cluster as neighbors) were considered similar transcriptional states and were combined. Clusters associated with long-term cycling states and their neighbors were then aligned with our annotations. Clusters that corresponded to doublets or technical artefacts were also removed from this analysis. The bar represents the shortest time range that encompassed 80% of the “long-term” cells and their transcriptionally similar ε-neighbors.

Gene module identification

Identification of gene modules is sensitive to the noise inherent in scRNAseq data. To ameliorate this, we used a form of imputation/smoothing based on k-nearest neighbor networks ⁵⁹. Briefly, PCA was conducted on the data and the 5 nearest neighbors were identified for each cell based on the top 30 PCs using the command knn_smoothing (k=5, d=30, seed=42) from the R implementation of the Wagner, Yan, and Yanai approach.

Due to limitations on the addressable size of a matrix for the clustering package chosen, we focused on a subset of genes and downsampled the cells used as input. Genes were limited to those that were expressed in 0.1% – 75% of cells in the data to exclude genes who were too lowly expressed to produce meaningful results and to focus on genes that exhibited cell-type specificity by excluding genes that were mostly ubiquitous. To maximize retention of cellular complexity, downsampling was performed to focus on eliminating cells from overrepresented populations while preserving rare cell types and changes over time. Each cluster was divided according to its major “stage groups” (3–4 hpf, 5–6 hpf, 7–9 hpf, 10–12 hpf, 14–21 hpf, 24–34 hpf, 36–46 hpf, 48–58 hpf, 60–70 hpf, 72–82 hpf, 84–94 hpf, 96–106 hpf, 108–118 hpf, 120 hpf). Per cluster–stage-group (i.e. the cells defined by the intersection of cluster and stage group), 50 cells or 20% (whichever was larger) was retained.

Fuzzy c-means clustering was then performed using the R package Mfuzz ⁶³. Briefly, data was standardized (Mfuzz::standardise) and then clustered with fuzziness parameter 1.04 (empirically determined) and 200 modules requested (Mfuzz::mfuzz, c = 200, m = 1.04). These were then filtered for technical quality. First, modules that were extremely similar (member gene loadings had correlation >0.95) were combined by summing their member gene loadings (eliminating 48 of 200 modules). Second, intra-cluster variation in expression patterns was reduced by limiting gene memberships to core members (“α-core members” in Mfuzz parlance) that contribute the most strongly to the overall ‘expression’ of a module—gene loadings <0.2 were converted to 0. Third, any modules that had fewer than 5 core member genes were eliminated (eliminating 5 of 152 remaining modules). Finally, new cell embeddings were generated by multiplying the new gene loading matrix against the original expression data (new.cell.embedding <- Matrix::t(data.unlogged) %*% membership.adjusted).

The resultant 147 modules were analyzed for their expression across tissues and how completely individual genes within these modules were shared across tissues. Gene modules were then annotated to identify the functional roles of their constituent genes based on literature and information from public repositories such as ZFIN. Genes from each module were grouped and analyzed together based on their previously reported functions or based on the family that they belonged to. For example, in GEP-193 (Table S5), two groups of genes were identified: one associated to Megalin-mediated endocytosis and the other represented a family of SLC transporters. Gene modules that (i) represented outliers, (ii) consisted of member genes expressed only within one tissue subset, and (iii) and contained primarily technical member genes that were mitochondrial or ribosomal were excluded from our downstream analysis eliminating 57 of the resultant 147 modules.

Gene modules were calculated using all cells as well as cells constituting each tissue type. To compare the gene modules recovered between the global dataset versus the tissue groups, we calculated correlation and cosine similarity between all modules that were calculated either globally or on a particular tissue. Modules with a cosine correlation of at least 0.25 were sorted and clustered using the functions hclust(method = “ward.D2”) and cutree(h = 0.75). All modules calculated on individual tissues were either: (1) highly correlated with a gene module calculated on the global data set, or (2) not recovered in the global analysis, but also not shared with another tissue in the dataset. Thus, while this approach identified many more cell-type specific GEPs, it did not yield any additional shared GEPs that were not already captured from the global analysis.

