Divergence of immune cell types in chordate blood

Summary

Evolutionary adaptations often occur at the level of cell types and cellular function. Innate immune cells are a promising system for studying cell type evolution, as they are widespread across metazoans, have several conserved functions, and are under selective pressure from pathogens. However, molecular characterizations of invertebrate immune cells are limited, and it remains unclear whether invertebrate immune cell types are homologous to those in vertebrates. Here we use single-cell RNA sequencing, in situ hybridization, and live reporters to define the identity of blood cell states in Ciona robusta, a member of the tunicate subphylum, the sister group to vertebrates. We find evidence that C. robusta circulating blood contains a differentiation hierarchy with at least five major lineages, constituting more than 75% of circulating cells. The mature cell states include phagocytes, as well as cells variously expressing vanadium-binding proteins, carbonic anhydrases, pattern recognition receptors, cytokines, and complement factors. Despite the expression of homologs to vertebrate immune components, extensive divergence between tunicate and vertebrate immune cells obscures cell state homology. Altogether, this work modernizes blood cell classifications in C. robusta and extends the known repertoire of immune cells within chordates.

Graphical Abstract

Introduction

Evolutionary adaptations occur at different scales, from genes to body plans. Intermediate to these is the diversification of gene expression and cell types^1–3. Novel cell types have facilitated major changes in organismal lifestyle and physiology, as observed with vertebrates’ adaptive immune lymphoid cells and oxygen-transporting red blood cells⁴. New cell types can arise with the emergence of novel proteins^2,5, like red blood cells and hemoglobin⁶, or from changes in the expression of existing proteins^3,7.

A tissue of practical interest for studying cellular diversification is the blood and its constituent immune cells. Immune cells are broadly present across metazoans^8,9, with widely conserved cellular functions including pathogen recognition, phagocytosis, and cytotoxicity¹⁰. Immune systems affect ecosystem dynamics¹¹ and can serve as a source of antimicrobials¹². They are also under selective pressure from rapidly evolving pathogens¹³, driving evolution of cellular sensing and memory^14,15.

Our understanding of immune cell types is largely drawn from studies of vertebrates¹⁶. Less is known about invertebrate immune cells, where knowledge largely focuses on features shared with vertebrates—including homologous pattern recognition receptors (PRRs)^15,17,18, cytokines^19,20, and complement factors^21–23; as well as shared phagocytic^24–26 and cytotoxic^9,27,28 cellular functions. Some immune proteins are diversified in invertebrates, particularly PRRs^10,17. However, it is unclear whether invertebrates and vertebrates have homologous immune cell types. Immune cells in Drosophila and other invertebrates have been identified as macrophages based on shared phagocytic function^29–31, but vertebrate macrophages also play a role in antigen presentation³². For most invertebrates, we lack the molecular information to assess cell type homology. Immune cells in most species are classified by morphology^30,33,34, which may conceal heterogeneous cell states. Our understanding of these cells would benefit from detailed molecular characterization.

We focus on the blood cells of tunicates, whose last common ancestor with vertebrates lived ~550 million years ago³⁵. As our closest relatives without adaptive immunity, tunicates sit at a key juncture in immune cell evolution. Tunicate immune cells mainly reside in the blood, which is pumped by a heart through a semi-open circulatory system^36,37. Blood cells are morphologically diverse and are capable of phagocytosis, cytotoxic cell killing, foreign body encapsulation, allorecognition, and regeneration^38–41. Like most invertebrates to date, tunicate blood cells have been defined based on morphology, with limited molecular characterization^34,38.

Here, we have generated a combined transcriptomic and morphological atlas of circulating blood cells from the tunicate species Ciona robusta (Figure 1A). We find a differentiation hierarchy in C. robusta blood, and confirm the identity of progenitor and differentiated cell states in vivo using live cell reporters. We show that these blood cells are substantially more diverse than previous morphological classifications suggest. We begin to characterize several mature cell states, including cells expressing homologs of vertebrate immune effector genes. However, we find that C. robusta blood cells have diverged considerably from vertebrate blood cells, offering no clear evidence of the nature of ancestral immune cell types. Altogether, these observations extend a description of chordate immune cells beyond vertebrates and make clear the evolutionary plasticity of immune cell identity and gene expression.

Figure 1: scRNA-seq reveals a complex set of transcriptional states in C. robusta blood.

(A) Photo of adult Ciona robusta (formerly Ciona intestinalis type A⁸¹) from Queen Anne’s Battery Marina, Plymouth, UK. Image was kindly shared by John Bishop.

(B) Approach for associating blood cell morphotypes with single cell transcriptomics.

(C) Table summarizing the number of libraries, animals, and high-quality cells collected on each experimental day.

(D) UMAP of scRNA-seq data, colored by cluster.

(E) Heatmap showing each cluster’s top 20 enriched genes. Genes were chosen using the Wilcoxon rank-sum test (log2 fold-change >1, FDR <0.05), then filtered to require a large difference between their highest and second highest mean expression in a cluster (log-transformed difference above 0.5). The 20 genes with the highest log2 fold-change are plotted.

(F) The number of detected SNP profiles per cluster. A SNP profile is detected if it contributes to >0.5% of a cluster.

See also Figure S1.

Results

A single-cell transcriptomic atlas of C. robusta blood

We set out to document the cells in C. robusta blood through transcriptional profiling and imaging (Figure 1B). This informally establishes a census of “cell types”, but definitions of a “cell type” are evasive and a matter of some debate^42,43. To avoid ambiguity, we here use the term “morphotype” when referring to cells by their morphology and “cell state” when referring to gene expression state. A single cell might take on multiple cell states or morphotypes throughout its lifetime. We use “cell type” only when discussing theories related to cellular diversification, or when referring to cells in well-studied organisms which are widely referred to as cell types (e.g. T and B cells).

Previously characterized C. robusta blood cell morphotypes are diverse^34,39,44,45, including small round cells that resemble lymphocytes, hyaline ameboid-like cells, a range of granule-containing cells, and cells with features that differ strongly from vertebrate cell morphologies—morula cells (MC) and compartment cells with several large vacuoles, and signet ring cells (SRC) and unilocular refractile granulocytes (URG) that each have a single large vacuole. Classifications of tunicate blood cells vary, with studies defining six³⁹, seven^36,45, or nine^34,44,46 morphotypes. Cellular nomenclature is not standardized, with over 30 designations used across studies^34,39.

To obtain a harmonized definition of cell states, we generated a transcriptomic atlas of C. robusta blood cells. We collected circulating cells from 12 adult animals across two separate experimental days (Figure 1C) and profiled them by single-cell RNA sequencing (scRNA-seq). We refer to these circulating cells as blood cells, though they might include cells from other lineages that enter circulation, and might exclude hematopoietic cells that reside in organs^41,47,48. Collection of this data required optimization to accommodate the high salinity of C. robusta blood⁴⁹. Blood cells die when exposed to standard buffers for scRNA-seq, but cell viability can be rescued by a buffer containing mannitol, as we have reported in a separate technical paper⁴⁹. In this way, we sampled 15,575 cells post-filter, with a median of 6,138 transcripts detected per cell.

UMAP embedding of the scRNA-seq data reveals that C. robusta blood contains a highly complex mixture of transcriptional states (Figure 1D,E). We fractionated cell states to a resolution of 35 transcriptional clusters. All cell clusters are represented in data from both experimental days (Figure 1F), and they transcriptionally resemble circulating blood cells: all clusters express known markers of C. robusta blood (≥1 count per 10,000 [CP10k]), and all but 3 lack expression of cadherins and selected markers of non-blood tissues (<0.1 CP10k) (Figure S1A).

Since blood was sampled from wild-caught animals, there might be animal-specific differences in cell state resulting from differences in genotype, environment, symbionts, or infection. To evaluate the reproducibility of cell states across animals, we took advantage of natural single nucleotide polymorphisms (SNPs) within profiled mRNAs to assign cells to individual animals^50,51. We distinguished nine SNP profiles: eight representing one animal each, and a ninth representing four animals with fewer cells sampled (Figure S1B). Of the 35 clusters, 26 are represented by at least 8/9 SNP profiles (Figure 1F). The remaining clusters are represented by at least 5/9 SNP profiles and may represent environmentally-induced cell states. This SNP analysis also revealed two pairs of transcriptionally similar clusters partitioned by animal-specific differences (clusters 1/7 and 0/14.1; Figure S1C–I); we combined the cluster pairs, resulting in 33 total cell states. Thus, the number of common, transcriptionally distinct cell states in C. robusta blood is between 26 and 33—a substantially larger number than the 6–9 morphotypes in current classifications^{34,36,39,44,45}.

Linking transcriptional state to cell morphology

In vertebrates, the first scRNA-seq atlases of blood could immediately be associated with canonical cell types, owing to decades of work in purifying cell populations and identifying genes specific to each⁵². In tunicates, there is some prior information associating blood morphotypes with gene expression (see Table S1). However, these previously identified markers are insufficient to annotate clusters: several markers are not cluster specific (e.g. CrGal-a, CrGal-b), other markers of multiple morphotypes are expressed in only one cluster (e.g. CiEM-a), and many clusters do not express any previously proposed markers (Figure S2A). We therefore used microscopy to associate morphologies with new sets of markers derived from the scRNA-seq data (Figure 1B).

To characterize cell morphology free from fixation artifacts (Figure S2B), we used differential interference contrast (DIC) imaging of live blood cells ex vivo. We then fixed cells in place with paraformaldehyde and labeled select genes using hybridization chain reaction fluorescent in situ hybridization (HCR FISH)⁵³. We selected 1–2 genes to mark each cluster, along with a ubiquitously expressed reference gene (Figure S2C,D). We chose clusters with distinct markers, including those that were not ubiquitous across animals (as documented in Figure 1F). Cells were manually matched across multiple rounds of imaging, allowing association of live-cell morphologies with gene expression (Figures 2A, S2E).

Figure 2: Linking transcriptional clusters to morphotype demonstrates limitations of morphological categories.

(A) Workflow for associating live morphologies with marker expression.

(B) UMAP annotated by representative images and morphological descriptions. Fluorescence images for cluster 16 show high vs. low marker gene expression. See Data S1 for each cluster’s fluorescence images.

(C) Bottom: Heatmap showing the frequency of observed morphotypes per cluster. Top: Bar graph indicating the number of marker-positive cells counted per cluster. 16-high and 16-low refer respectively to high or low marker gene fluorescence.

Scale bars = 5 μm. Abbreviations: HA, hyaline amoebocyte; GA, granular amoebocyte; RA, refractile amoebocyte; SRC, signet ring cell; SGH, small granules hemocyte; LGH, large granules hemocyte; MC, moral cell; ICC, irregular compartment cell; URG, unilocular refractile granulocyte; Hem., hemocyte; Amoebo., amoebocyte.

See also Figure S2.

Using this approach, we annotated 24 of the 33 clusters in the scRNA-seq data (Figure 2B, Data S1). We assigned a transcriptional identity to all previously documented tunicate blood morphotypes except for orange cells, which are rare (0.7% of cells observed across 3 separate days, n=2,021 cells) and were not labeled by any marker genes tested (Figure 2B). Based on these integrated measurements, we propose an updated joint cell state and morphotype classification of C. robusta blood, summarized in Table 1 (see also Figure S3A,B, Table S2). We propose names for each cell state, incorporating both classical morphotype names and functional annotations that will be justified in subsequent results sections. Where possible, we have aligned nomenclature with prior definitions^34,36,44,45.

Table 1: C. robusta blood cell categorization.

Table showing each morphotype’s description, a representative image, previously used nomenclature, and the scRNA-seq cell states matched to it. scRNA-seq clusters are listed by name, with cluster numbers (as in Figure 1D) in square brackets. Cell state abundances are averaged across SNP profiles, and means are listed with standard error. Scale bars = 5 μm.

See Table S2 for a full conversion from cluster number to cell state name, and for per-SNP profile and per-experimental day abundance estimates. See also Figure S3.

