SAMap iterates between two modules. The first module constructs a gene-gene bipartite graph with cross-species edges connecting homologous gene pairs, initially weighted by protein sequence similarity (Figure 1C). In the second module, SAMap uses the gene-gene graph to project the two single-cell transcriptomic datasets into a joint, lower-dimensional manifold representation, from which each cell’s mutual cross-species neighbors are linked to stitch the cell atlases together (Figure 1D). Then, using the joint manifold, the expression correlations between homologous genes are computed and used to reweight the edges in the gene-gene homology graph in order to relax SAMap’s initial dependence on sequence similarity. The new homology graph is used as input to the subsequent iteration of SAMap, and the algorithm continues until convergence, defined as when the cross-species mapping does not significantly change between iterations (Figure 1E).

This algorithm overcomes several challenges inherent to mapping single-cell transcriptomes between distantly related species. First, complex gene evolutionary history often results in many-to-many homologies with convoluted functional relationships (Briggs et al., 2018; Nehrt et al., 2011). SAMap accounts for this by using the full homology graph to project each dataset into both its own and its partner’s respective principal component (PC) spaces, constructed by the SAM algorithm, which we previously developed to robustly and sensitively identify cell types (Tarashansky et al., 2019). The resulting within- and cross-species projections are concatenated to form the joint space. For the cross-species projections, we translate each species’ features into those of its partner, with the expression for individual genes imputed as the weighted average of their homologs specified in the gene-gene bipartite graph. Iteratively refining the homology graph to only include positively correlated gene pairs prunes the many-to-many homologies to only include genes that are expressed in the same mapped cell types.

Second, frequent gene losses and the acquisitions of new genes result in many cell type gene expression signatures being species-specific, limiting the amount of information that is comparable across species. Restricting the analysis of each dataset to only include genes that are shared across species would result in a decreased ability to resolve cell types and subtypes with many species-specific gene signatures. SAMap solves this problem by constructing the joint space through the concatenation of within- and cross-species projections, thus encoding all genes from both species.

Third, the evolution of expression programs gradually diminishes the similarity between homologous cell types. To account for this effect, SAMap links cell types across species while tolerating their differences. Cells are mapped by calculating each of their k mutual nearest cross-species neighbors in the combined projection. To establish more robust mutual connectivity, we integrate information from each cell’s local, within-species neighborhood (Figure 1D), overcoming the inherent stochasticity of cross-species correlations. Two cells are thus defined as mutual nearest cross-species neighbors when their respective neighborhoods have mutual connectivity. It is important to note that the magnitude of connections is not directly calculated from their expression similarity, allowing cell types with diverged expression profiles to be tightly linked if they are among each other’s closest cross-species neighbors.

Lastly, SAMap is robust to technical batch effects between datasets that are collected through different platforms. For instance, we have succeeded in running SAMap on datasets containing hundreds of thousands of cells that were collected with different single-cell platforms, including 10X genomics, Drop-Seq, SmartSeq, and MARS-Seq. SAMap runtimes were typically less than an hour on an average desktop computer for the largest dataset we tested (Figure 1—figure supplement 1). Further, SAMap overcomes potential memory issues when running on large datasets by chunking its computationally intensive operations into smaller blocks, saturating the memory usage with respect to the number of cells (Figure 1—figure supplement 1B).

We first applied SAMap to the Xenopus and zebrafish atlases, which both encompass embryogenesis until early organogenesis (Briggs et al., 2018; Wagner et al., 2018). Previous analysis linked cell types between these two organisms by matching ontogeny, thereby providing a reference for comparison. SAMap produced a combined manifold with a high degree of cross-species alignment while maintaining high resolution for distinguishing cell types in each species (Figure 2A). We measured the mapping strength between cell types by calculating an alignment score (edge width in Figure 2B and color map in Figure 2C), defined as the average number of mutual nearest cross-species neighbors of each cell relative to the maximum possible number of neighbors.