Embryo pretreatment and fixation

Zebrafish larvae (3–5 dpf) were collected and fixed in 4% paraformaldehyde at 4°C overnight and stored in methanol at −20°C. Larvae were then rehydrated to PBST (phosphate buffered saline, 0.1% Tween-20 pH 7.3) in 3 graded steps. After rehydration, embryos were further permeabilized by a 50–55-minute proteinase K treatment (10 μg/mL in PBST), post-fixed for 20 mins with 4% paraformaldehyde at RT, and then washed with PBST (5 times).

Generation of probes for in situ hybridization of best4+ cells

For genes expressed in the intestine and best4⁺ cells, antisense best4, otop2, pbx3a and cdx1b probes were generated by in vitro transcription. Whole-larvae cDNA was generated by isolating total RNA from 3–5 dpf zebrafish larvae using the E.Z.N.A Total RNA kit (Omega, Bio-Tek INC, Cat# R6834–01) and reverse transcribing using the iScript cDNA synthesis kit (BioRad, Cat# 1708891). Fragments of coding sequences of these genes were amplified by PCR using genespecific primers as follows:

   best4: 949 bp; F: 5’–TGATGATGGTGGTCTCTGGA and R: 5’–

   CTTCCAATAGCAGCGTCCAT.

   otop2: 1011bp; F: 5’–TGATGGCTGTGACTGAGGAG and R: 5’–

   GTGGTAAACATCGGAATGCC.

   pbx3a: 958bp; F: 5’–AGCAGGACATCGGAGACATT and R: 5’–

   AACTGGACGCAGCAGAAGAT.

   cdx1b: 641bp; F: 5’–CCGTAAGACACCCAAGCCTA and R: 5’–

   CTCAGCACTACCAGGCAATG.

PCR products were then inserted into the pSC plasmid using the Agilent Strataclone Kit (Cat# 240205) to generate plasmids JFP524 (best4), JFP526 (otop2), JFP549 (pbx3a) and JFP542 (cdx1b). Plasmids were linearized with NotI (pSC-best4, pSC-otop2) or HindIII (pSC-pbx3a, pSC-cdx1b) and in vitro transcribed using T7 (pSC-pbx3a, pSC-cdx1b) or T3 polymerase (pSC-best4, pSC-otop2) and fluorescein or digoxygenin RNA labeling kits (digoxygenin: Roche, Cat# 11277073910; fluorescein: Roche, Cat# 11685619910). These reactions were cleaned using the NEB Monarch RNA Cleanup protocol (NEB, Cat# T2040L) and used as anti-sense RNA probes for fluorescent in situ hybridization.

Two-color fluorescent in situ hybridization

Tyramide-mediated fluorescent in situ hybridization was primarily used to characterize best4⁺ cells in the zebrafish intestine. Post-fixation, animals were prehybridized in hybridization buffer (50% formamide, 5X SSC, 0.1% Tween-20, 1M citric acid, pH 6.0, 50μg/mL heparin, and 500 μg/mL tRNA) for 2 hours at 70°C in a dry heat block. Hybridization was then performed at 70°C overnight in hybridization buffer with 3 ng/μL of each probe. Following hybridization, the next day, larvae were washed with the following series of buffer washes at 70°C: 75%, 50%, and 25% prehybridization buffer diluted in 2X SSC for 10 minutes each, 2X SSC for 15 minutes, then twice in 0.2X SSC for 30 minutes each. This was followed by a dilution series of 0.2X SSC:PBST (3:1, 1:1, and 1:3 PBST) for 5 minutes each at room temperature. Next, animals in PBST were rocked in 2% blocking buffer (5 g of blocking reagent, Roche, 11096176001) in 1X maleate buffer (150mM maleic acid, 100mM sodium chloride) for at least 2 hours and then incubated in anti-fluorescein-POD F_ab fragments (Roche, 11207733910) diluted at 1:400 overnight at 4°C. Upon retrieval, the antibody solution was washed off using PBST and the antibody was developed in dark for 45 minutes without agitation using a tyramide staining solution (TSA PLUS Fluorescein Reagent, Akoya Biosciences, TS-000200) diluted 1:50 in amplification buffer (1X Plus Amplification Diluent, Akoya Biosciences, FP1135). Next, the peroxidase was inactivated in 1% hydrogen for 20 minutes peroxide followed by elution of the antibody with 0.1M glycine (pH 2.2). Larvae were then washed in PBST and incubated in blocking solution for 2 hours at room temperature. The blocking solution was replaced with anti-DIG-POD F_ab fragments (Roche 11207733910) diluted 1:500 in blocking reagent. Digoxygenin staining was then developed using the Tyramide PLUS staining solution (TSA PLUS Cy3 reagent, Akoya Biosciences, TS-000202) diluted 1:50 in Amplification buffer for 45 minutes at room temperature. To counterstain DNA, Hoecsht 33342 (Invitrogen, H1399) was added at 1:1000 dilution in PBST and animals were incubated overnight at 4°C then washed 6 times with PBST for 15 minutes each.