Morphotype	scRNA-seq annotation	Abundance	Live DIC images	Previously used names
Hemoblast-like cell: round, hyaline, high nucleus to cytoplasmic ratio, large nucleolus	candidate multipotent progenitor (cMPP) [0/14.1]	13.0% ± 0.7%		hemoblast36; stem cell45; lymphocyte-like cell34,46
Round cell (RC): round, hyaline, smaller cell and nucleus than hemoblast-like	candidate lineage-restricted progenitor (cLRP-1) [1/7]pre-refractile amoebocyte (pre-RA) [4] RC [21]	14.6% ± 0.9% 6.4% ± 0.4% 1.9% ± 0.1%		lymphocytes36; lymphocyte-like cell46; hemoblast34
Round spreading cell (RSC): similar size and features to small round morphology, but shape is teardrop or slightly spreading	RSC [10]	3.6% ± 0.3%		lymphocytes36; lymphocyte-like cell46; hemoblast34
Hyaline amoebocyte (HA): spreading, hyaline, amoeboid motion suggested by shape and often directly observed	HA-1 [3] HA-2 [9] HA-3 [13] HA-4 [15] HA-5 [31]	4.3% ± 0.3% 2.1% ± 0.4% 2.6% ± 0.3% 3.5% ± 0.6% 0.24% ± 0.03%		hyaline amoebocyte34,36,45,46
Blebbing-like cell (BLC): hyaline, blebbing-like shape with small protrusions	BLC [11]	2.6% ± 0.3%		hyaline amoebocyte (?)34,36,45,46
Granular amoebocyte (GA): spreading, granulated, amoeboid motion observed, agranular region at leading edge of motion	GA [12]	2.8% ± 0.4%		granular amoebocyte45; granulocyte34,46
Large granules hemocyte (LGH): round, contains larger granules	large granules hemocyte/morula cell (LGH/MC) [16]	1.4% ± 0.2%		large granules granulocyte34,46
Small granules hemocyte (SGH): round or near-round, contains smaller granules	SGH-1 [24] SGH-2 [29]	0.83% ± 0.07% 1.0% ± 0.2%		small granules granulocyte34,46
Signet ring cell (SRC): round, large vacuole takes up most of cell, thin cytoplasm around, bulge at cite of nucleus	SRC [22]	1.0% ± 0.2%		vesicular cell36; signet ring cell34,45,46
Unilocular refractile granulocyte (URG): round, large or small, one refractile vacuole occupying most/all of cell	URG-1 (large) [19] URG-2 (small) [27] URG-3 (small) [28]	2.1% ± 0.1% 0.44% ± 0.06% 0.83% ± 0.05%		compartment cell44; unilocular refractile granulocyte34,46
Morula cell (MC): round or near-round, almost completely filled with round refractile compartments	large granules hemocyte/morula cell (LGH/MC) [16]	1.4% ± 0.2%		morula cell44; morula/compartment cell, distinguished by acidophilic stain34,46
Irregular compartment cell (ICC): round, filled with non-circular refractile compartments	ICC [18]	2.5% ± 0.1%		compartment cell (?)34,46
Refractile amoebocyte (RA): spreading, refractile compartments similar to morula cell	RA-1 [8] RA-2 [25]	4.1% ± 0.4% 1.45% ± 0.01%		refractile amoebocyte45
Orange cell: round or irregular shape, pigmented orange, many small orange compartments	not identified in the scRNA-seq dataset	–		orange cell

Table 1 and Figure 2 show several relationships between transcriptional state and morphotype. Some morphotypes consist of multiple transcriptional clusters (Figure 2C), with four clusters mapping to URGs (18, 19, 27, and 28), five to hyaline amoebocytes (HA) (3, 9, 13, 15, and 31), and five to small round cells (0/14.1, 1/7, 21, 4, 22). Despite their morphological similarity, these clusters vary in expression of hundreds of genes and express unique sets of transcription factors (Figure S3C). Conversely, ten transcriptional states each associate with more than one morphotype (clusters 0/14.1, 3, 4, 8/25, 9, 16, 18, 19, 22, and 28; Figure 2B,C). This morphological heterogeneity might align with transcriptional heterogeneity within clusters. In cluster 16, the genes KY21.Chr1.1992 and KY21.Chr8.1350 have a gradient of expression (Data S1–12a,b), and HCR FISH showed high expression in large granules hemocytes (LGH) and low expression in MCs (Figure 2B bottom right, 2C “16 (high)” vs. “16 (low)”, Data S1–12c–f). In another case, cluster 22 maps to both the small round morphology and SRCs in a range of sizes. This morphological heterogeneity is consistent with small round cells giving rise to small SRCs, which then grow to large SRCs, as was suggested seventy years ago³⁶. In total, the map of cell states to morphotypes shows that morphological definitions fail to reflect the diversity of C. robusta blood cell states, and that some morphologically distinct cells are transcriptionally similar.

These morphological annotations come with caveats. First, morphologies can depend on sample preparation; for instance, HA-3 morphologies differ across buffer conditions (Figure S2F). Second, our annotations lack histological staining, which some studies have used to distinguish morphotypes⁴⁵. Finally, our annotation is not yet comprehensive: we did not label markers for all 33 clusters, and some markers failed to label any observed cells (markers of 17, 23, 30, and 14.2; see Figure S2C). We also failed to annotate the orange cell, possibly because their pigmentation blocks some fluorescence wavelengths (Figure S2G). Alternatively, orange cells could be missing from the scRNA-seq dataset if they are sensitive to wash steps or cell encapsulation—no unannotated clusters express pigment cell markers from other tunicate species⁴¹ (Figure S2H). Nevertheless, our results offer a morphotype annotation of most cell clusters.

Observing in vivo cell behavior

To observe cell morphologies and behaviors in live animals, and to enable future in vivo analysis of cell states, we designed fluorescent reporters for six cell states: GA, HA-1, HA-2, HA-3, URG-1, and cLRP-1. We isolated intergenic DNA upstream of cell state-specific marker genes (Figure S4A). We cloned these DNA elements upstream of green fluorescent protein (GFP), introduced them into C. robusta embryos by electroporation, then grew animals until the juvenile stage, at which point the circulatory system has developed⁵⁴ (Figure 3A).

Figure 3: Live reporters label transcriptional states for in vivo observation.

(A) Approach for labeling live blood cells with fluorescent reporters.

(B) Frames from Video S1A showing an HA-3 reporter GFP+ cell (white arrowheads) circulating in a juvenile. Grey arrows indicate the heart.

(C) Table listing genes used per cell state for reporter design. Observed in vivo circulation and/or crawling is indicated with corresponding sections of Video S1.

(D) Frames from Video S1G showing an HA-3 reporter GFP+ cell crawling in a juvenile.

(E) Single-cell morphologies observed, either in vitro (HA-3, GA, and HA-1) or in vivo (URG-1, cLRP-1, and HA-2). Scale bars = 5 μm.

See also Figure S4.

For each reporter, we checked whether GFP+ cells in juvenile animals circulated. Four of the transgenes labeled cells in circulation, predicted to be HA-3, HA-1, URG-1, and cLRP-1 (Figure 3B,C, Video S1A–D). Cells labeled by the remaining two transgenes, predicted to be amoebocytes GA and HA-2, were observed crawling within the juvenile body (Figure 3C, Video S1E,F). Cells expressing HA-3 and HA-1 reporters also exhibited this motility and alternated between circulating and crawling (Figure 3C,D, Video S1G,H). Additionally, GFP+ cells for all six labeled states included the expected morphologies (Figures 3E, S4B–G). These results support the classification of observed cell states, demonstrate in vivo circulation or ameboid-like motility of some cells, and identify reporters for live analysis of several blood cell states.

Characterizing identity and hierarchy in C. robusta blood

With access to the transcriptional and morphological atlas of blood cells, we explored these cells’ functions by gene set enrichment analysis (Figure S3D,E, Table S3), trajectory analysis (Figure 4A–C), live cell tracing (Figure 4D), and functional assays (Figure 4E,F). We find evidence of a hematopoietic hierarchy in circulation. We also identify phagocytes, as well as cells expressing vanadium-binding proteins, enzymes associated with gas transport, and genes related to immune function (Figures 4G, S5). The cell states’ predicted hierarchy and functional roles are summarized in Figure 4H.

Figure 4: C. robusta blood contains a likely hematopoietic hierarchy and immune cell states.

(A) UMAPs showing expression of homologs to MKI67 (KY21.Chr2.1280) and TOP2A/B (KY21.Chr9.846).

(B) RNA Velocity of the C. robusta dataset. Arrows pointing away from or towards cMPPs are black or light blue, respectively.

(C) Coarse grain tree showing the inferred hematopoietic hierarchy.

(D) Images showing that a cMPP/cLRP reporter labels round cells in swimming larvae (left) and a morphologically diverse population after metamorphosis (right). See also Figure S4H,I and Video S1I,J.

(E) Approach for testing HA-2 and HA-1 for phagocytic function.

(F) Images showing that cells with phagocytosed zymosan express markers of HA-2 and HA-1.

(G) Heatmap showing vanadium-binding gene expression. See Table S1 for gene IDs.

(H) Schematic summarizing the predicted hematopoietic hierarchy and cell state functions. Cell size is relative to the scale bar. See Figure S3D for immune pathway enrichment, Figure S5E for AMP expression, and Figure 3C for motility.

Scale bars are 5 μm unless otherwise noted.

In this and subsequent figures, cell states present in ≤6/9 SNP profiles (see Figure S3B) which were not verified by HCR FISH (Figure 2B) are labeled with italic text.

See also Figures S3, S4, and S5.

A hematopoietic hierarchy in circulation

Unlike vertebrates, tunicates are thought to have an abundant population of circulating hematopoietic progenitors, which are proliferative and have a hemoblast-like morphology^36,39,45. We indeed find evidence of a hematopoietic hierarchy in circulation. Hemoblast-like and small round cells (clusters 0/14.1, 1/7, and 21; see Figure 2B) are part of a continuum of states enriched for markers of proliferation and for cell cycle and ribosome biogenesis pathways (Figures 4A, S3D). Based on RNA velocity analysis⁵⁵, cells appear to differentiate away from cluster 0/14.1 hemoblast-like cells (Figure 4B). We therefore labeled these cells as candidate multi-potent progenitors (cMPP) and labeled other proliferative clusters as candidate lineage-restricted progenitors (cLRP) (Tables 1, S2). cMPPs might be self-renewing or might be replenished from a hematopoietic niche not sampled in this study^41,47,48.

To infer C. robusta’s hematopoietic hierarchy, we used a k-nearest neighbor graph of single cell transcriptomes to construct a coarse-grained tree connecting cMPPs to differentiated cell states. The resulting tree (Figure 4C) connects 20 of the 33 clusters in five major lineages. Five clusters not linked to the tree form two separate continua: HA-3 and HA-5; and separately irregular compartment cells (ICC), URG-2, and a cluster whose morphology was not determined (ND-2) (Figure 4C, bottom). The remaining nine clusters—including SRC, GA, and LGH/MC—are disconnected from the rest. These cells may represent long-lived cell states that lack constitutive precursors, cell states generated in a non-circulating niche, or cells of non-hematopoietic origin. Altogether, by counting the number of hierarchy branches and separated clusters, we estimate that C. robusta blood contains at least 15 mature cell states.

To collect further evidence that cMPP/cLRP cells are progenitors, we generated a live-cell reporter, Chr4.593>GFP, from a gene broadly expressed across these clusters (Figure S4A). We used the reporter to track the earliest appearance of these cells during development, as well as their capacity to give rise to differentiated cells. In several tunicate species, blood cells arise from mesodermal progenitors forming pouch-like clusters in tailbud-stage embryos⁵⁶, and they are thought to do the same in C. robusta^54,57. We indeed observed Chr4.593>GFP+ cells in tailbud-stage mesodermal pouches, while reporters for differentiated cells HA-3 and GA labeled no cells at this stage (Figure S4H). From this point onwards, Chr4.593>GFP+ cells were continuously present. In swimming larva, the cells dispersed, with a subset migrating towards the trunk anterior (Figure S4H). One day after the start of metamorphosis, labeled cells were concentrated both at the trunk base and in an anterior organ called the stolon. A few hours later, some Chr4.593>GFP+ cells were highly motile and extravasated through the epidermis into the tunic, as has been previously described⁵⁸ (Video S1I). After metamorphosis, at the early juvenile stage, some Chr4.593>GFP+ cells circulated in the bloodstream, while others remained static in the tunic, between the ciliary gill slits, and in the dorsal region of the body (Video S1J). At this stage, Chr4.593>GFP+ cells showed a range of morphologies, including round cells, vacuolated cells, and crawling amoebocytes (Figures 4D, S4I). Some Chr4.593>GFP+ cells might be transcriptionally distinct from adult cMPP/cLRP cells. Nevertheless, these results are consistent with cMPP/cLRP-like cells developing from the embryonic mesoderm and being capable of giving rise to a diverse population of cells.