SAMap revealed broad agreement between transcriptomic similarity and developmental ontogeny, linking 26 out of 27 expected pairs based on previous annotations (Figure 2B and Supplementary file 2; Briggs et al., 2018). The only exception is the embryonic kidney (pronephric duct/mesenchyme), potentially indicating that their gene expression programs have significantly diverged. In addition, SAMap succeeded in drawing parallels between the development of homologous cell types and matched time points along several cell lineages (Figure 2C). While the concordance was consistent across cell types, we noticed that the exact progression of developmental timing can vary, suggesting that SAMap can quantify heterochrony with cell type resolution.

SAMap also linked a group of secretory cell types that differ in their developmental origin, some even arising from different germ layers (highlighted edges in Figure 2B). Within ectoderm, frog cement gland cells map to zebrafish muc5ac⁺ secretory epidermal cells, and frog small secretory cells (SSCs) map to zebrafish pvalb8⁺ mucous cells (Janicke et al., 2010). Across germ layers, SSCs also map weakly to zebrafish endodermal cells, and frog ectodermal hatching gland maps to zebrafish mesodermal hatching gland. These cell types are linked through a large set of genes, including proteins involved in vesicular protein trafficking and several conserved transcription factors (TFs) such as myb, foxa1, xbp1, and klf17 (Figure 2D), which all have documented functions in controlling the differentiation of secretory cell types (Bennett et al., 2007; Dubaissi et al., 2014; Pan et al., 2014). For example, klf17 is expressed in zebrafish and frog hatching glands, and plays essential roles in regulating gland cell specification in both species (Kurauchi et al., 2010; Suzuki et al., 2019). Together, the conserved cell type specification programs (Erwin and Davidson, 2009) between developmentally distinct secretory cells support the notion that they may be transcriptionally and evolutionarily related despite having different developmental origins (Arendt et al., 2016).

To benchmark the performance of SAMap, we used eggNOG (Huerta-Cepas et al., 2019) to define one-to-one vertebrate orthologs between zebrafish and frog and fed these gene pairs as input to several broadly used single-cell data integration methods, Seurat (Stuart et al., 2019), LIGER (Welch et al., 2019), Harmony (Korsunsky et al., 2019), Scanorama (Hie et al., 2019), and BBKNN (Polański et al., 2019). We found that they failed to map the two atlases, yielding minimal alignment between them (Figure 2E and Figure 2—figure supplement 1). We also compared the results when restricting SAMap to using the one-to-one orthologs instead of the full homology graph. Even when removing the many-to-many gene homologies and the iterative refinement of the homology graph, we identified similar, albeit weaker, cell type mappings. This suggests that, at least for the frog and zebrafish comparison, SAMap’s performance is owed in large part to its robust, atlas stitching approach.

Finally, to test if SAMap is robust to incomplete cell type atlases, we downsampled the frog and zebrafish data by systematically removing individual cell types. We found that cell types whose homologous partners were removed mapped weakly to closely related cell types, some of which were already present in the original mapping. For example, optic cells from both species were also connected to eye primordium, frog skeletal muscles to zebrafish presomitic mesoderm, and frog hindbrain to zebrafish forebrain/midbrain (Supplementary file 3). We observed several new mapping pairs, but their alignment scores were all barely above the detection threshold of SAMap. Moreover, most of these edges were mapped between cell types with similar developmental origins, with the only exception being the zebrafish neural crest mapped to the frog otic placode in the absence of frog neural crest cells. Examining the genes that support this mapping revealed that both cell types express sox9 and sox10, two TFs previously implicated to form a conserved gene regulatory circuit common to otic/neural crest cells (Betancur et al., 2011). Taken together, these results suggest that SAMap is more sensitive in linking homologous cell types compared to other existing methods, exhibits high robustness when applied to incomplete datasets, and yields mapping results that are well supported by conserved gene expression programs.