Hybridization Chain Reaction (HCR) RNA in situ hybridization

HCR was performed for characterizing pericyte populations, intestinal smooth muscle populations, and the pneumatic duct. HCR probes were designed using the Özpolat lab probe generator ¹³⁹, available at https://github.com/rwnull/insitu_probe_generator. Probes were designed with amplifiers B1, B2, B3, and B5, skipping 10 bases from the beginning of the cDNA and by choosing the maximum poly A/T and poly G/C homopolymer length as 5. For each gene, 20 probe pairs were ordered in OPools format (Integrated DNA Technologies) and resuspended in nuclease-free water to a working concentration of 1 μM. Oligos constituting probe sets used in this study are consolidated in Table S7. Following rehydration and fixation, larvae were prehybridized in HCR probe hybridization buffer (Molecular Instruments, Lot# BPH02724) for 0.5–2 hours at 37°C with shaking at 300 rpm. Then, hybridization was performed with 1 μL of each 1 μM probe diluted in 500 μL probe hybridization buffer at 37°C with shaking at 300 rpm overnight (12–16 hours). Probes were washed off using the HCR probe wash buffer (Molecular Instruments, Lot# BPW02624) and subsequently pre-amplified with fresh hairpin amplification buffer (Molecular Instruments) for 30 mins to 2 hours at room temperature. Hairpins (3 μM; Molecular Instruments) were pre-annealed (98°C for 90 seconds and 25°C for 30 mins with a ramp rate of −0.1°C per second) to create hairpin secondary structure. Hairpin mixtures were then diluted 1:100 in hairpin amplification buffer (Molecular Instruments) and larvae were incubated for 12–16 hours at 24°C with 300 rpm shaking. Following amplification, animals were rinsed in 5X SSCT at 24°C followed by washes in PBST and, in some cases, immunostained for a fluorescent protein.

Immunostaining After HCR

Immunostaining was performed for flk::mCherry-CAAX animals to visualize blood vessel architecture in 5 dpf zebrafish larvae. After HCR, samples were further permeabilized with 0.1% Triton X-100 in phosphate buffered saline. Samples were blocked for 2 hours in blocking solution (5% Normal Donkey Serum (Jackson Labs), 10% Bovine Serum Albumin, 1% DMSO, and 0.1% Triton X-100 in phosphate buffered saline). Primary staining was performed with anti-RFP (rabbit, Rockland 600–401-379) and anti-GFP (rabbit, Invitrogen A11122) antibodies diluted 1:1000 in blocking buffer. Secondary staining was performed with goat Alexa Fluor 405-anti-rabbit (Molecular Probes, Cat#: A-31556) diluted 1:1000 in PBST.

Transgenesis and generation of F0 mosaics

Transgenic lines were established using a standard Tol2 protocol ¹⁴⁰. For the kcnk18 1.8kb:sfGFP transgene, a 1.8kb sequence ~3.6 kb upstream of the transcriptional start site was PCR amplified off of genomic DNA using primers 5’–CATCCTACGACGAGCGAAAC–3’ and 5’–GGACTCAGGAGGACAACACC–3’ and then TOPO TA cloned into a pENTR Gateway vector (Invitrogen), using the pENTR^™ 5’-TOPO TA Cloning Kit (Invitrogen, Cat#K59120) to generate plasmid JFP703 (p5E-kcnk18–1.8). JFP703 p5E-kcnk18–1.8, JFP236 pME-sfGFP, Tol2Kit302 p3E-SV40 polyA, and Tol2Kit394 pDEST-Tol2-pA2 plasmids were then recombined using LR Clonase II Plus (Invitrogen, Cat# 12538120) to generate JFP704 pTol2-kcnk18(1.8)-sfGFP-SV40polyA. Embryos were injected at the one-cell stage with 40pg Tol2 transposase mRNA (produced from Tol2Kit396 pCS2-FA-transposase) and 25pg of pTol2-kcnk18(1.8)-sfGFP-SV40polyA plasmid. Injected F0 embryos from three independent clutches were assessed between 5–6 dpf, and embryos were selected for imaging if there were GFP+ cells present in the intestinal tract. F0 injected animals were compared to uninjected control embryos, which did not exhibit similar fluorescence in the intestine. Injected mosaic F0 animals and uninjected controls were co-stained with SiR700-Actin Kit (Cytoskeleton, Inc, Cat# CY-SC013) and Hoecsht33342 (Invitrogen Cat# H1399) prior to imaging.