Hyaline amoebocytes HA-2 and HA-1 are phagocytes

There is broad conservation of metazoan phagocytosis and associated transcriptional regulators²⁶, and past work identified hyaline amoebocytes and granulated cells as phagocytes in C. robusta^34,44. We show here that HA-2 and its hypothesized precursor state, HA-1 (see Figure 4C), are phagocytes. HA-2 shows evidence of being phagocytic based on (i) its expression of known C. robusta markers of phagocytosis, including the transcription factor Cebpa²⁶ (Figure S5A), and (ii) images from HCR FISH experiments resembling engulfed cells (Data S1–6c,d). To test HA-1 and HA-2 for phagocytic function, we incubated blood cells with fluorescent zymosan, then labeled HA-1 and HA-2 marker genes using HCR FISH (Figure 4E). Cells with engulfed zymosan indeed expressed HA-1 and HA-2 markers, confirming that these states are phagocytes (Figures 4F, S5B,C). HA-1 and HA-2 accounted for 85% of observed zymosan-positive cells (240/282 cells), so they account for most but not all phagocytes in circulation (Figure S5D).

Hyaline amoebocytes HA-5 are likely vanadocytes

Some tunicate species contain vanadocytes, blood cells which accumulate high levels of vanadium, though these cells’ function is unclear^59,60. Vanadocytes have not been documented in C. robusta blood⁵⁹ and have been proposed to have an SRC morphology in other species⁶⁰. In the scRNA-seq data, HA-5 cells are enriched for genes encoding vanadium binding proteins⁶¹ (150 CP10k of Van1, 51 CP10k of Van3, and 21 CP10k of Van5), while SRCs do not express these genes (Figures 4G, S3E). This expression pattern suggests that HA-5 cells are vanadocytes. HA-5 cells show weak yellow pigmentation (Figure 2B, Data S1–21c), matching the +5 oxidation state of vanadium, while the +3 and +4 states (blue and green) have been identified in vanadocytes of other species⁶⁰. HA-5 might thus represent a distinct vanadocyte population.

Antimicrobial peptide expression in ICC, URG-1, and ND-1

ICC, URG-1, and ND-1 are enriched for antimicrobial peptides (AMPs) that have been predicted and confirmed in C. robusta⁶², with the most highly expressed being Ci-MAM-E in ICCs (415 CP10k), Ci-PAP-A in URG-1s (81 CP10k), and KY21.Chr13.185 in ND-1 (80 CP10k) (Figures S3E, S5E). Given ICC’s and URG-1’s large intracellular compartments, we hypothesize that these cells’ compartments store and release AMPs to kill pathogens. This hypothesis is consistent with previously documented URG cytotoxicity⁶³ and AMP localization to URG compartments⁶⁴.

GA express enzymes associated with gas exchange

A common function of blood is to facilitate gas transport. C. robusta have no red blood cells, and although globin genes have been identified⁶⁵, their blood does not store or accumulate oxygen⁵⁹. Consistent with this, we find that globin gene expression in C. robusta blood is several orders of magnitude lower than that found in vertebrate red blood cells⁵² (Figure S5F). Red blood cells also play a role in carbon dioxide transport, expressing carbonic anhydrases (CAs) which catalyze the conversion of carbon dioxide into soluble bicarbonate⁶⁶. GA is enriched for carbon dioxide transport genes (Figure S3E), including CA (42 CP10k) and SLC4A1 (0.68 CP10k) (Figure S5G). Other cell states — including HA-3, URG-2, and blebbing-like cells (BLC) — also express CAs. One or more of these cells might play a role in gas transport and respiration.

Expression of genes with homology to vertebrate immune factors

As expected, C. robusta blood cell states are enriched for immunity-related gene sets, including complement factors, cytokine and chemokine signaling pathways, PRR signaling pathways, and the contents of neutrophil granules (Figure S3D,E). Several of the immune gene families comprising these gene sets appear to have expanded in C. robusta. The complement factor C6, a single gene in vertebrates, has 13 copies across 5 chromosomal clusters in C. robusta (Table S1). Of these copies, seven are expressed in the blood, specifically in BLCs, ICCs, round spreading cells (RSCs), and ND-6 (Figure S5H). This dramatic expansion might be associated with the acquisition of novel protein functions.

The expression of C6 and some other complement factors in C. robusta blood cells does not align with characterized expression patterns in vertebrates. In humans, C6, complement factor B (CFB), and mannose-binding lectin (MBL2) are expressed and secreted by hepatocytes in the liver, rather than by blood cells^67,68. In C. robusta, all three factors are expressed by blood cells, with CFB and MBL2 expressed by HA-5 (putative vanadocytes) and GA, respectively (Figure S5H). The expression of these factors in tunicate mesoderm (blood) vs. vertebrate endoderm (liver) suggests plasticity in the cellular deployment of these genes. This aligns with plasticity documented across other phyla—CFB is expressed by circulating cells in echinoderms⁶⁹ and by the liver-like hepatic cecum in the chordate amphioxus⁷⁰.

Expression of genes with no detected vertebrate homologs

So far, we have hypothesized cell function based on expression of genes with predicted vertebrate homologs. However, in the C. robusta genome, 40% of genes have no vertebrate homolog detected by either orthology inference (OrthoFinder⁷¹) or BLAST (Table S1). Three C. robusta blood cell states are significantly further enriched for such genes (FDR<0.01, Fisher’s Exact Test): ICCs (56% of differentially expressed genes (DEGs) are non-vertebrate genes), URG-1s (69%), and refractile amoebocytes (70% in pre-RA, 60% in RA-1) (Figure S5I, genes listed in Table S1). In SRCs, 7 out of 12 highly-expressed DEGs have no detected homology (Table S1). The prevalence of non-vertebrate genes in these and other cells suggests the potential to uncover novel immune functions.

Evidence of major divergence between invertebrate chordate and vertebrate blood cells

scRNA-seq data allows us to search for homology between cell states through whole transcriptome comparison. We evaluated homology between C. robusta, human⁶⁸, and zebrafish⁷² blood cells using the Self-Assembling Manifold mapping (SAMap) algorithm⁷³ (Figure 5A). As a benchmark, we identified SAMap similarity scores (>0.3) that aligned expected homologous cell states between hematopoietic cells in human and zebrafish (species diverged 420 million years ago³⁵) (Figure S6A).

Figure 5: Divergent immune cell states between C. robusta and vertebrates.

(A) Approach for aligning datasets from C. robusta, human, and zebrafish using SAMap.

(B) SAMap similarity matrix for human vs. C. robusta. See Figure S6A for other species pairs’ similarity matrices.

(C, D) Generalized Sankey (GS) plots showing the number of DEGs shared between cell states across species. Horizontal bar widths represent the number of DEGs per cell state; solid-filled areas within these bars indicate the number of DEGs with a homolog in the other species; and connection widths indicate how many DEGs in connected cell states are homologous. A cell state of interest is shown with the top 5 most similar states in another species. Cell states of interest are (C) human neutrophils, (D, top) zebrafish neutrophils, and (D, bottom) C. robusta HA-1 and GA.

(E) Schematic for calculation of a specificity score, which is high (>1) if a cell state matches closely to a single cell state in another species, and low (~1) if it matches closely to more than one cell state.

(F) Specificity scores for human cell states when compared against C. robusta or zebrafish.

See also Figure S6.

Using this benchmark, both human and zebrafish neutrophils matched to C. robusta GA and HA-1 (phagocyte), monocytes/macrophages to BLCs, hematopoietic progenitors to cMPP and cLRP-4, and erythrocytes to pre-RA (Figures 5B, S6A). Of the remaining C. robusta cell states, 19/33 do not map to any cell state in human or zebrafish with similarity >0.3, and 7/14 human and 4/9 zebrafish cell states do not map to any state in C. robusta.

However, SAMap is liable to emphasize weak similarities driven by a small number of genes when comparing cells across large evolutionary distances. Here, SAMap identified only 3 genes co-enriched in pre-RA and vertebrate erythrocytes (Table S1), while there are hundreds of genes enriched in erythropoiesis. We therefore developed an approach using generalized Sankey (GS) plots and an associated gene overlap score (a Jaccard index) to critically evaluate shared enriched genes between species’ cell types (Figure 5C,D, see Figure 5 legend for GS plot description).

We focus on the relationship between neutrophils and C. robusta GA and HA-1, which were among the most similar based on SAMap. As a control, we first confirmed that zebrafish neutrophils are the most similar to human neutrophils by a large margin (Figure 5C top), with a Jaccard index 3.9 times higher than the second most similar zebrafish cell state; we call this ratio a specificity score (Figure 5E,F). By contrast, human neutrophils mapped evenly to multiple C. robusta cell states (Figure 5C bottom), with a specificity score of 1.03 (Figure 5F). A converse analysis likewise shows that HA-1 and GA cells, which SAMap matches to human neutrophils, are not significantly more similar to neutrophils than to other cell types (Figure 5D), with specificity scores of 1.5 and 1.3, respectively. These results suggest there is no clear homologous cell state in C. robusta to human neutrophils. Instead, genes defining neutrophil identity in humans appear mixed across C. robusta cell states, and C. robusta cell state identities appear mixed across human cells. We see similar results comparing human monocytes and erythrocytes to C. robusta blood cells (Figures 5F, S6B,C). Altogether, there appears to be no clear blood cell homology between C. robusta and vertebrates.

Evidence of divergence in hematopoietic regulation

The divergence observed between vertebrate and tunicate blood cells also extends to the transcription factors (TFs) specifying cell types. In vertebrate hematopoiesis, fate specification is mediated by the combinatorial expression of several TFs^74,75 listed in Figure 6A. Homologs of these TFs are variably expressed across C. robusta blood cells (Figure 6B). However, the patterns of TF expression are divergent between human and C. robusta based on two pieces of evidence. First, we analyzed differentially expressed TFs at branch points in the hypothesized C. robusta hierarchy (Figure 6C). These TFs include homologs of known fate regulators in mammalian hematopoiesis (Spi1/B/C, Nfkb, Cebpz, Egr, and Stat.b), but also TFs not expressed in mammalian hematopoietic cells^52,68, such as Sox14/15/21 (upregulated in HA-4) and Hey (upregulated in ND-4). Second, TFs homologous to those involved in vertebrate fate specification show diverged coexpression patterns in C. robusta (Figure 6B). In mammals, SPI1 is coexpressed with IRF8 in monocytes⁷⁵; in C. robusta, coexpression of SPI1 and IRF8 is absent. Similarly, three TFs coexpressed in neutrophils—CEBPA, GFI1, and LEF1—are expressed in three separate cell states in C. robusta. Examining all coexpression pairings of these canonical hematopoietic TFs (listed in Figure 6A) further shows that TF coexpression patterns have diverged between tunicates and vertebrates (Figure 6D left). As a control, coexpression is better conserved between humans and zebrafish (Figure 6D right).

Figure 6: Divergent transcription factor expression between C. robusta and vertebrates.

(A) TF combinations important for fate specification in mammalian hematopoiesis⁷⁵.

(B) Heatmap showing expression of C. robusta homologs to TFs in (A). Human homologs are listed in parentheses for each gene, with TFs listed in (A) bolded.

(C) Heatmaps showing differentially expressed TFs at branch points in the C. robusta hypothesized hierarchy.

(D) Scatter plot showing expression correlation of TF pairs from (A) for human vs. (left) C. robusta or (right) zebrafish. Correlation of each TF pair within each species is calculated across mean log-transformed expression in all cell states. In cases of one-to-many TF homology, multiple points are plotted, e.g. three points for C. robusta Cebpa which has 3 human homologs. Pearson correlation, R, is listed with a p-value.

(E) Schematic for assessing conservation of gene coexpression, based on methods introduced by Crow et al.⁷⁶

(F, H) Heatmaps showing changes in coexpression for (F) human E2F8 vs. C. robusta E2f and (H) human STAT5B vs. C. robusta Stat.b. Left: top 30 human coexpression partners for the TF. Right: homologs of those coexpression partners in C. robusta. TF expression is the first row of each heatmap. Coexpression partners are ordered by decreasing correlation with the C. robusta TF.

(G) Bar plot with the number of coexpression partners shared across species for E2F8/E2f (grey) and several immune-related TFs (gold). See also Figure S6.

See Table S1 for TF gene IDs.