The key benefit of using the full homology graph is to enable the systematic identification of gene paralogs that exhibit greater similarity in expression across species than their corresponding orthologs. These events are expected to arise as the result of gene duplications followed by diversification of the resulting in-paralogs (Studer and Robinson-Rechavi, 2009). In an alternative scenario, genetic compensation by transcriptional adaptation, where loss-of-function mutations are balanced by upregulation of related genes with similar sequences (El-Brolosy et al., 2019), could also result in this signature.

In total, SAMap selected 8286 vertebrate orthologs and 7093 paralogs, as enumerated by eggNOG, for manifold alignment. Paralogs were identified as non-orthologous genes that map to the same eggNOG orthology group ancestral to Vertebrata. Among these, 565 genes have markedly higher expression correlations (correlation difference >0.3) with their paralogs than their orthologs (see Figure 3A for examples), and 209 of them have orthologs that are either completely absent or lowly expressed with no cell-type specificity (Supplementary file 4). We term these events as ‘paralog substitutions’, as the orthologs may have lost or changed their functional roles at some point and were compensated for by their paralogs. Substituting paralogs were identified in most cells types with some (e.g. dorsal organizer) exhibiting higher rates than others (Figure 3B), suggesting uneven diversification rates of paralogs across cell types. SAMap also linked an additional 297 homologous pairs previously unannotated by orthology or paralogy, but which exhibit sequence similarity and high expression correlations (>0.5 Pearson correlation). These likely represent unannotated orthologs/paralogs or isofunctional, distantly related homologs (Gabaldón and Koonin, 2013).

We next asked whether paralog substitution rates depend on the evolutionary time since gene duplication. We categorized paralogs by the taxonomic level of their most recent shared orthology group and found that more recent paralogs substitute orthologs at higher rates than more ancient paralogs (Figure 3C). This observation is consistent with the expectation that less diverged genes may be more capable of functionally compensating for each other. To rule out the possibility that these paralogs were linked spuriously during the homology refinement steps of SAMap, we repeated the paralog substitution analysis on an aligned manifold constructed using only one-to-one orthologs. We identified 70% of the paralog substitutions and observed similar patterns in evolutionary time and cell type dependencies (Figure 3—figure supplement 1A–B). The other 30% of substitutions had smaller correlation differences on the border of our detection threshold (i.e. correlation difference >0.3) (Figure 3—figure supplement 1C). Failure to detect these substitutions was due to inaccurate imputation of gene expressions across species when restricting the mapping to one-to-one orthologs, which resulted in weaker alignment with fewer cross-species edges. Altogether, these results illustrate the potential of SAMap in leveraging single-cell gene expression data for pruning the networks of homologous genes to identify evolutionary substitution of paralogs and, more generally, identify non-orthologous gene pairs that may perform similar functions in the cell types within which they are expressed.

To test if we can identify homologous cell types in animals with radically different body plans, we mapped the cell atlases of two flatworms, the planarian Schmidtea mediterranea (Fincher et al., 2018), and the trematode Schistosoma mansoni, which we collected recently (Li et al., 2021). They represent two distant lineages within the same phylum but have remarkably distinct body plans and autecology (Laumer et al., 2015; Littlewood and Waeschenbach, 2015). While planarians live in freshwater and are known for their ability to regenerate (Reddien, 2018), schistosomes live as parasites in humans. The degree to which cell types are conserved between them is unresolved, given the vast phenotypic differences caused by the transition from free-living to parasitic habits (Laumer et al., 2015).

SAMap revealed broad cell type homology between schistosomes and planarians. The schistosome had cells mapped to the planarian stem cells, called neoblasts, as well as most of the differentiated tissues: neural, muscle, intestine, epidermis, parenchymal, protonephridia, and cathepsin⁺ cells, the latter of which consists of cryptic cell types that, until now, have only been found in planarians (Fincher et al., 2018; Figure 4A). These mappings are supported by both known cell-type-specific marker genes and numerous homologous transcriptional regulators (Figure 4B and Figure 4—figure supplement 1).