Microscopy and image analysis

Stained larvae were mounted in 1% low melting agarose diluted in 1X Danieau buffer (58 mM NaCl, 0.7 mM KCl, 0.4 mM MgSO₄, 0.6 mM Ca(NO₃)₂, 5 mM HEPES, pH 7.6) and imaged on either a Nikon Eclipse Ti2 inverted microscope with a Nikon DS-Ri2 camera or an inverted Zeiss LSM 880 Airyscan using a 10X air, 20X air, and 40X long working distance water/oil objective. Image processing was performed in ImageJ/Fiji ¹⁴¹ and Photoshop (Adobe). The acquired z-stacks were projected using Fiji, and brightness and contrast were set in Fiji using only a linear relationship in the lookup table. Cropping and resizing was performed using Photoshop (Adobe) and figures were assembled using Illustrator (Adobe).

Inference of developmental trajectories using URD

Transcriptional trajectories were constructed using URD v1.1.2 as previously described ² from (1) the putatively non-neural crest derived (i.e. foxc1a/foxc1b negative and prrx1a/b negative) viSMCs, vaSMCs, and myofiboblasts and (2) intestinal derivatives to determine the molecular events that occur as cells diversify and differentiate in these tissues. Cells belonging to each group were selected by cluster identities and used to create an URD object using the URD::seuratToURD2() function. Previously identified highly variable genes were maintained.

To identify and remove outlier cells, a k-nearest neighbor network was calculated between cells using Euclidean distance in gene expression space with 100 nearest neighbors. Cells were then removed based on their distance to their nearest neighbor or unusually high distances to the 20^th nearest neighbor using the function URD::knnOutliers (SMCs: x.max = 25, slope.r = 1.1, int.r = 4, slope.b = 1.85, int.b = 8; intestine: x.max = 23, slope.r = 1.2, int.r = 5, slope.b = 0.85, int.b = 7.5).

URD uses user-defined starting (‘root’) and endpoints (‘tips’) for building trajectories. The earliest stage cells from each subset (36–58 hpf for SMCs/myofibroblasts and 14–21 hpf for intestinal subtypes) were selected as the ‘root’ or starting point of the tree. Endpoints were defined from late stage cells (120 hpf from the intestinal tissues and 108–120 hpf for the SMCs/myofibroblasts), based on their cluster identity. Clusters were excluded that represented clear progenitor or precursors based on gene expression of cell cycle genes (e.g., intestine: cluster 14).

To identify cells along trajectories between the ‘root’ and each ‘tip’, first a diffusion map was calculated using the function URD::calcDM with parameters nn = 100, sigma = 12 (non-skeletal muscle) or sigma = 8.2 (intestine), which draws on the R package destiny ¹⁴². Next, pseudotime was computed using the function URD::floodPseudotime (n = 100, minimum.cells.flooded = 2). Then, biased random walks were simulated starting from each terminal using the function URD::simulateRandomWalk, with the following parameters—SMCs/pericytes: optimal cells forward = 10, max.cells.back = 20, n.per.tip = 25000, root.visits = 1, max.steps = 5000; intestine: optimal cells forward = 20, max.cells.back = 40.

Next, we fit a branching tree structure to each set of trajectories using the URD::buildTree command with the following parameters—SMCs/pericytes: divergence.method = “preference”, cells.per.pseudotime.bin = 10, bins.per.pseudotime.window = 8, p.thresh = 0.001; intestine: divergence.method = “preference”, cells.per.pseudotime.bin = 25, bins.per.pseudotime.window = 5, p.thresh = 0.1. To visualize the trajectories, we generated force-directed layouts on cells that had been robustly visited during the random walks (i.e., cells that have a minimum visitation frequency of 0.5) using the command URD::treeForceDirectedLayout function with the following parameters num.nn = 87 (intestine) or num.nn = 110 (SMCs/pericytes), cells.to.do = robustly.visited.cells, cut.outlier.cells = NULL, cut.outlier.edges = NULL, cut.unconnected.segments = 2, min.final.neighbors = 4.