We next systematically probed the conservation of regulatory relationships between TFs and their possible targets using the proxy of gene coexpression. Using a method established by Crow et al.⁷⁶, we quantified the degree to which coexpression relationships are maintained between species (Figure 6E). The TFs with the most highly conserved coexpression relationships include those which regulate the cell cycle, such as human E2F8 and C. robusta ortholog E2f. Out of the top 30 genes coexpressed with E2F8 in the human dataset, 24 have C. robusta homologs which remain correlated with E2f (Figure 6F). By contrast, TFs associated with mammalian hematopoiesis have much lower levels of coexpression conservation (Figure 6G). In a particularly divergent case, C. robusta Stat.b appears to be anti-correlated with several homologs of human STAT5B coexpression partners (Figure 6H). Coexpression is significantly better conserved between humans and zebrafish (Figure S6D). These results hold when conservation is quantified with a second statistical metric introduced by Crow et al.⁷⁶ (Figure S6E,F). Altogether, these results suggest that, in addition to divergence of mature cell states, the regulatory relationships governing hematopoiesis have diverged significantly between vertebrates and tunicates.

Discussion

In this study, we establish a joint transcriptional and morphological atlas of C. robusta blood cells. We introduce live reporters that allow observing cell states in vivo, and we present evidence that some differentiated cells arise from a hematopoietic hierarchy in circulation. Our atlas demonstrates that C. robusta blood cells are transcriptionally far more diverse than suggested by previously identified morphotypes. We then show that blood cells have diverged significantly since the last common ancestor of vertebrates and tunicates. This divergence is reflected in poor homology to human cell states, distinct transcription factor coexpression, and cell state–specific expression of genes that have no detected vertebrate homologs.

It was to be expected that the number of transcriptional states would exceed the number of morphotypes. In mammals, the discovery of B, T, and dendritic cells lagged decades after initial morphological classifications, and it took decades more work to identify subsets of T cells, dendritic cells, monocytes, and neutrophils^77,78. Cell state diversity reflects the existence of multiple developmental lineages, as well as differences in cell age, responses to inflammatory signaling, and infection.

Nonetheless, the diversity of cell states documented here extends well beyond that in vertebrates. This diversity suggests that C. robusta may support qualitatively different strategies for sensing, responding to, and adapting to pathogens. Such strategies are unlikely to be discovered by focusing only on features common to vertebrates. Complex, still unobserved immune cell repertoires may be common across invertebrates. Defining the roles of these cells may be fruitful: historically, major breakthroughs in immunity—including the discoveries of phagocytosis, antimicrobial peptides, and toll-like receptors—emerged from studies of invertebrates^25,79.

Our study leaves open at least two problems. First, we do little to clarify the function of identified cell states. Several functional assays exist for tunicate blood cells^24,38,41,45, and these can now be revisited while tracking transgenically labeled cells in live animals. Sessile tunicates, including the juveniles of C. robusta, are stationary and transparent, making them practical models for live immune assays. The second problem arises from the difficulty of determining gene and cell state homologies. State-of-the-art methods for identifying gene homology are imperfect⁷⁶, in our case failing to identify known homologs of toll-like receptors in C. robusta⁸⁰. Methods like SAMap seek to identify homologous cell states with flexible gene homology, but with distant species the results are hard to interpret, as seen from the GS plots introduced here. Cell state homology might be better determined by clustering cell states from many species, which would require blood cells sampled from multiple phyla. Such data can now be collected and used to test whether certain blood cell states in vertebrates and tunicates derive from shared ancestral cell types.

Finally, we reflect on what can be learnt about immune cell evolution. New cell types are thought to arise by a ‘sister cell’ mechanism: duplication of a developmental lineage, followed by divergence in the set of genes turned on in each lineage^1–3. Here we find that the divergence between C. robusta and humans is too large to identify sister cell relationships, if they exist—genes specific to human immune cell states appear split across multiple C. robusta cell states, and vice versa C. robusta cell state identity is split across human cells. It is still possible that vertebrate and tunicate blood cells descended from common ancestral cell types, or, alternatively, that they evolved from distinct ancestral populations. In either case, the divergence between vertebrate and tunicate blood cells suggests a pliability to which genes and functions can be expressed in the same cell. Altogether, this atlas of C. robusta blood provides a step towards documenting the diversity of immune systems and towards understanding how that diversity arises during evolution.

Resource Availability

Lead contact

Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Allon M. Klein (Allon_Klein@hms.harvard.edu).

Materials availability

Plasmids generated in this study are available from C. J. Pickett (cj.pickett@salve.edu).

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Ciona robusta

Ciona robusta hermaphrodite adults were collected from the wild in Carlsbad, California, USA by M-REP. They arrived in the lab <24 hours after collection, where they were kept at 17–18°C in 35–40 ppt artificial seawater and fed 6–7 days a week with 15 mL of Phyto-Feast. Blood was collected for experiments 1–14 days after arrival in the lab, during which time animals remained healthy by visual inspection.

Animals for scRNA-seq experiments were used 1 day (April 2022 collection) and 7 days (May 2023 collection) after arrival. Though the precise age of these wild-caught adults is unknown, animal size is recorded in Table: Animal metadata for scRNA-seq experiments.

Table.

Animal metadata for scRNA-seq experiments

Dissection animal ID	Length (cm)	Library	Estimated fraction of library by volume (%)	scRNA-seq animal ID, if known
Animal A	13.5	220428	71.9	animal_01 (determined by the skewed ratio of volume during cell prep and of cell counts in the scRNA-seq dataset)
Animal B	13.5	220428	14.4	animals_02-05
Animal C	10.0	220428	1.4	animals_02-05
Animal D	13.5	220428	1.4	animals_02-05
Animal E	11.5	220428	10.8	animals_02-05
Animal F	10.0	230518_1 and 230518_2	25.0 of 230518_1 25.0 of 230518_2	animal_06 (determined by being present in both May 2023 libraries)
Animal G	12.0	230518_1	25.0	(unknown) animal_07, 08, or 09
Animal H	9.5	230518_1	25.0	(unknown) animal_07, 08, or 09
Animal I	11.0	230518_1	25.0	(unknown) animal_07, 08, or 09
Animal J	11.5	230518_2	25.0	(unknown) animal_10, 11, or 12
Animal K	8.5	230518_2	25.0	(unknown) animal_10, 11, or 12
Animal L	11.0	230518_2	25.0	(unknown) animal_10, 11, or 12

METHOD DETAILS

Blood Cell Collection

To reduce movement during dissection, animals were relaxed by incubating 3–5 adults for 3–4 hours at 4 °C in 400 mL artificial seawater from the aquarium system with 5 g of 99% L-menthol crystals. Volume and weight measurements were approximate. To collect blood, animals were dissected to expose a large blood vessel adjacent to the heart and running along the endostyle (“ventral vessel” in Millar 1953³⁶). Blood was drawn from the vessel with a zero dead space tuberculin syringe with a 25G needle, then transferred to a microcentrifuge tube on ice.

Single-Cell RNA Sequencing

Cell Barcoding and Library Preparation

Methods for scRNA-seq collection for animals A–E (see Table: Animal metadata for scRNA-seq experiments), harvested in April 2022, are described in a separate technical paper⁴⁹ (the “PBS-M” sample). For the May 2023 scRNA-seq collection, samples were processed according to the same approach with just one change noted below: first, blood was harvested from the 7 animals F–L. Equal blood volumes were then pooled for animals F,G,H,I, and then again for animals F,J,K,L, such animal F contributed blood to both pools (see Table: Animal metadata for scRNA-seq experiments). Blood was then filtered through a 40 μm cell strainer and transferred to a microcentrifuge tube, which had been coated with bovine serum albumin (BSA) by washing with 0.5 mL of 10% (v./v.) BSA in Dulbecco’s phosphate-buffered saline (DPBS). This BSA in DPBS solution was not adjusted to match the isotonicity of C. robusta blood because the volume remaining in the tube when cells were added (<10 μL) was small relative to the volume of blood (100–100 μL).

Blood cells were pelleted by spinning at 800 g for 10 minutes at 4 °C in a swinging bucket centrifuge (Eppendorf Centrifuge 5810R). The supernatant was removed, and cells were resuspended in 0.7 M D-mannitol in 1X PBS (PBS-M). This PBS-M buffer was optimized for scRNA-seq of cells adapted to high salinity environments, as documented in the separate technical paper⁴⁹. Cells were then filtered through a 40 μm cell strainer once more to remove large clumps. The two filtration steps represent the only protocol changes between this May 2023 collection and the previously reported April 2022 collection.

Cell encapsulation, reverse transcription (RT), cDNA amplification, and library preparation was done with the 10X Genomics Chromium Next GEM Single Cell 3’ GEM Kit v3.1, Chromium Next GEM Chip G Single Cell Kit, and Library Construction Kit. We followed the manufacturer’s guidelines except for the preparation of the RT mix prior to cell encapsulation. The manufacturer’s protocol directs users to add nuclease-free water, followed by cell suspension; instead of adding nuclease-free water, we added an equivalent volume of PBS-M. scRNA-seq was provided by the Single Cell Core at Harvard Medical School, Boston, MA.

Sequencing

Both libraries were sequenced on an Illumina NovaSeq 6000, and 10X Genomics Cell Ranger (version 6.1.2) was used to generate a count matrix from FASTQ files. Reads were aligned to a reference genome which combined the KY21 gene models⁹³ for the HT2019 assembly⁹⁴ (downloaded from Ghost Database⁸²) with mitochondrial genes from the Ensembl KH genome assembly⁹⁵. To make it compatible with Cell Ranger, the KY21 gene model GFF3 file was converted to a GTF file using GffRead⁸⁶. Genome files compatible with Cell Ranger for the combined genome are available at https://github.com/AllonKleinLab/paper-data/tree/master/Scully_mannitol_2025/C_robusta_genome_HT2019_KY21_with_Ensembl_mito

All further analyses were done on libraries from both experimental days: one library from April 2022 (from the technical paper⁴⁹) and two from May 2023 (just described).

Cell Preparation and Imaging pre-HCR FISH

In preparation for cell imaging, 8-well polymer coverslip pretreated with poly-l-lysine (PLL) were coated with concanavalin A as follows: 150 μL of 1 mg/mL concanavalin A (ConA) in nuclease free water was added to each well to coat the bottom for at least 90 minutes or overnight; the solution was then removed, and wells were left to dry for at least 3 hours or overnight. The coverslips were then used within <1 day.

Blood cells were then collected as described above (“Blood cell collection”), with 4–5 animals contributing cells for each experiment. After collection, blood samples from all animals were kept on ice, then pooled and filtered through a 40 μm cell strainer. The cell concentration was estimated using the BioRad TC20 Automated Cell Counter, then the suspension was diluted with Ca²⁺- and Mg²⁺-free artificial seawater (CMF-ASW; 0.5 M NaCl, 9 mM KCl, 5 mM HEPES, pH 7.4) to a concentration of 2e6 cells/mL. DAPI was added to a concentration of 3 μM for identification of dead cells during subsequent imaging. Next, 150 μL of cell suspension was added to each well of the pre-treated 8-well polymer PLL+ConA coated coverslips. The coverslips were then taped to the bottom of a swinging bucket centrifuge and spun at 400g for 10 minutes at 4 °C to precipitate cells.

The coverslip wells containing live cells were then imaged on a widefield Nikon Ti2 inverted microscope with a Lumencor Sola light engine, the Perfect Focus System, and a Nikon Plan Apo VC 100x Oil 1.40 NA objective. Morphological information was collected using DIC imaging, with color information collected by illuminating cells with red, green, and blue light in series. DAPI signal was collected using excitation filter Chroma ET395/25x, dichroic mirror Chroma T425lpxr, and emission filter Chroma ET460/50m. Images were collected using a Hamamatsu Flash4.0 LT camera (6.5 μm² photodiode) with NIS-Elements image acquisition software. Imaging was carried out at room temperature, and a 9 by 9 grid of images was captured for each sample using a Nikon motorized xy-stage. Immediately after each well was imaged, cells were fixed by gently adding 150 μL of 8% paraformaldehyde (PFA) in CMF-ASW (final concentration 4% PFA in 300 μL). After a 30–45 min incubation at room temperature, PFA was washed twice with 150 μL of 1X PBS. Fixed and washed cells were imaged once more across the same fields of view. Across all samples on all days, 60–75 minutes passed between the first animal’s blood collection and cell suspension centrifugation onto the coverslip, and 45–90 minutes passed between centrifugation and the start of PFA fixation.