We next determined if cell type homologies exist at the subtype level. For this, we compared the stem cells, as planarian neoblasts are known to comprise populations of pluripotent cells and tissue-specific progenitors (Fincher et al., 2018; Zeng et al., 2018). By mapping the schistosome stem cells to a planarian neoblast atlas (Zeng et al., 2018), we found that the schistosome has a population of stem cells, ε-cells (Wang et al., 2018), that cluster with the planarian’s pluripotent neoblasts, both expressing a common set of TFs (e.g. soxp2, unc4, pax6a, gcm1) (Figure 4C–D). The ε-cells are closely associated with juvenile development and lost in adult schistosomes (Wang et al., 2018; Nanes Sarfati et al., 2021), indicating pluripotent stem cells may be a transient population restricted to their early developmental stages. This is consistent with the fact that, whereas schistosomes can heal wounds, they have limited regenerative ability (Wendt and Collins, 2016). SAMap also linked other schistosome stem cell populations with planarian progenitors, including two populations of schistosome stem cells – denoted as μ (Tarashansky et al., 2019) and μ’ – to planarian muscle progenitors, all of which express myoD, a canonical master regulator of myogenesis (Scimone et al., 2017). These likely represent early and late muscle progenitors, respectively, as μ-cells do not yet express differentiated muscle markers such as troponin, whereas μ’-cells do (Figure 4—figure supplement 2).

To compare cell types across broader taxonomic scales, we extended our analysis to include juvenile freshwater sponge (Spongilla lacustris) (Musser et al., 2019), adult Hydra (Hydra vulgaris) (Siebert et al., 2019), and mouse (Mus musculus) embryogenesis (Pijuan-Sala et al., 2019) atlases. In total, SAMap linked 1051 cross-species pairs of cell types, defined by the annotations used in each respective study. Of the cell type pairs, 95% are supported by at least 40 enriched gene pairs, and 87% are supported by more than 100 gene pairs, indicating that SAMap does not spuriously connect cell types with limited overlap in transcriptional profiles (Figure 5—figure supplement 1A).

We next extended the notion of cell type pairs to cell type trios, as mapped cell types gain additional support if they share transitive relationships to other cell types through independent mappings, forming cell type triangles among species. The transitivity of a cell type pair (edge) or a cell type (node) can be quantified as the fraction of triads to which they belong that form triangles (Figure 5A). The majority (81%) of cell type pairs have non-zero transitivity independent of alignment score and the number of enriched gene pairs (Figure 5—figure supplements 1–2). Cell type pairs with fewer than 40 enriched gene pairs tend to have lower (<0.4) transitivity (Figure 5—figure supplement 1B). The transitivity measure can also be used to identify potentially spurious connections. 16% of mapped cell type pairs have zero edge transitivity but non-zero node transitivity. These cell types are connected to only a single member of an interconnected cell type group (motifs 2 and 3 in Figure 5B). Such links may be of lower confidence as they should connect to other members of the group and are thus excluded from downstream analysis.

Among the interconnected groups of cell types, we identified families of neural cells and contractile cells (Figure 5C). Both cell type families are highly transitive compared to the overall graph transitivity (bootstrap p-value<1×10⁻⁵), meaning that their constituent cell types have more transitive edges within the group than outside the group (Figure 5D). In addition, the dense, many-to-many connections within the contractile and neural families are each supported by at least 40 enriched gene pairs (Figure 5E). Consistent with the nerve net hypothesis suggesting a unified origin of neural cell types (Tosches and Arendt, 2013), the neural family includes vertebrate brain tissues, both bilaterian and cnidarian neurons, cnidarian nematocytes that share the excitatory characteristics of neurons (Weir et al., 2020), and Spongilla choanocytes and apopylar cells, both of which are not considered as neurons but have been shown to express postsynaptic-like scaffolding machinery (Musser et al., 2019; Wong et al., 2019). The contractile family includes myocytes in bilaterian animals, Hydra myoepithelial cells that are known to have contractile myofibrils (Buzgariu et al., 2015), and sponge pinacocytes and myopeptidocytes, both of which have been implicated to play roles in contractility (Musser et al., 2019; Sebé-Pedrós et al., 2018). In contrast to the families encompassing all seven species, we also found a fully interconnected group that contains invertebrate multipotent stem cells, including planarian and schistosome neoblasts, Hydra interstitial cells, and sponge archeocytes (Alié et al., 2015). The lack of one-to-one connections across phyla is in keeping with recent hypotheses that ancestral cell types diversified into families of cell types after speciation events (Arendt et al., 2016; Arendt et al., 2019). Our findings thus suggest that these cell type families diversified early in animal evolution.