Identifying gene cascades along developmental trajectories

Gene cascades were constructed using two different methods for intestinal SMCs and intestinal derivatives. For intestinal SMCs, each major group (or clade) of branches from the end of the branching tree were considered as a single entity and compared against each other pairwise to identify differentially expressed genes. In addition, using URD’s branching tree as a framework, the URD::aucprTestAlongTree() function was also used to find genes that are differential markers of each lineage within each group using the parameters: must.beat.sibs = 0.6, auc.factor = 1.1, log.effect.size = 0.4. These differentially expressed genes were further curated based on other criteria as described previously ⁵¹. All genes that had a fold change of 0.6 along the trajectory pursued and a classifier score of 1.05 were used for downstream analysis. Genes along the individual intestinal SMCs branches were normalized to the maximum observed for each gene within this tissue and ordered based on the pseudotime that they enter “peak” expression (defined as 50% higher expression than the minimum expression value). Gene expression dynamics within the intestinal SMCs were fit using smoothed spline curves using the function URD::geneSmoothFit() with parameters method = “spline”, spar = 0.5, moving.window = 5, cells.per.window = 8, pseudotime.per.window = 0.005. For the intestinal SMC trajectory, genes were further compared between the two branches to select genes that are specific to one branch or another or markers of both (i.e., this strategy was used to differentiate between intestinal SMC precursors, circular and longitudinal SMC markers). In the intestine trajectory, for each intestinal population, cells in each segment were compared pairwise with cells from each of the segment’s siblings and children using the function URD::aucprTestAlongTree(log.effect.size = 0.4, auc.factor = 0.6, max.auc.threshold = 0.85, frac.must.express = 0.1, frac.min.diff = 0, must.beat.sibs = 0.6). Genes were called differentially expressed if they were expressed in at least 10% of cells in the branch under consideration and were 0.6 times better than a random classifier for the population as determined by the area under a precision-recall curve. Gene expression of each marker gene in each trajectory within the intestine was then fit using an impulse model using the function URD::geneCascadeProcess() with parameters: moving.window = 5, cells.per.window = 18, pseudotime.per.window = 0.01. Cells in each trajectory were grouped using a moving window through pseudotime within which mean expression was calculated. Expression was then normalized to the maximum observed expression for each gene within the intestine. The parameters of the impulse model were then used to calculate the onset time for each gene and order genes according to the pseudotime of their expression onset.

Simulation of artificial pericyte doublets

To simulate artificial pericyte doublets, we mixed gene expression signatures of the general pericyte cluster (pericyte-0: C9) that did not express any unique markers with that of other cell types that expressed the characteristic markers of the two pericyte populations identified in our study (C20: pericyte-1; C4: pericyte-2). We identified the top 3 genes expressed by C20 and C4 respectively and extracted cells from the global dataset other than pericytes that strongly expressed these genes (log-normalized expression >2). Out of these cells, only ones between 60–120 hpf were used for downstream analysis as the two putative pericyte clusters consisted of cells encompassing those stages. For each group, we then randomized the pool of C9 (pericyte-0) cells and the extracted non-pericyte cells to create 5000 random cell pairs to use for creating artificial doublets. To create a mixed gene expression signature characteristic of doublets in scRNAseq, for each cell pair, we unlogged the expression data, took the mean of the pericyte-0 and non-pericyte cells’ expression, and re-calculated the log. We limited the genes used to calculate average expression to the union of the highly variable genes calculated on the global dataset and the non-skeletal muscle atlas. To understand how similar or different these doublets are to the distinct pericyte subpopulations, we calculated Euclidean distances in variable gene expression space for cells within a pericyte population (i.e. pericyte-1 to pericyte-1) and between a pericyte population and its simulated doublets (i.e. pericyte-1 to pericyte-1simulated-doublets) using the “dist” function. Simulated doublets were always more distant from the pericyte subpopulations than other cells within the population.

To further confirm that the pericyte-1 and −2 clusters are not artefacts, we further performed differential gene expression analysis between the individual pericyte clusters versus their corresponding artificial doublet signatures. For each pericyte cluster, based on the distances calculated above, we chose the nearest 5% artificial doublet cells (i.e. those most similar to the putative pericyte population) and compared their gene expression using the Seurat::FindMarkers function. Even the most similar simulated doublets did not recapitulate the expression of the pericyte-1 and pericyte-2 populations.