For most samples, we proceeded immediately to HCR FISH (see next section) on the same day. For labeling markers of clusters 0/14, 8/25, 10, 11, 12, 13, 22, and 24, we dehydrated cells with methanol and stored them at −20 °C before beginning HCR FISH. We found that methanol dehydration did not affect the quality of in situ hybridization in these cells, so this was used as a convenience to split experiments across multiple days when needed. Cells were dehydrated by washing three times for 5 minutes in 150 μL of 100% methanol chilled to −20 °C. Dehydrated cells were stored overnight or up to 5 days at −20 °C.

HCR FISH

We followed an HCR FISH protocol adapted from previously published protocols^53,96. Unless otherwise stated, a volume of 150 μL was used for all washes and incubations. For dehydrated samples, cells were rehydrated with a series of washes: 5 minutes with 75% methanol in PBST (1X PBS with 0.1% Tween-20), 5 minutes with 50% methanol in PBST, 5 minutes with 25% methanol in PBST, and then 5 5-minute washes in PBST. All samples were permeabilized with a 5–7 minute wash in 1X PBS with 0.1% Triton X-100, followed by a wash in PBST.

Next, cDNA probes were hybridized to target mRNA in each sample. Before hybridization, samples were incubated in probe hybridization buffer for 30 minutes at 37 °C. Then hybridization began by replacing the buffer with probe solution (4 nM concentration of each gene’s probe set in probe hybridization buffer) which had been preheated to 37 °C. Samples were incubated with probes for 20-25 hours at 37 °C. Probes were then removed by washing 4 times for 15 minutes at 37 °C with probe wash buffer, then washing 2 times for 5 minutes at room temperature with 5X SSCT buffer (5X SSC buffer with 0.1% Tween-20).

Fluorescence signal was then added and amplified using metastable fluorescent hairpins⁵³. These fluorescent hairpins were first snap-cooled by heating to 95 °C for 5 minutes, then cooling to room temperature for 30 min. A hairpin solution was prepared by adding snap-cooled hairpins to amplification buffer at a ratio of 1:50 for each hairpin. See Table S4 for the hairpins used for each transcriptional cluster. Amplification was done at room temperature: samples were first incubated in 5X SSCT buffer for 30 min, then incubated in the hairpin solution for 14.5–17.5 hours. Excess hairpins were removed by washing samples 5 times for 5 minutes in 5X SSCT buffer at room temperature. Finally, cells were labeled with DAPI by washing with PBST, incubating with 3 μM DAPI in PBST for 20–35 minutes, then washing twice more for 3 minutes with PBST. Samples were stored at 4 °C.

HCR FISH and DAPI fluorescence was imaged using a Yokogawa W1 spinning disk confocal on an Nikon Ti inverted microscope with the Perfect Focus System and a Nikon Plan Apo Lambda D 100x Oil 1.45 NA objective. Fluorescence was excited and collected with the lasers, dichroic mirrors, and emission filters listed in Table: Confocal imaging information. Images were acquired using a Hamamatsu ORCA-Fusion BT CMOS camera (6.5 μm² photodiode) with NIS-Elements image acquisition software. The same fields of view were imaged as for the live imaging. To achieve this, at the start of the experiment, one well on each 8-well coverslip was left empty, and an “X” was etched into the center using a diamond pen. Coordinates for each field of view were saved relative to the center of this X during live imaging, then used to find the same fields of view during HCR FISH imaging using a Prior Proscan III motorized stage. At each field of view, 9 z-slices were collected with a step size of 1 μm. DIC images were additionally captured at each field of view to aid in matching cells across rounds of imaging. Fluorescence z-stacks are displayed as maximum-intensity z-projections, and contrast limits were adjusted (identically for all cells within a sample) using a custom image analysis pipeline built in Python using napari⁸⁷ (available at https://github.com/AllonKleinLab/paper-data/tree/master/Scully_Ciona_blood_2025/HCR_FISH/napari_image_analysis_pipeline). Using this pipeline, cells were manually matched between the live images and HCR FISH fluorescence images to link marker-positive cells to live morphologies (Figure S2E).

Table.

Confocal imaging information

Fluorophore	Excitation LASER	Dichroic Mirror	Emission Filter
DAPI	80 mW 405 laser from a Nikon LUN-F XL solid state laser combiner	Di01-T405/488/568/647	ET455/50m
Alexa Fluor 488	80 mW 488 laser from a Nikon LUN-F XL solid state laser combiner	Di01-T405/488/568/647	ET525T/50m
Alexa Fluor 546	65 mW 561 laser from a Nikon LUN-F XL solid state laser combiner	Di01-T405/488/568/647	ET605/52m
Alexa Fluor 647	60 mW 640 laser from a Nikon LUN-F XL solid state laser combiner	Di01-T405/488/568/647	ET705/72m

We found that some cells tended to non-specifically accumulate HCR FISH signal (example in Figure S2E, grey arrow). We showed that this fluorescence was non-specific by using a probe set of zebrafish col5a1, which still labeled some cells (Figure S7A). This non-specific fluorescence was distinguishable from FISH-labeled mRNA in that it (1) was much brighter than HCR FISH signal, (2) tended not to have puncta but rather to have a large uniform region of fluorescence, and (3) was fully overlapping in all channels. Fluorescence matching these features was ignored when identifying marker-positive cells.

Live in vivo Imaging

Molecular Cloning

To construct reporter transgenes, HMW C. robusta genomic DNA was PCR-amplified with oligos listed in Table: Live reporter constructs. Putative enhancer fragments were restriction-digested and subcloned into Mesp>GFP (described by Davidson et al.⁸⁴) after restriction digest removal of the Mesp enhancer fragment. All transgenes were sequenced prior to use. Mesp>RFP was described previously by Davidson et al.⁸⁵.

Table.

Live reporter constructs

Name (best human hit)	Marker gene KY21 ID	Cell type (cluster)	Promoter sequence coordinates	Fwd oligo	Rev oligo
FBL	KY21.Chr4.593	cMPP, cLRPs (0/14.1, 1/7, 2, 5, 6)	Chr4:3796342..3797276	GATTTTATGTGAAACTCAGAAACTAATC	GATAACTTGAGTTTATATTCACGATCTTGATTC
SGO1	KY21.Chr8.361	HA-3 (13)	Chr8:2196169..2196850	CAGCATATTATTAAGAGAACCATATATTTATCC	CCACACGATGATAAATTAAAGCC
SLC175A	KY21.Chr4.869	cLRP-1 (1/7)	Chr4:5400469..5400793	GGTTGTTTTATCGCCATATAGTTATAATACACC	GTTGAGCTGTAAAAAAAGATCGTTAAAACG
TSC22D4	KY21.Chr2.603	URG-1 (19)	Chr2:3341917..3343007	GCTAAAATTGTTTCAAATATTTATTCAGTTAATATATTGGG	GAATTATTGTTTTTTATTTGTTCATTTGTTTCTGCAG
CA7	KY21.Chr13.352	GA (12)	Chr13:3123233..3125428	CGCGCAAATTTGATGACG	CGGGGTCAACAAGTTCG
SELL1	KY21.Chr1.1154	HA-1 (3)	Chr1:8413913..8415811	CCAGTTAGTTTTTAATGGGACATCAACG	CACAATGAATTTGCTGAGAAG
ADGRD1b	KY21.Chr2.1414	HA-2 (9)	Chr2:8797610..8800322	GCACGAACATTCCTTCTATTTAGCC	GGTGTTGACGATTTTCTTCGC

Electroporation

Electroporation procedures followed the method described by Zeller et al.⁹⁷. Transgene DNAs in water were combined and diluted in 0.77M D-mannitol not exceeding a final concentration of 125ug/mL. Embryos and juveniles were subsequently cultured at 18 °C in 0.22 μm-filtered artificial seawater (FASW) supplemented with penicillin (10 U/mL) and streptomycin (10 μg/mL).

Imaging

For in vivo imaging, transgenic C. robusta juveniles were paralyzed by incubation in menthol-treated FASW. Samples were imaged in menthol-treated FASW in glass-bottomed dishes. For ex vivo imaging, transgenic juvenile C. robusta in FASW were roughly dissociated using a plastic tube pestle. Samples were mounted on slides. Images were captured on a Zeiss LSM 980 via confocal imaging or fluorescent widefield with an Axiocam 305 camera.

Zymosan Phagocytosis Assay

To test phagocytic function of clusters HA-1 and HA-2, we incubated cells with zymosan, a yeast cell wall extract which is readily engulfed by phagocytes⁹⁸. The cells were then fixed and stained by HCR FISH with markers of HA-1/2.

pH-sensitive (pHRhodo) dye–labeled zymosan was provided as a kind gift from Lillian Horin. It was prepared by washing unlabeled zymosan 3 times with 0.1 M sodium bicarbonate, then incubating the pellet with 200 μM pHrodo amine-reactive dye in PBS with 0.02% sodium azide for 30 minutes, rocking at room temperature. The zymosan was then washed 5 more times with PBS with 0.02% sodium azide, then stored for 3 years at 30 mg/mL in PBS with 0.02% sodium azide at 4 °C. It was diluted in CMF-ASW to a concentration of 1 mg/mL shortly before being added to cells.

Next, blood was collected as described above (“Blood Cell Collection”), then mixed with diluted zymosan in a ratio of 9:1 blood:zymosan by volume. The cells were then incubated for 2 hours at 18 °C in a rocking microcentrifuge tube, followed by filtration through an 40 μm cell strainer and centrifugation onto a PLL+ConA coated coverslip for live imaging as described above (“Cell Preparation for HCR FISH with Live Imaging”). We confirmed that intracellular zymosan, but not extracellular zymosan, was fluorescent (Figure S7B). We then proceeded to label HA-1 and HA-2 marker genes as described in “HCR FISH”.

QUANTIFICATION AND STATISTICAL ANALYSIS

Single-Cell RNA-seq Processing

Initial Cell Barcode Filtration

Cell barcodes corresponding to viable cells were filtered based on (i) the number of mRNA molecules detected with unique molecular identifiers (UMI), (ii) the fraction of mitochondrial reads, and (iii) the likelihood of being a doublet. For (i), cells were kept that had more than 2,000 UMIs. This threshold was set based on clear separation from background as seen by plotting a log-scale histogram of UMIs per cell barcode for each library. For (ii), a mitochondrial fraction below 10% was required. For (iii), we used the SCRUBLET package⁸⁸ to identify likely doublets, which were removed.

Processing and Dimensionality Reduction

Preprocessing and dimensionality reduction were performed with Scanpy⁸⁹ (version 1.8.1), using default parameters except where otherwise noted. Counts were normalized to the median counts per cell across all libraries (scanpy.pp.normalize_total). Highly variable genes were identified as described in Klein et al.⁹⁹ (code available at https://github.com/AllonKleinLab/paper-data/blob/master/Scully_Ciona_blood_2025/helper_functions/highly_variable_genes.py). Counts were log transformed (scanpy.pp.log1p, base=10), z-scores were calculated (scanpy.pp.scale), then principal component (PC) analysis was performed (scanpy.tl.pca, use_highly_variable=True). Batch integration was performed using BBKNN¹⁰⁰ to generate a single-cell network, treating the three libraries as the batches to be integrated (scanpy.external.pp.bbknn, batch_key=‘library’, neighbors_within_batch=3). This network was used for Leiden clustering of cells (scanpy.tl.leiden) and UMAP visualization (scanpy.tl.umap).

Lowering filtration thresholds did not reveal additional clusters that span both experimental collections (Figure S1J–L).

Genotyping and Demultiplexing

We assigned cells in each library to individual wild-caught animals using naturally present SNPs. Variant SNPs were identified using Cellsnp-lite⁵⁰ (mode 2, i.e. without an a priori list of SNPs; cellsnp version 0.3.2 and cellsnp-lite version 1.2.3). Cell barcodes were then assigned to animals using Vireo⁵¹ (vireosnp version 0.5.6). Initial inspection revealed problems with Vireo assignments for the April 2022 library, so we developed a custom pipeline. The inspection and new pipeline are described below.

Quality checks for Vireo’s assignments

Initial inspection of Vireo assignments showed that some transcriptional clusters were called by Vireo as having a large proportion of between-genotype doublets. These clusters were abundant and showed unique gene expression, which is not consistent with them being doublets. Specifically, Vireo labeled over 50% of cells in four clusters as doublets in the April 2022 library (Figure S7C). Further inspection showed that Vireo made use of SNPs associated with genes expressed specifically in these clusters, which suggested that Vireo was assigning clusters within one animal to different SNP profiles. This problem may occur because some animals in the April 2022 collection contributed <2% of the initial blood sample by volume. For the two scRNA-seq libraries generated from the May 2023 collection, we did not observe similar problems and thus directly used the assignments provided by Vireo without using the custom pipeline described below.