The high interconnectedness between cell types across broad taxonomic scales suggests that they should share ancestral transcriptional programs (Arendt et al., 2016). SAMap identified broad transcriptomic similarity between bilaterian and non-bilaterian contractile cells that extends beyond the core contractile apparatus. It links a total of 23601 gene pairs, connecting 5471 unique genes, which are enriched in at least one contractile cell type pair. Performing functional enrichment analysis on these genes, we found cytoskeleton and signal transduction functions to be enriched (p-value<10⁻³) based on the KOG functional classifications (Tatusov et al., 2003) assigned by eggNOG (Figure 6A). These genes include orthology groups spanning diverse functional roles in contractile cells, including members of the adhesion complex that connects cells, actomyosin networks that drive contractility, and signaling pathways that stimulate contraction (Figure 6B and Supplementary file 5). This observation suggests that contractile cells near the beginning of animal evolution already possessed the broad assemblage of gene modules associated with different functional aspects of derived muscle cell types in extant animals.

We also identified several transcriptional regulators shared among contractile cells (Figure 6B). Previously known core regulators involved in myocyte specification (Brunet et al., 2016) were enriched only in bilaterian (e.g. myod and tcf4/E12) or vertebrate contractile cells (e.g. mef2). In contrast, we found homologs of Muscle Lim Protein (Csrp) and Forkhead Box Group 1 (Larroux et al., 2008) enriched in contractile cells from all seven species. The Fox proteins included FoxC, which is known to regulate cardiac muscle identity in vertebrates (Brunet et al., 2016) and is contractile-specific in all species except schistosome and Spongilla. Notably, we also identified FoxG orthologs to be enriched in three of the four invertebrates (Figure 6—figure supplement 1), suggesting that FoxG may play an underappreciated role in contractile cell specification outside vertebrates. Whether the most conserved regulators are positioned at the center of the cell type specification regulatory complex remains an important question to address in future studies.

For the family of invertebrate multipotent stem cells, we identified 3343 genes that are enriched in at least one cell type pair and observed significant enrichment (p-value<10⁻³) of genes involved in translational regulation such as RNA processing, translation, and post-translational modification (Figure 6C). We also observed cell cycle and DNA replication genes, as expected for actively dividing cells, though these were not the most enriched categories. Shared stem cell genes comprise 979 orthology groups, 17% of which are enriched in all cell types of this family (Supplementary file 5). Importantly, these genes have consistently higher expression in the multipotent stem cells (MSCs) compared to lineage-restricted stem cells (LSCs) and differentiated cells (DCs) for all four species (Figure 6D), and may represent a large, deeply conserved gene module specifically associated with multipotency.

Next, we asked whether there is an MSC-specific transcriptional regulatory program. Notably, we identified a number of transcriptional regulators enriched in MSCs across all four invertebrates. This included TFs from sox and klf orthology groups, the transcriptional coactivator sub1, and several chromatin modifiers, including ssrp1, a subunit of the FACT complex, and kat7, a histone acetyltransferase (Figure 6D and Supplementary file 6). sox and klf are well-known pluripotency factors (Bialkowska et al., 2017; Sarkar and Hochedlinger, 2013), whereas the others have been studied in various processes associated with stem cell functions (MacPherson et al., 2020; Sikder et al., 2019; Zeng et al., 2013). The conserved enrichment of these transcriptional regulators in invertebrate MSCs suggests that their functional role in stem cells may be more phylogenetically ancient than previously appreciated. Determining their roles in establishing and maintaining multipotency across diverse animal taxa is an important avenue of future research.