Comparison of gene expression between species

Gene expression comparisons were performed between all annotated clusters in the zebrafish intestine against that in a human colon ¹⁴ and a small intestine ¹⁰⁸ scRNAseq dataset. Genes specific to any zebrafish intestinal cell population (either markers of a single cell type when compared to the rest of the intestine, or markers of all absorptive/secretory cells when compared to the other class), were used, resulting in comparison using a set of 3,317 genes. Gene orthologs between zebrafish and humans were obtained using the HGNC Comparison of Orthology Predictions (available: genenames.org/tools/hcop/) and bioMart ¹⁴³ to identify zebrafish orthologs of human genes and vice versa. In cases where there were multiple zebrafish orthologs for a single human gene or multiple human orthologs for a single zebrafish gene, normalized logarithmic expression values of the orthologs were summed (unlog, sum, re-log). Gene expression of intestinal cells in zebrafish and humans were then scaled such that the mean expression across cells is 0 and variance across cells is 1. Scaled gene expression was then averaged across clusters using the R base function “aggregate” and used to calculate correlation in expression profiles between cell types using the function “cor()” to establish how transcriptionally similar they are. Since orthology predictions were not 1-to-1, orthologs and downstream correlation was calculated in both directions (human-to-zebrafish and zebrafish-to-human) and averaged.