Custom SNP demultiplexing

To address the problem for the April 2022 library, we implemented a custom pre-processing step, which avoids focusing on SNPs that are cell type specific and instead uses SNPs broadly detected across cell states. To this end, we selected a subset of cells whose transcriptional clusters lack highly unique SNPs (cell barcodes listed in Table S5). We carried out principal component analysis on the variable SNPs for this subset of cells, and then projected the SNP profiles for all cells onto the first principal components (PCs). The PC space of variable SNPs was generated by log-transforming the SNP count matrix for this cell subset (using scanpy.pp.log1p), then selecting highly variable SNPs (using scanpy.pp.highly_variable_genes, min_mean=0.0125, max_mean=3, min_disp=0.5), and running principal component analysis (scanpy.pp.pca, with default settings and use_highly_variable=True). PC1 separated the animal with the most cells from the rest (Figure S7D). We thus defined new animal assignments to be “animal 1” for cells with a PC1 value < 0.3, “animals 2–5” for cells with a PC1 value > 0.7, and “undetermined” for all other cells (Figure S7E). We did not fractionate cells from animals 2–5 as the cell count for these animals was low.

Animal SNP profile present in two libraries

As noted above, one animal (animal F) contributed cells to both libraries in the May 2023 collection (see section ‘Single-Cell Barcoding, Library Preparation, and Sequencing’). To associate the cells from this animal in both libraries, Vireo-annotated animals were matched across libraries using vireoSNP.utils.vcf_utils.match_VCF_samples. A single SNP profile showed a close match between the libraries, and was taken to be the shared animal.

Identifying and Removing HA-3/SRC Doublets

We identified a subset of cell barcodes in the scRNA-seq dataset as doublets or multiplets of cells from the HA-3 and SRC clusters. Below, we describe (1) defining and removal of likely doublets, and (2) further the evidence that this subset represents doublets or multiplets.

Defining likely HA-3/SRC doublets

We found that Leiden cluster 19 (pre-doublet removal) comprises two populations—while all cluster 19 cells express markers of HA-3, a subset additionally expresses markers of SRC (corresponding to Leiden cluster 2 pre-doublet removal). We defined new cluster boundaries to reflect the presence of these two populations. For the cells in clusters 19 and 2, we categorized cells into 3 subsets: one expressing only HA-3 markers, one expressing only SRC markers, and one with coexpression of HA-3 and SRC markers.

Specifically, we focused on HA-3 marker KY21.Chr9.653 and SRC marker KY21.Chr11.830, which were used in HCR FISH experiments. We determined which cells coexpress these genes at the level of fine-grain clusters using the following approach. For Leiden clusters 2 and 19, we generated a new kNN graph (scanpy.pp.neighbors, n_neighbors=10), then assigned cells to fine-grain clusters with Leiden clustering at a high resolution (scanpy.tl.leiden, resolution=3). We classified each fine-grain cluster as “Chr9.653+” if they express Chr9.653 but not Chr11.830, as “Chr11.830+” if they express Chr11.830 but not Chr9.653, or as “double-positive” if they express both genes. Specifically, we considered a gene to be expressed in fine-grain cluster F if its mean log-transformed expression in cluster F is more than double the mean across cells in the full dataset which are not in Leiden clusters 2 or 19. The resulting redefinition of cell cluster boundaries are shown in Figure S7G,H (barcodes listed in Table S5).

Cell barcodes of the “double-positive” population were removed, and we then repeated the post-filtration processing steps as described in “Processing and Dimensionality Reduction”. Figure 1 shows regenerated Leiden clusters after doublet removal.

Evidence that these are doublets

We expand here with a supplemental note giving three pieces of evidence suggesting that the “double-positive” cells are doublets or multiplets. (i) The genes enriched in these cells are all shared with Chr9.653+ (HA-3) or Chr11.830+ (SRC) cells (Figure S7I). (ii) The double-positive cells have a 1.95-times and 4.39-times higher mean number of transcripts detected per cell as compared to Chr9.653+ and Chr11.830+ cells, respectively (Figure S7J). (iii) HCR FISH staining revealed no cells that coexpress Chr9.653 and Chr11.830, despite these genes being both detected in the double-positive population (p<0.05) (Figure S7K,L). The double-positive cells include cell barcodes from both experimental days (Figure S7M), so we speculate that it reflects a reproducible physical association of these two cell types.

Manual Sub-Clustering

After manual inspection of the Leiden clusters, we additionally sub-clustered Leiden clusters 2 and 14, due to within-cluster heterogeneity. In both cases, we split each of 2 and 14 into two by Leiden clustering with a resolution of 0.1; we called the bigger subclusters 2.1 and 14.1 and the smaller subclusters 2.2 and 14.2, respectively.

Cluster-Specific Marker Genes Selection

Filtering genes for HCR FISH

For choosing HCR FISH marker genes, we required that transcripts be long enough to accommodate 10 probes hybridizing to it. To allow 10 probe pairs of 53 nucleotides each, with a spacing of at least 10 nucleotides between each pair, transcripts must be at least 610 nucleotides long. We only considered marker genes from the KY21 gene models⁹³ whose shortest annotated transcript was at least 610 nucleotides in length.

Filtering genes for live reporters

For choosing live reporter marker genes, we wanted to avoid genes which might be expressed in the developmental precursors of blood cells; if GFP protein remained in these cells’ descendents, GFP+ blood cells could include cells not currently expressing the marker gene. Where possible, we selected genes with low or no expression in scRNA-seq data of C. robusta developmental mesenchyme¹⁰¹. Specifically, when possible, we chose genes with log-transformed expression log₁₀(CP10k +1) below 0.1 in all mesenchyme clusters.

Selecting marker genes

Marker genes for each cluster were selected to (i) be expressed highly enough to be detectable with HCR FISH or live reporters; (ii) show a large dynamic range with near-binary in expression, such that cells can be easily categorized as expressing or not expressing the gene; and (iii) be specific to each cluster.

To achieve (i), we ran preliminary experiments labeling genes with a range of expression levels to identify the detection limit. In the following, log-transformed expression refers specifically to ${log}_{10}\left(\right(countsperM)+1),$ where M=6,138 is the median unnormalized counts per cell across the dataset, evaluated before doublet cluster removal. In preliminary HCR FISH experiments, we found genes with a mean log-transformed expression below 1 were not detectable, but genes with expression above 2 were easily detected: KY21.Chr3.306 (mean log expression of 0.70 in GA) was not detected; Chr8.361 (1.6 in HA-3) was weakly detected; and all of Chr13.352 (2.8 in GA), Chr9.653 (3.5 in HA-3), and Chr12.617 (4.2 in GA) were easily detected. Based on these experiments, we required that candidate marker genes for each cluster of interest have a mean log-transformed expression above 1 in that cluster.

For (ii), we defined a score that we reasoned would identify genes that appear binary by HCR FISH as follows. Let $r_u:=min\{xϵX:x>u\}–max\{xϵX:x<u\}$, where $X=\{x_1,x_2,...,x_N\}$ and $x_k$ is the log-transformed mean expression of the gene in cluster k. The value of $r_{u=1}$ gives the log-ratio in expression between clusters above and below the detection limit (u=1). With this score, marker genes for each cluster were selected by first filtering genes to consider only those with mean log expression >1 in at least one cluster and <0.35 in at least one other, then further filtering genes with mean log expression above 1 in the cluster, and finally selecting the top several genes with highest $r_{u=1}$ values. Finally, we carried out a manual inspection of expression of these genes on the UMAP embeddings, to select 1–2 genes with highly specific expression for each cluster.

In some cases, no genes passing these filters uniquely labeled a cluster. For cluster pairs 1/7 and 8/26, which are transcriptionally similar, we chose marker genes which labeled both clusters together (Data S1–2b,5b). For cluster pair 0/14, also transcriptionally similar, no filtered genes labeled only 0/14 and no other clusters; we chose two genes which are only coexpressed in 0/14 (Data S1–1b). The selected marker genes are listed in Figure S2C (HCR FISH) and Figure S4A (live reporters).

Selection of Ubiquitous Control Gene

We selected a ubiquitously expressed gene to be used as a positive control in HCR FISH experiments. We required that such a gene have a mean $lo{g}_{10}(CP10k+1)$ above 1 in at least 30 out of 33 Leiden clusters. Four genes pass this threshold, from which we chose KY21.Chr11.687 (homologous to ACTB/G1) as a control.

Probe Design for HCR FISH

Once genes were selected for HCR FISH experiments, HCR amplifier sequences^53,102 (B1, B2, or B3) were chosen for each gene. These sequences determine which fluorescent hairpins hybridize to a gene’s probe set. The ubiquitous control gene KY21.Chr11.687 was given amplifier B2, and all cluster marker genes were given either B1 or B3 as reflected in Table S4. Probe sequences (split initiator probe pairs) were designed for each gene and HCR amplifier sequence as described by Choi et al.⁵³. Between 10 and 20 probe pairs were evenly spaced across each gene’s longest transcript sequence (excluding introns, including untranslated regions), with a minimum of 10 nucleotides between probe pairs. Sequences for each gene can be found in Table S4. Probes were ordered from Integrated DNA Technologies.

OrthoFinder-Based Vertebrate Homologs

We used OrthoFinder⁷¹ (version 2.5.5) to identify likely vertebrate homologs of C. robusta genes. OrthoFinder was run on the following proteomes with default parameters: vertebrate species Homo sapiens, Mus musculus, Gallus galllus, Xenopus tropicalis, Danio rerio, and Petromyzon marinus; and tunicate species Ciona robusta, Ciona savignyi, Styela clava, and Botrylloides diegensis. If genes from two species are in the same orthogroup as defined by OrthoFinder, they are considered homologous.

Gene Set Enrichment Analysis

For Figure S3D,E, gene enrichment analysis was carried out to identify KEGG pathways¹⁰³ and manually curated gene sets which were significantly enriched in each cell type’s list of differentially expressed genes (DEGs). DEGs for each cell type were determined using the Wilcoxon rank-sum test (scanpy.tl.rank_genes_groups, method='wilcoxon', requiring log2 fold-change > 1 and FDR<0.01). Lists of C. robusta genes associated with each KEGG pathway were generated based on their homology to human genes associated with that KEGG pathway, as described below. Fisher’s exact test and separately Gene Set Enrichment Analysis (GSEA) were then used to identify KEGG pathways and manually curated gene sets which were enriched in each cell type’s DEG list. Code for this enrichment analysis is available at https://github.com/AllonKleinLab/paper-data/tree/master/Scully_Ciona_blood_2025/scRNA-seq_analysis/gene_enrichment_analysis.

For each KEGG pathway, we defined the C. robusta gene list as the set of C. robusta genes which have at least one human homolog (defined by OrthoFinder) in that KEGG pathway’s human gene list. We avoided pathways represented by small numbers of genes by only testing enrichment of pathways with both of the following: (i) at least 10 C. robusta genes in the final gene list, and (ii) at least 10 distinct human genes with a homolog in C. robusta in the original human list. For manually curated gene sets, gene lists with citations and C. robusta homologs used can be found in Table S3.

For performing gene enrichment analysis, we considered only C. robusta genes which were detected in more than 20 cells in the scRNA-seq data. Genes expressed in fewer cells were removed from cell DEG and gene sets. Fisher’s exact test was run using scipy.stats.fisher_exact (alternative=‘greater’). Multiple hypothesis correction was the performed using the Benjamini-Hochberg procedure, requiring a false discovery rate (FDR) below 0.05 (statsmodels.stats.multitest.fdrcorrection, alpha=0.05). Separately, GSEA was run using the GSEApy package⁹² prerank function (rnk=‘scores’ output by scanpy.rank_genes_groups, permutation_num=1000, min_size=5). The full gene enrichment analysis results are in Table S3.

RNA Velocity Analysis

RNA velocity⁵⁵ was used to generate hypotheses of differentiation trajectories. Velocyto⁵⁵ (version 0.17.17) was used to quantify spliced and unspliced expression levels for each gene. We then used scVelo⁹⁰ (version 0.3.0) to compute a splice-aware neighbors graph of all cells and estimate velocities for all genes. These were used to compute a velocity graph whereby the transition probability between pairs of cells was estimated. This graph was overlaid on the pre-existing UMAP coordinates of the C. robusta barcoded cells (Figure 4B). Arrows pointing away from the cMPP cluster were manually selected and differently colored after vector image generation.