The source code for SAMap is publicly available at Github (https://github.com/atarashansky/SAMap; copy archived at swh:1:rev:c696585f8fe41ec1599b0720df579f3cb14f935b; Tarashansky et al., 2021), along with the code to perform the analysis and generate the types of plots presented in the figures. We also provide a wrapper function to launch a graphical user interface provided by the SAM package to interactively explore both datasets in the combined manifold. The datasets analyzed in this study are detailed in Supplementary file 1 with their accessions and annotations provided.

The SAMap algorithm contains three major steps: preprocessing, mutual nearest neighborhood alignment, and gene-gene correlation initialization. The latter two are repeated for three iterations, by default, to balance alignment performance and computational runtime.

To annotate the primary zebrafish and Xenopus cell types, the cell subtype annotations provided by the original publications (Briggs et al., 2018; Wagner et al., 2018) are coarsened using a combination of the manual matching and developmental hierarchies. For example, as ‘involuting marginal zone’ in Xenopus is manually matched to ‘non-dorsal margin’, 'dorsal margin' 'non-dorsal margin involuted', and ‘dorsal margin involuted’ in zebrafish, we label these cells as ‘involuting marginal zone’. In cases where the matching is insufficient to coarsen the annotations, we use the provided developmental trees to name a group of terminal cell subtypes by their common ontogenic ancestor. Cell types that do not cluster well in the manifold reconstructed by SAM are excluded from the comparison. These include germline, heart, and olfactory placode cells, as they are mixed with other cell types in the Xenopus atlas. The germline cells are scattered across the reconstructed manifold and do not concentrate in a distinct cluster. The heart cells and olfactory placode cells are inextricably mixed with larger populations of intermediate mesoderm and placodal cells, respectively. Similarly, the iridoblast, epiphysis, nanog⁺, apoptotic-like, and forerunner cells are excluded because they do not cluster distinctly in the zebrafish atlas.

The annotations provided by their respective studies are used to label the cells in the Spongilla, Hydra, planarian, and mouse atlases. To annotate the schistosome cells, we use known marker genes to annotate the main schistosome tissue types (Li et al., 2021). Annotations for all single cells in all datasets are provided in Supplementary file 1.

The combined manifold N is embedded into 2D projections using UMAP implemented in the scanpy package (Wolf et al., 2018) by scanpy.tl.umap with the parameter min_dist = 0.1. The sankeyD3 package (https://rdrr.io/github/fbreitwieser/sankeyD3/man/sankeyD3-package.html) in R is used to generate the sankey plots. Edge thickness corresponds to the alignment score between mapped cell types. The alignment score between cell types a and b is defined as sab=1|ca|+|cb|(∑i∈ca∑j∈cbC1,ij+∑i∈cb∑j∈caC2,ij), where ca and cb are the set of cells in cell types a and b, respectively. Cell type pairs with alignment score less than z are filtered out. By default, z is set to be 0.1.

The network graphs in Figure 5C are generated using the networkx package (https://networkx.github.io) in python. To focus on densely connected cell type groups, we filter out cell type pairs with alignment score less than 0.05.