QUANTIFICATION AND STATISTICAL ANALYSIS

For all graphs except Figure S1, error bars report mean ± S.E.M. Figure S1 is plotted with ggplot2::geom_boxplot ¹⁴⁴, where the bar represents the median, the lower and upper hinges correspond to the first and third quartiles (the 25th and 75th percentiles), and the whiskers extend from the hinge to the value most distant from the mean and no further than 1.5 * IQR from the hinge (where IQR is the inter-quartile range, or distance between the first and third quartiles). Venn diagrams were generated using the R package “bioVenn”. All computational analysis and statistical tests were performed in R. The rest of the data processing steps are described in their respective methods sections. Figure legends describe the details of the data plotted including what statistical tests were performed, significance, and sample sizes. All code used to generate the figures and plots have been deposited at Zenodo (https://doi.org/10.5281/zenodo.10048114) and Github (https://github.com/farrelllab/2023_Sur/) and software information is presented in the Key Resources table.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit anti-GFP	Invitrogen	Cat# A11122
Rabbit anti-RFP	Rockland	Cat# 600–401-379
Goat anti-rabbit Alexa 405	Invitrogen	Cat# A-31556
Goat anti-rabbit Alexa 488	Invitrogen	Cat# A-11008
anti-fluorescein-POD Fab fragments	Roche	Cat# 11207733910
anti-DIG-POD Fab fragments	Roche	Cat# 11207733910
Bacterial and Virus Strains
5-alpha competent E.coli	NewEngland Biolabs	Cat# C2987H
Chemicals, Peptides, and Recombinant Proteins
16% Paraformaldehyde	ProSciTech	Cat# C004
NEB Next High Fidelity 2X PCR Master Mix	NewEngland Biolabs	Cat #M0541L
Trypsin solution from porcine pancreas	SigmaAldrich	Cat # T4549
collagenase P	SigmaAldrich	Cat# H9269
Hank’s Balanced Salt Solution	SigmaAldrich	Cat # 11213865001
DMEM/F12	Gibco	Cat # 12500062
Propidium iodide/acridine orange	Logos Biosystems	Cat# F23001
Proteinase K	SigmaAldrich	Cat# 3115879001
Hoecsht33342	Invitrogen	Cat# H1399
Notl	NewEngland Biolabs	Cat# R0189S
HindIII	NewEngland Biolabs	Cat# R0104S
RNase A	ThermoFisher	Cat# EN0531
Blocking reagent	Roche	Cat# 11096176001
TSA PLUS Fluorescein Reagent	Akoya Biosciences	Cat# TS-000200
1X PLUS Amplification Diluent	Akoya Biosciences	Cat# FP1135
TSA PLUS Cy3 reagent	Akoya Biosciences	Cat# TS-000202
Gateway LR Clonase II Plus Enzyme Mix	Invitrogen	Cat# 12538120
SiR700-Actin Kit	Cytoskeleton, Inc	Cat# CY-SC013
Critical Commercial Assays
Chromium Single Cell 3’ Reagent Kits v3	10X Genomics	Cat # PN-1000075, PN-1000073, PN-120262
Agilent High Sensitivity DNA Kits	Agilent	Cat# 5067–4626
Monarch RNA Cleanup protocol	NewEngland Biolabs	Cat# T2040L
E.Z.N.A Total RNA Kit	Omega, Bio-Tek INC	Cat# R6834–01
iScript cDNA Synthesis Kit	BioRad	Cat# 1708891
Strataclone Kit	Agilent	Cat# 240205
Digoxygenin RNA labeling kits	Roche	Cat# 11277073910
Fluorescein RNA labeling kits	Roche	Cat# 11685619910
Hybridization chain reaction	Molecular Instruments	NA
pENTR™ 5’-TOPO TA Cloning Kit	Invitrogen	Cat# K59120
E.Z.N.A Plasmid DNA Mini Kit I	Omega, Bio-Tek INC	Cat# D6942–00S
E.Z.N.A Cycle Pure Kit	Omega, Bio-Tek INC	Cat# D6492–01
Deposited Data
Zebrafish 3–12 hpf Dropseq data	Farrell et al.2	GEO GSE106587
Human colon single-cell RNAseq data	Smillie et al.103	https://singlecell.broadinstitute.org/single_cell/study/SCP259/intra-and-inter-cellular-rewiring-of-the-human-colon-during-ulcerative-colitis#study-visualize
Human small intestine single-cell RNAseq data	Burclaff et al.108	GEO GSE185224
Raw and analyzed scRNAseq data	This study	GEO GSE223922
Daniocell: Single-cell atlas for navigation	This study	http://daniocell.nichd.nih.gov/
Original code	This study	https://doi.org/10.5281/zenodo.10048114 https://github.com/farrelllab/2023_Sur/
Experimental Models: Organisms/Strains
TL/AB	Zebrafish International Resource Center (ZIRC)	ZDB-GENO-031202–1
casper	Zebrafish International Resource Center (ZIRC)	ZDB-ALT-990423–22, ZDB-ALT-980203–444
Tg(flk:mCherry-CAAX)y171	Fujita et al.137	https://zfin.org/ZDB-TGCONSTRCT-110429-1
Tg(pdgfrb:eGFP)ncv22	Ando et al.83	https://zfin.org/ZDB-TGCONSTRCT-160609-1
Tg(acta2:mCherry)	Whitesell et al.82	https://zfin.org/ZDB-TGCONSTRCT-120508-2#summary
Tg(kcnk18 1.8kb:sfGFP)	This study	NA
Oligonucleotides
best4 F: 5’–TGATGATGGTGGTCTCTGGA – 3’; R: 5’–CTTCCAATAGCAGCGTCCAT – 3’	IDT	NA
otop2 F: 5’– TGATGGCTGTGACTGAGGAG and R: 5’– GTGGTAAACATCGGAATGCC	IDT	NA
pbx3a F: 5’– AGCAGGACATCGGAGACATT and R: 5’–AACTGGACGCAGCAGAAGAT	IDT	NA
cdx1b F: 5’– CCGTAAGACACCCAAGCCTA and R: 5’– CTCAGCACTACCAGGCAATG	IDT	NA
-3.6kb kcnk18 (1.8 kb) F: 5’–CATCCTACGACGAGCGAAAC–3’ and R: 5’–GGACTCAGGAGGACAACACC–3’	IDT	NA
MULTI-seq anchor and co-anchor nucleotides	McGinnis et al35	Oligos used were a gift from lab of Zev Gartner, now available Sigma-Aldrich LMO001
Probes for hybridization chain reaction	This study	Table S7
Recombinant DNA
P5E-kcnk18–1.8	This study	JFP703
pME-sfGFP	This study	JFP236
p3E-SV40 polyA	Tol2Kit145	302
pDEST-Tol2-pA2	Tol2Kit145	394
pTol2-kcnk18(1.8)-sfGFP-SV40polyA	This study	JFP704
pSC-best4	This study	JFP524
pSC-otop2	This study	JFP526
pSC-cdx1b	This study	JFP542
pSC-pbx3a	This study	JFP549
Software and Algorithms
Cell Ranger (v4.0.0)	10x genomics	https://support.10xgenomics.com/single-cell-gene-expression/software/pipelines/latest/using/tutorial_in
deMULTIplex	McGinnis et al35	https://github.com/chris-mcginnis-ucsf/MULTI-seq/
Drop-seq Tools v 1.12	Macosko et al.138	https://github.com/broadinstitute/Drop-seq/releases/tag/v1.12
R v(3.6.3, 4.2.0)	R core team	https://www.r-project.org/
Rstudio	Rstudio	https://rstudio.com
Seurat (version 3.1.5, 4.1.0)	Satija et al.15	https://satijalab.org/seurat/articles/get_started.html
knn_smoothing	Wagner et al.59	https://github.com/yanailab/knn-smoothing
MFuzz	Kumar et al.63	http://mfuzz.sysbiolab.eu/
HCOP	HUGO Gene Nomeclature Committee	https://www.genenames.org/tools/hcop/
URD (version 1.1.2)	Farrell et al.2	https://github.com/farrellja/URD
FIJI	Schindelin et al. 141	https://fiji.sc/
Photoshop	Adobe	https://www.adobe.com/products/photoshop.html
Illustrator	Adobe	https://www.adobe.com/products/illustrator.html
Other
FlowMi 40 μm cell filter	Bel-Art	H13680–0040
Luna FL automated cell counter	Logos	F23001
Confocal microscope	Nikon Zeiss	Eclipse Ti2 LSM 880 Airyscan
C1000 Touch Thermocycler	Biorad	C1000
Lawson Lab transcriptome (v4.3.2)	Lawson et al.36	https://www.umassmed.edu/lawson-lab/reagents/zebrafish-transcriptome/