Differentiation Hierarchy Analysis

To build a tree connecting cell types in a differentiation hierarchy (Figure 4C), we adapted a method in Wagner et al. 2018¹⁰⁴. The tree is constructed as a maximum spanning tree (MST, implemented with networkx.maximum_spanning_tree¹⁰⁵) on an undirected, weighted cell type graph. The graph is constructed with edge-weights biased towards connections between cycling cell clusters (assumed to be progenitors) and non-cycling clusters (assumed to be differentiated non-progenitors). The bias is achieved by combining two graphs: a graph connecting all clusters to all clusters, and a bipartite graph connecting progenitors to non-progenitors. This biased approach enables the use of different parameters (thresholds on edge weights) for progenitor and non-progenitor clusters, as we observed that there is otherwise a dominance of connections within progenitor clusters. The method is detailed below, with Python code implementing the graph construction available at https://github.com/AllonKleinLab/paper-data/tree/master/Scully_Ciona_blood_2025/scRNA-seq_analysis/differentiation_hierarchy_analysis. The final tree was visualized in Fig 4C using Graphviz’s neato spring model layout¹⁰⁶, with some node positions manually adjusted to improve readability.

For building the differentiation hierarchy, we used cell type labels as defined in Table S2. We built the maximum spanning tree, $MST\left(G\right)$, of graph $G=(V,E,W)$, where vertices V are the set of cell clusters; edges E are the set of all possible edges, making a fully connected graph; and weights W reflect the connections between cell clusters. For simplicity, we omit $E$ from further definitions. This graph $G=(V,W)$ was constructed from two other graphs: (1) an unbiased graph $G^{\left(0\right)}=(V,W^{\left(0\right)})$, and (2) a bipartite graph $G^{\left(pd\right)}=(V,W^{\left(pd\right)})$ of progenitor-to-differentiated cell cluster connections. Within G, the weight between cell clusters i and j was defined as $w_{i,j}=max(w_{i,j}^{(0)},w_{i,j}^{(pd)})$.

We now define the construction of $G^{\left(0\right)}$ and $G^{\left(pd\right)}$ below. For each, we used the batch-corrected k-nearest neighbors (kNN) graph on the scRNA-seq data generated by BBKNN (see section “Processing and Dimensionality Reduction”).

(1) To build $G^{\left(0\right)}$, we established an undirected coarse-grained graph between annotated cell clusters as follows: Within the cell kNN graph, let $n(i,j)$ be the number of edges connecting individual cells between cell cluster i and cell cluster j, where $i\ne j$. Let $N(i)={\sum }_{i,k\ne i}n(i,k)$ be the total number of outgoing edges from cell cluster i. We define a coarse-grained graph as having one node per cell cluster, with edge weights, $w_{i,j}$, between cell cluster i and j given by:

$$w_{i,j}=\frac{n(i,j)}{\sqrt{N\left(i\right)N\left(j\right)}}$$

We then excluded edge weights which are less than 30% of the maximum edge weight to obtain final edge weights: ${\widehat{w}}_{i,j}=\left\{w_{i,j}\text{for}{w}_{i,j}>\alpha max(w_{i,j});0\text{otherwise}\right\}$, where $\alpha =0.3$.

(2) We built $G^{\left(pd\right)}$, a bipartite graph connecting progenitor to non-progenitor cell clusters, as follows: We labeled each cell cluster as a progenitor or non-progenitor cluster based on expression of cell cycle markers. Specifically, cell clusters were labeled as progenitors if their median expression of the MKI67 ortholog, KY21.Chr2.1280, was above zero; all other cell clusters were labeled as non-progenitors. For building this bipartite graph $G^{\left(pd\right)}$, we only considered edges in the cell kNN graph which connected a progenitor to a non-progenitor. The process for determining edge weights otherwise followed the process for building $G^{\left(0\right)}$. Note that since only a subset of edges from the kNN graph are used for building $G^{\left(pd\right)}$, the values of $n(i,j)$ and $N\left(i\right)$ are different from those used for building $G^{\left(0\right)}$. The same value $\alpha =0.3$ was used.

BLAST-Based Likely Vertebrate Homologs

To identify likely C. robusta homologs of human genes of interest, we employed a BLAST-based approach rather than relying solely on OrthoFinder results. This methodological choice was necessitated by OrthoFinder’s failure to identify several known immune-related homologs in C. robusta, including previously documented Toll-like receptor (TLR) homologs⁸⁰. The BLAST-based approach enabled a more comprehensive identification of functional homologs between human and C. robusta genomes for subsequent expression analysis.

Our BLAST-based homology detection pipeline accounted for vertebrate- and tunicate-specific gene duplication events by allowing one-to-many gene matching. Specifically, we identified each C. robusta gene’s best human gene hit based on BLASTp. To ensure specificity and prevent spurious matches, we filtered the results to exclude cases where the identified human hit had substantially closer sequence similarity to a different C. robusta gene, specifically: for each C. robusta gene $g^{Cr}$ with best human hit $g^{Hs}$, the gene pair $\left(g^{Cr},g^{Hs}\right)$ was excluded if the bit score between $g^{Cr}$ and $g^{Hs}$ was less than half the bit score between $g^{Hs}$ and its best C. robusta hit.

Comparison of Cell Types using SAMap

We ran SAMap to align human (Homo sapiens), zebrafish (Danio rerio), and C. robusta mature and differentiating blood cells. For the human dataset we used the Tabula Sapiens blood and bone marrow data⁶⁸. To improve the interpretability of results, we combined similar human cell type annotations and performed most analyses using this coarse-grain cell clustering (Table S5). For the zebrafish dataset, we used data of zebrafish kidney marrow from Wang et al.⁷². The cells for this dataset were manually re-annotated by inspecting gene expression, and new annotations for each cell barcode are listed in Table S5. Some cell barcodes were identified as likely kidney progenitors and were excluded from analysis.

SAMap was run (version 1.0.15) with samap.run (pairwise=True). Similarity matrices were generated with samap.analysis.get_mapping_scores (n_top=0). To display similarity matrices (Figures 5B, S6A), one species’ cell types were ordered from highest to lowest similarity score. For the second species, we used the following ordering procedure: For each cell type $i$ in the ordered list of the first species, we identified all cell types from the second species that had their maximum similarity with cell type $i$. These matching cell types were then added to the ordered list of the second species, arranged from highest to lowest maximum similarity score. Lists of co-enriched genes between cell states (see Table S1) were generated with samap.analysis.GenePairFinder.find_all (align=0.1).

Jaccard Index

We define a cell type similarity score based on the Jaccard index to quantify the fraction of cell type DEGs which overlap between two species. Let A be the set of DEGS in cell type 1 of species 1, excluding genes with no homolog in species 2. Let B be the set of DEGs in cell type 2 of species 2, excluding genes with no homolog in species 1. Given the many-to-many gene homology, we calculate the intersection of A and B in two ways: We define $\{A\cap B{\}}_1$ as the set of species 1 genes in A with at least one homologous gene in B, and $\{A\cap B{\}}_2$ as the set of species 2 genes in B with at least one homologous gene in A We then calculate two Jaccard indexes, $J_1=\frac{\left|\{A\cap B{\}}_1\right|}{\left|A\right|+\left|B\right|-\left|\{A\cap B{\}}_1\right|}$ and $J_2=\frac{\left|\{A\cap B{\}}_2\right|}{\left|A\right|+\left|B\right|-\left|\{A\cap B{\}}_2\right|}$. The final used Jaccard index is defined as the mean of these two values: $J=\frac{1}{2}(J_1+J_2)$.

We identified DEGs for each cell type in the human, zebrafish, and C. robusta datasets using the following approach. For the human and zebrafish data, we normalized the count matrix to 1e4 counts per cell (scanpy.pp.normalize_total, target_sum=10^4) and log-transformed the data (scanpy.pp.log1p, base=10). For all three species datasets, we identified enriched DEGs using the Wilcoxon rank-sum test (scanpy.tl.rank_genes_groups, method=‘wilcoxon’), keeping genes with a log2 fold-change > 2 and a false discovery rate < 0.05.

Generalized Sankey (GS) Plots

Definition of GS plots

GS plots are a generalization of Sankey plots, which visualize flows between nodes of a directed graph. A canonical Sankey plot visualizes incompressible flows, and the objects flowing between nodes are not individually identifiable. GS plots make two generalizations: they allow flows to be compressible (expanding or contracting), and they trace flow of identifiable elements between nodes. In a limit of incompressible flow and arbitrary element identity, a GS plot is a Sankey plot.

In Sankey and GS plots, graph nodes are represented by rectangles, and graph edges are represented by ribbon-like connectors¹⁰⁷. Canonical Sankey plots visualize flow magnitude by the widths of connectors. Flows do not overlap, and the sum of connector widths into and out of a node are constant to reflect incompressible flow. For the general case represented by GS plots, the underlying flows can overlap when the same elements (in our case, genes) from one node (in our case, a cell type) flow to elements in more than one node. Second, compressible flows are represented by connector widths changing between start and end. Third, we consider the case in which nodes contain additional elements not present in the flows; graphically, this is represented by allowing node rectangle widths to exceed the total width of incoming flows.

In the following, we define GS plots for the case of a directed graph with two layers (two species, in our case), rooted in a single node (one cell type, in our case) in one layer, which flows to one or more nodes in a second layer (multiple cell types). The code for making these plots is available on Github (https://github.com/AllonKleinLab/paper-data/tree/master/Scully_Ciona_blood_2025/species_comparisons/gs_plots). These GS plots span three levels of ontology: two layers, which contain one or more nodes, which contain multiple elements. Elements may have many-to-many relationships across layers. The GS plots compare a focus node (FN) from one layer (Layer A) to a set of nodes from another layer (Layer B). Each node contains a number of elements and is represented by a rectangle whose width corresponds to the number of those contained elements.

GS plots use ribbon-like flows whose widths represent the number of shared elements between the FN in Layer A and each node in Layer B. For the flows between two nodes, the width where a flow meets the top (FN) rectangle represents the number of the FN’s elements which are related to at least one element in the connected bottom node; the width at the bottom represents the number of the bottom node’s elements which are related to at least one element in the FN. Since we allow many-to-many element relationships, the flows’ widths at the top and bottom may not be the same. Flows are plotted to overlap in the FN when the same elements are present in multiple categories of Layer B. The flows are specifically rendered as the area between two Bézier curves.

Application of GS plots to cross-species comparison (Figure 5C,D)

To use GS plots in cross-species comparisons (Figure 5C,D), each layer represents a species, nodes represent cell types, and elements represent enriched DEGs. DEGs were determined as described for the Jaccard index calculation (see “Jaccard Index”). Many-to-many homologous relationships between genes were determined by OrthoFinder (see “Identification of C. robusta Vertebrate Homologs with OrthoFinder”). For each cell type, we additionally represented genes with no homology (i.e. node elements which can never be part of flows) through rectangle coloration: filled-in regions represent the number of genes which have at least one homolog in the other species’ genome, and empty regions represent genes with no detected homolog.

When visualizing relationships between a focus cell type (FCT) and cell types from another species, we selected the five cell types with the highest Jaccard indices, excluding cell types with fewer than 10 shared genes. Cell types were arranged from left to right in order of decreasing Jaccard index.

Lineage-Enriched Transcription Factors

We first generated a list of C. robusta TFs by combining a curated list of C. robusta TFs from the Ghost Database⁸² with 70 additional genes which OrthoFinder mapped to known human TFs¹⁰⁸. Table S1 contains the full list of 381 C. robusta TFs, along with names used in this paper. Next, using the coarse-grain differentiation hierarchy tree (Figure 4C), we found TFs which might be involved in fate decisions by identifying differentially expressed TFs at each branching point (Figure 6C). There are three branch points in the tree (i.e. nodes with degree > 2): cMPP, cLRP-2, and cLRP-4. At each of these branch points, we identified differentially expressed TFs across the subset of cell clusters which are child nodes of the branching point (e.g. for the branch point cLRP-3, we compared cLRP-5, pre-RA, RC, and URG-1). Specifically, we used the Wilcoxon rank-sum test (scanpy.rank_genes_groups, method=‘wilcoxon’, corr_method=‘benjamini-hochberg’) to compare TF expression in each cell cluster to other cell clusters in the subset. TFs with a log fold-change > 2 and a false discovery rate < 0.01 were labeled as enriched in the given cell cluster and are plotted in Figure 6C.