We define g1 and g2 to contain SAMap-linked genes from species 1 and 2, respectively. Note that a gene may appear multiple times as SAMap allows for one-to-many homology. Let Xa1b2 denote the set of all cells with cross species edges between cell types a1 and b2. We calculate the average standardized expression of all cells from species i that are in Xa1b2: Yi,gi=1|{x,x∈Xa1b2}|∑x∈Xa1b2Z~i,x,gi, where Z~i,x,gi∈R|gi| is the standardized expression of genes gi in cell x. The correlation between Y1,g1 and Y2,g2 can be written as Corr(Y1,g1,Y2,g2)=∑j=1|g1|S(Y1,g1)j∘S(Y2,g2)j, where S(Z) standardizes vector Z to have zero mean and unit variance. We use the summand to identify gene pairs that contribute most positively to the correlation. We assign each gene pair a score: hg=T(S(Y1,g1))∘T(S(Y2,g2)), where T(Z) sets negative values in vector Z to zero in order to ignore lowly-expressed genes. To be inclusive, we begin with the top 1000 gene pairs according to hg and filter out gene pairs in which one or both of the genes are not differentially expressed in their respective cell types (p-value > 10⁻²), have less than 0.2 SAM weight, or are expressed in fewer than 5% of the cells in the cluster. The differential expression of each gene in each cell type is calculated using the Wilcoxon rank-sum test implemented in the scanpy function scanpy.tl.rank_genes_groups.

We use the eggNOG mapper (v5.0) (Huerta-Cepas et al., 2019) to assign each gene to an orthology group with default parameters. For the zebrafish-to-Xenopus mapping, genes are considered orthologs if they map to the same vertebrate orthology group. For the pan-species analysis, we group genes from all species with overlapping orthology assignments. In Figure 6B, each column corresponds to one of these groups. As each group may contain multiple genes from each species, we present the expression of the gene with the highest enrichment score per species. All gene names and corresponding orthology groups are reported in Supplementary file 5.

SAMap outputs gene-gene correlations across the combined manifold for all pairs of genes in the homology graph. As determined by eggNOG, genes that map to the same orthology group for the two species’ most recent common ancestor are considered orthologs, and those that map to the same orthology group more ancestral than Vertebrata are considered as paralogs. We note that as eggNOG does not provide an orthology group corresponding to the osteichtyan ancestor, our analysis does not include the paralogs that duplicated in between the osteichtyan and the vertebrate ancestors. If a gene has significantly higher correlation to one of its paralogs than its ortholog (>0.3 by default), we consider its ortholog to have been substituted. Paralog substitutions are identified using the samap.analysis.ParalogSubstitutions function provided by the SAMap package.

The evolutionary time period in which paralogs were duplicated can be inferred by identifying their most recent shared orthology group. We calculate the enrichment of paralog substitutions for each taxonomic level (i.e. Chordata, Bilateria, Metazoa, Opisthokonta, and Eukaryota) using the eggNOG orthology group assignments. We normalize the number of substituting paralogs by the total number of paralogs at each level to calculate the rate of paralog substitution across evolutionary time.

To quantify the enrichment of substituting paralogs in each cell type, we define a cell type-specific substitution score. We first assign paralog substitution events to cell types if the paralogous gene pairs are enriched in any of their mappings. Each cell type k then has a set of substituting paralogs Pk. The score Sk for cell type k is calculated as Sk=∑i∈Pk1-nimk, where ni is the number of paralogs of ortholog i normalized by the maximum number of paralogs observed across all genes to accounts for the fact that genes with more paralogs are more likely to match with substituting paralogs by random chance, and mk is the number of differentially expressed genes in cell type k. Similarly, the denominator accounts for the fact that cell types with more differentially expressed genes are more likely to have paralog substitutions by random chance. The substitution scores for cell types with annotated homologs across species are averaged.

We generate gene trees to validate the identity of genes involved in putative examples of paralog substitution and of Fox and Csrp transcriptional regulators that are identified as enriched in contractile cells. For this, we first gather protein sequences from potential homologs using the eggnog version 5.0 orthology database (Huerta-Cepas et al., 2019). For the Fox and Csrp phylogenies, we include all Fox clade I (Larroux et al., 2008) and Csrp/Crip homologs, respectively, from the seven species included in our study.