Supplementary Material

2

Table S1: Sample collection, fish breeding information, library preparation information, and quality metrics for single-cell transcriptomes. Related to Figure 1.

3

Table S2: Compiled cluster annotations. Related to Figure 1.

Contains several levels of categorization, including cluster, tissue subset (“subset”) as in Figure S1H & Daniocell subsets, tissue (as in Figure 1, Figure 2, Daniocell dot plots), ‘celltype’ clusters (“cluster.celltype”) where similar clusters or clusters that separate by stage have been combined, and a more detailed annotation (“identity.super” and “identity.sub”). Also includes most similar anatomy annotations from ZFIN (“zfin”) when known. In addition, identifier genes that were used to help annotate each cluster are listed (“ident.gene”) and the top differentially expressed genes are listed (“roc” and “wilcox”).

4

Table S3: Genes categorized based on spatiotemporal variation. Related to Figure 1D–F. Provides lists of genes from “Housekeeping”, “Ubiquitous varying”, “Tissue-specific/restricted”, and “Cell type-specific/restricted” categories as in Figure 1D–F. Columns are labeled as such: CV.cluster.mean.log = Coefficient of variation, calculated across cluster means, and log-transformed.

5

Table S4: Genes differentially expressed in each long-term cycling state, related to Figure 2, Figure S2.

Differential gene expression analysis was performed for each long-term cycling state against the rest of the cells present in the cluster (short-term cycling, short-term non-cycling, and long-term non-cycling).

6

Table S5: Compiled annotations for each of the 147 gene expression programs calculated using fuzzy c-means clustering, related to Figure 3 and Figure S3.

7

Table S6: Genes expressed along the URD cascade for the intestinal smooth muscle subtypes (circular and putative longitudinal) and best4⁺ cells, related to Figure 5, Figure S7, Figure 7, and Figure S10. Rows indicate temporally ordered gene names and columns indicate pseudotime windows. Expression values are color coded such that high expression in red and low expression is yellow. Transcription factors along each cascade are highlighted in green and genes that are discussed in the manuscript are highlighted in purple.

8

Table S7: HCR probes used in this study to profile pericytes, myofibroblasts, intestinal smooth muscle, and pneumatic duct, related to Figure 4, Figure 5, Figure 6, Figure S4, Figure S5, Figure S6, Figure S8.

Highlights:

• Daniocell: a 0.5M-cell, 62-stage scRNAseq atlas of wild-type zebrafish development.

• An epas1a+ perivascular cell subtype associated with the posterior cerebral vein.

• Transcriptionally distinct intestinal smooth muscle populations.

• best4+ intestinal cell similarity to humans, anatomy, and developmental trajectory.