Coexpression Conservation

Conservation of gene coexpression relationships was calculated following methods established in Crow et al.⁷⁶ with one minor modification. For completeness we describe the approach here. We first built a gene coexpression network within each species (C. robusta, human, and zebrafish) as follows: the scRNA-seq data was clustered at a high resolution (scanpy.tl.leiden, resolution=20), and then for all pairs of genes we calculated the Spearman correlation of mean log-transformed expression across clusters. These correlation coefficients were ranked (descending) across all gene pairs to generate a weighted graph of gene coexpression relationships. We next quantified gene coexpression conservation based on how the networks centered around each gene compares across species using the “AUROC score” described in Crow et al.⁷⁶, allowing N-to-M homology. Gene homology was determined using OrthoFinder. This approach recapitulates the method in Crow et al.⁷⁶, with only a difference in the original high-resolution clustering step.

As a complementary approach, we calculated a second conservation score, which shows similar results (Figure 6G, Figure S6D). The second score, labeled as the “number of shared coexpression partners”, represents the number of coexpression partners shared across species, and was calculated as follows. We compared a human gene of interest ($g_h$) to one homologous gene ($g_x$) in a non-human species (C. robusta or zebrafish). Taking the top 30 human coexpression partners of $g_h$ (the 30 genes with the largest edge weights connected to $g_h$), we identified the non-human species’ genes which are homologous to at least one human coexpression partner. There were more or fewer than 30 homologous genes in cases of N-to-M homology. We then defined the score as the number of these coexpression partner homologs which are highly correlated with $g_x$, specifically the number of these genes whose Pearson correlation with $g_x$ is above 0.5.

For cases where the genes of interest $g_h$ or $g_x$ have more than one homolog in the other species, we calculated separate coexpression conservation scores for all homologous gene pairs. For example, human IRF4 and IRF8 are both homologous to C. robusta Irf-r.b, so we calculated two conservation scores—one each for IRF4 vs. Irf-r.b and IRF8 vs. Irf-r.b (Figure 6G).

Supplementary Material

table_s1_homology

Table S1. C. robusta gene names and homology to vertebrates based on BLAST, OrthoFinder, and SAMap gene matchings, including genes for which no homologs were detected, related to Figures 4 and 6.

table_s2_state_names_abundance

Table S2. Cell state definitions and relative abundances, related to Table 1

video_S1

Video S1. Reporter-positive cell behavior in vivo, related to Figures 3 and 4

(A) HA-3 reporter-positive cells circulate in a juvenile. Green represents the HA-3 reporter, Chr8.361>GFP. Magenta represents a heart reporter, Mesp>RFP.

(B) HA-1 reporter-positive cells circulate in a juvenile. Green represents the HA-1 reporter, Chr1.1154>GFP.

(C) URG-1 reporter-positive cells circulate in a juvenile. Green represents the URG-1 reporter, Chr2.603>GFP.

(D) cLRP-1 reporter-positive cells circulate in a juvenile. Green represents the cLRP-1 reporter, Chr4.869>GFP.

(E) GA reporter-positive cell crawling in a juvenile. Green represents the GA reporter, Chr13.352>GFP.

(F) HA-2 reporter-positive cell crawling in a juvenile. Green represents the HA-2 reporter, Chr2.1414>GFP.

(G) HA-3 reporter-positive cell crawling in a juvenile. Green represents the HA-3 reporter, Chr8.361>GFP.

(H) HA-1 reporter-positive cell crawling in a juvenile. Green represents the HA-1 reporter, Chr1.1154>GFP.

(I) cMPP/cLRP reporter-positive cells during metamorphosis. Green represents the cMPP/cLRP reporter, Chr4.593>GFP, and is shown as a maximum-intensity projection across z-slices.

(J) cMPP/cLRP reporter-positive cells are static in the tunic and circulating in a juvenile. Green represents the cMPP/cLRP reporter, Chr4.593>GFP. Magenta represents a heart reporter, Mesp>RFP.

table_s3_gsea

Table S3. Gene enrichment analysis for KEGG pathways and manually curated gene sets, related to Figure 4

table_s4_hcr_fish

Table S4. HCR FISH probe set sequences and combinations used per experiment, related to Figure 2

table_s5_barcode_lists

Table S5. Cell barcode lists for April library SNP demultiplexing, for the HA-3/SRC doublet cluster, and for human and zebrafish cell type annotations, related to STAR Methods

Data S2

Data S1. HCR FISH images for each cluster, related to Figure 2

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Chemicals, peptides, and recombinant proteins
Artificial seawater	Instant Ocean	Cat# SS15-10
Phyto-Feast	Reef Nutrition	N/A
99% L-menthol crystals	Thermo Scientific	Cat# A10474
Bovine serum albumin	Sigma Aldrich	Cat# 9048-46-8
Dulbecco’s phosphate-buffered saline	Gibco	Cat# 14200075
D-mannitol	Sigma Aldrich	Cat# M4125
10X PBS	Gibco	Cat# 70011-044
Concanavalin A	Sigma Aldrich	Cat# C2010-1G
Nuclease free water	Invitrogen	Cat# 10977-023
DAPI	Sigma Aldrich	Cat# D9542
16% paraformaldehyde	Electron Microscopy Sciences	Cat# 15710
Methanol	OmniSolv	Cat# MX0484-1
Tween-20	Millipore Sigma	Cat# P9416-50ML
Triton X-100	VWR	Cat# VW3929-2
Probe hybridization buffer	Molecular Instruments	N/A
Probe wash buffer	Molecular Instruments	N/A
20X SSC	Invitrogen	Cat# AM9763
Amplification buffer	Molecular Instruments	N/A
Metastable fluorescent hairpins: B1-Alexa Fluor 488, B1-Alexa Fluor 546, B1-Alexa Fluor 647, B2-Alexa Fluor 488, B2-Alexa Fluor 546, B3-Alexa Fluor 488, and B3-Alexa Fluor 647	Molecular Instruments	N/A
Penicillin-Streptomycin	Thermo Fisher Scientific	Cat# 15070063
Zymosan	Invivogen	Cat# tlrl-zy
Sodium bicarbonate	Sigma Aldrich	Cat# S6014-1KG
pHrodo amine-reactive dye	Invitrogen	P36011
Critical commercial assays
Chromium Next GEM Single Cell 3’ GEM Kit v3.1	10X Genomics	Cat# PN-1000123
Chromium Next GEM Chip G Single Cell Kit	10X Genomics	Cat# PN-1000127
Library Construction Kit	10X Genomics	Cat# PN-1000190
Deposited data
Ciona robusta blood scRNA-seq data, May 2023 libraries	This paper	GEO: GSE296253
Ciona robusta blood scRNA-seq data, April 2022 library	Scully et al.49	GEO: GSE292926
Ciona robusta genome assembly, HT2019 assembly plus the mitochondrial chromosome from the Ensembl KH assembly	Scully et al.49	https://github.com/AllonKleinLab/paper-data/tree/master/Scully_mannitol_2025/C_robusta_genome_HT2019_KY21_with_Ensembl_mito
Ciona robusta combined gene annotations, KY21 gene models plus the mitochondrial genes from the Ensembl KH annotations	Scully et al.49	https://github.com/AllonKleinLab/paper-data/tree/master/Scully_mannitol_2025/C_robusta_genome_HT2019_KY21_with_Ensembl_mito
Tabula Sapiens blood and bone marrow dataset	The Tabula Sapiens Consortium68	TS_Blood.h5ad.zip and TS_Bone_Marrow.h5ad.zip from https://figshare.com/articles/dataset/Tabula_Sapiens_release_1_0/14267219
Zebrafish kidney marrow scRNA-seq dataset	Wang et al.72	Genome Sequence Archive CRA017585
List of C. robusta TFs	Ghost Database82	https://ghost.zool.kyoto-u.ac.jp/HT_TF_KY21.html
Homo sapiens proteome (GRCh38)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/homo_sapiens/pep/
Mus musculus proteome (GRCm39)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/mus_musculus/pep/
Gallus gallus proteome (GRCg7b)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/gallus_gallus/pep/
Xenopus tropicalis proteome (UCB_Xtro_10)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/xenopus_tropicalis/pep/
Danio rerio proteome (GRCz11)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/danio_rerio/pep/
Petromyzon marinus proteome (Pmarinus_7.0)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/petromyzon_marinus/pep/
Ciona savignyi proteome (CSAV 2.0)	Ensembl	https://ftp.ensembl.org/pub/release-114/fasta/ciona_savignyi/pep/
Styela clava proteome (NCBI_ASM1312258v2)	NCBI	https://www.ncbi.nlm.nih.gov/datasets/genome/GCF_013122585.1/
Botrylloides diegensis proteome (listed as Botrylloides leachii)	Blanchoud et al. 201883, listed as Botrylloides leachii	https://www.aniseed.fr/aniseed/download/download_data?module=aniseed&action=download:download_data
Experimental models: Organisms/strains
Ciona robusta: wild-caught from Carlsbad, California, USA	M-REP	N/A
Oligonucleotides
HCR FISH probes for C. robusta genes	This paper	See Table S4
Recombinant DNA
Heart reporter: Mesp>GFP	Davidson et al.84	N/A
Heart reporter: Mesp>RFP	Davidson et al.85	N/A
HA-3 reporter: KY21.Chr8.361>GFP	This paper	N/A
HA-1 reporter: KY21.Chr1.1154>GFP	This paper	N/A
URG-1 reporter: KY21.Chr2.603>GFP	This paper	N/A
cLRP-1 reporter: KY21.Chr4.896>GFP	This paper	N/A
GA reporter: KY21.Chr13.352>GFP	This paper	N/A
HA-2 reporter: KY21.Chr2.1414>GFP	This paper	N/A
Software and algorithms
GffRead (version 0.12.7)	Pertea and Pertea86	https://f1000research.com/articles/9-304/v1
Cell Ranger (version 6.1.2)	10X Genomics	https://www.10xgenomics.com/support/software/cell-ranger/latest
NIS-Elements	Nikon	https://www.microscope.healthcare.nikon.com/products/software/nis-elements
napari (version 0.4.14)	Sofroniew et al.87	https://napari.org/stable/
SCRUBLET (version 0.2.3)	Wolock et al.88	https://github.com/AllonKleinLab/scrublet
Scanpy (version 1.8.1)	Wolf et al.89	https://scanpy.readthedocs.io/en/stable/
Cellsnp-lite: cellsnp (version 0.3.2) and cellsnp-lite (version 1.2.3)	Huang and Huang50	https://cellsnp-lite.readthedocs.io/en/latest/
Vireo: vireosnp (version 0.5.6)	Huang et al.51	https://vireosnp.readthedocs.io/en/latest/
OrthoFinder (version 2.5.5)	Emms and Kelly71	https://github.com/davidemms/OrthoFinder
Velocyto (version 0.17.17)	La Manno et al.55	https://velocyto.org/
scVelo (version 0.3.0)	Bergen et al.90	https://scvelo.readthedocs.io/en/stable/
BLAST+ (version 2.14.0 for Ciona robusta vs zebrafish; 2.12.0 for all other runs)	Camacho et al.91	https://ftp.ncbi.nlm.nih.gov/blast/executables/blast+/
GSEApy (version 1.1.3)	Fang et al.92	https://gseapy.readthedocs.io/en/latest/introduction.html
SAMap (version 1.0.15)	Tarashansky et al.73	https://github.com/atarashansky/SAMap
Other
Zero dead space tuberculin syringe with a 25G needle	Exelint	Cat# 26046
40 μm cell strainer	pluriSelect	Cat# 43-10040-60
8-well polymer coverslip pretreated with poly-L-lysine	ibidi	Cat# 80824

Data and code availability

• The scRNA-seq datasets generated and analyzed in the current study have been deposited at NCBI Gene Expression Omnibus (GEO) as Series GSE296253 (May 2023 libraries) and Sample GSM8869531 of Series GSE292926 (April 2022 library). An interactive version of the scRNA-seq dataset is available at https://kleintools.hms.harvard.edu/paper_websites/scully_ciona_robusta_blood/. Imaging data reported in this paper will be shared by the lead contact upon request.

• All original code has been deposited on Github at https://github.com/AllonKleinLab/paper-data/tree/master/Scully_Ciona_blood_2025 and is publicly available at doi:10.5281/zenodo.17209147 as of the date of publication.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.