Alignment of protein sequences is performed with Clustal Omega version 1.2.4 using default settings as implemented on the EMBL EBI web services platform (Madeira et al., 2019). Maximum likelihood tree reconstruction is performed using IQ-TREE version 1.6.12 (Nguyen et al., 2015) with the ModelFinder Plus option (Kalyaanamoorthy et al., 2017). For the Csrp tree, we perform 1000 nonparametric bootstrap replicates to assess node support. For Fox, we utilize the ultrafast bootstrap support option with 1000 replicates. For each gene tree we choose the model that minimizes the Bayesian Information Criterion (BIC) score in ModelFinder. This results in selection of the following models: DCMut+R4 (Csrp) and VT+F + R5 (Fox). The final consensus trees are visualized and rendered using the ETE3 v3.1.1 python toolkit (Huerta-Cepas et al., 2016) and the Interactive Tree of Life v4 (Letunic and Bork, 2019).

Using the eggNOG mapper, KOG functional annotations are transferred to individual transcripts from their assigned orthology group. For enrichment analysis, all genes enriched in the set of cell type pairs of interest are lumped to form the target set for each species. For example, the target set for Spongilla archaeocytes used in Figure 6C is composed of all genes enriched between Spongilla archaeocytes and other invertebrate stem cells. Note that this set includes genes from other species that are linked by SAMap to the Spongilla archeocyte genes. We include genes from other species in the target set to account for differences in KOG functional annotation coverage between species. As such, the annotated transcripts from all seven species are combined to form the background set. We use a hypergeometric statistical test (Eden et al., 2009) to measure the enrichment of the KOG terms in the target genes compared to the background genes.

For benchmarking, we use vertebrate orthologs as determined by eggNOG as input to Harmony (Korsunsky et al., 2019), LIGER (Welch et al., 2019), Seurat (Stuart et al., 2019), Scanorama (Hie et al., 2019), BBKNN (Polański et al., 2019), which are all run with default parameters. One-to-one orthologs are selected from one-to-many and many-to-many orthologs by using the bipartite maximum weight matching algorithm implemented in networkx. When using the one-to-one orthologs as input for SAMap, we run for only one iteration. The resulting integrated lower-dimensional coordinates (PCs for Seurat, Harmony, and Scanorama and non-negative matrix factorization coordinates for LIGER) and stitched graphs (BBKNN and SAMap) are all projected into 2D with UMAP (Figure 2—figure supplement 1A). The integrated coordinates are used to generate a nearest neighbor graph using the correlation distance metric, which is then used to compute the alignment scores in Figure 2—figure supplement 1B. The alignment scores for SAMap and BBKNN are directly computed from their combined graphs.

S. mansoni (strain: NMRI) juveniles are retrieved from infected female Swiss Webster mice (NR-21963) at ~3 weeks post-infection by hepatic portal vein perfusion using 37°C DMEM supplemented with 5% heat inactivated FBS. The infected mice are provided by the NIAID Schistosomiasis Resource Center for distribution through BEI Resources, NIH-NIAID Contract HHSN272201000005I. In adherence to the Animal Welfare Act and the Public Health Service Policy on Humane Care and Use of Laboratory Animals, all experiments with and care of mice are performed in accordance with protocols approved by the Institutional Animal Care and Use Committees (IACUC) of Stanford University (protocol approval number 30366). In situ hybridization experiments are performed as described previously (Tarashansky et al., 2019), using riboprobes synthesized from gene fragments cloned with the listed primers: collagen (Smp_170340): GGTGAAGAAGGCTGTTGTGG, ACGATCCCCTTTCACTCCTG; tropomyosin (Smp_031770): AAGCTGAAGTCGCCTCACTA, CATATGCCTCTTCACGCTGG; troponin (Smp_018250): CGTAAACCTGGTCAGAAGCG, ATCCTTTTCCTCCAGAGCGT; myosin regulatory light chain (Smp_132670): GAGACAGCGAGTAGTGGAGG, TGCCTTCTTTGATTGGAGCT; wnt11 (Smp_156540): TGTGGTGATGAAGATGGCAG, CCACGGCCACAACACATATT; frizzled (Smp_174350): CGAACAGGCGCATGACAATA, TGCTAGTCCTGTTGTCGTGT.