Cross-species integration of single-cell data reveals conserved pathology-associated cell populations across animal models and human samples Authors: Cancheng Li, Hongtao Sang, Dayong Yue, Rong Fu, Kunpeng Yang, Hongwei Zhang, Zixin Hu, Xun Gu, Huimin Zhang, Sidong Xiong, Hang Ruan Published: 2026 ABSTRACT SummarySingle-cell RNA sequencing (scRNA-seq) enables high-resolution profiling of cellular heterogeneity, but integrating data across species remains challenging due to technical variation and complex gene homology. We present TACMAN (transformer-based alignment of cross-species metapath aggregation network), a computational framework for cross-species scRNA-seq integration that combines a metapath-based heterogeneous graph neural network with an encoder-only transformer. TACMAN aligns conserved cell types across species under normal physiological conditions while preserving biological signals. We demonstrate its utility by integrating clinical human and mammalian model scRNA-seq data, revealing conserved cell subtypes in tumor, inflammatory, and infectious diseases. Notably, using our in-house single-cell transcriptomic atlas of an evolutionarily distant Caenorhabditis elegans germline tumor model, TACMAN identifies tumor-related cell populations conserved in human testicular germ cell tumor samples, enabling cross-species comparison under pathological conditions. TACMAN thus offers a powerful tool for comparative single-cell analysis, advancing translational research using animal models.Keywords: comparative transcriptomics, single-cell RNA sequencing, cross-species integration, single-cell deep learning, C. elegans germline tumor model, cell-type annotation transfer, disease model translation Introduction Cellular composition and gene expression vary widely across tissues and disease states. Understanding these dynamics at a high resolution is essential for advancing our knowledge of cell biology. Single-cell RNA sequencing (scRNA-seq) can profile every cell in a sample, revealing complex tissue ecosystems and rare cell phenotypes that bulk assays may miss. For example, macrophages are key immune cells involved in homeostasis, innate defense, and antigen presentation, bridging innate and adaptive immunity.1 They respond to environmental cues and can change phenotypes during disease, such as tumor-associated macrophages, which promote cancer and are potential therapeutic targets.2,3 In general, scRNA-seq has enabled mapping of diverse immune and non-immune cell populations in diseases such as cancer and inflammatory disorders, distinguishing conserved cell programs from pathology-specific cells at single-cell resolution. Studies of human diseases often rely on genetically tractable models (e.g., mice, hamsters) to overcome the limited access to patient tissues. Cross-species profiling is crucial in identifying which animal models most accurately reflect human pathology. With the advent of scRNA-seq, it is now possible to directly compare cell states across species. For example, a study comparing hamster models with COVID-19 patients has optimized animal model selection to match the varying severity levels of COVID-19.4 Recent research on human diseases, including cancers,5,6 infectious diseases,7 and autoimmune diseases,8 has used cross-species integrative analysis of scRNA-seq data from patients and disease animal models. However, differences in key cellular gene networks among species can make it difficult to generalize findings across organisms. Human genes have complex homologous relationships with their animal counterparts, and these species differences can confuse the direct interpretation of results. This aspect of species divergence is often overlooked, yet it is essential for accurately studying immune-mediated diseases.9 For more distant species, such as Caenorhabditis elegans, there are advantages over mammalian models, including a fast growth rate, lower cost, and non-invasive optical monitoring and manipulation methods. However, cross-species analysis between humans and C. elegans remains limited due to substantial biological differences. Combining scRNA-seq data from different experiments or species requires correcting for technical and biological batch effects. Variations in protocols and sequencing runs, as well as biological factors such as tissue source or species, can dominate the signal. To address this, one maps genes across species using orthologs and employs integration algorithms to align cell profiles. For instance, Harmony10 and Liger11 integrate datasets through an iterative clustering process. Several approaches utilize mutual nearest neighbors (MNN) to remove batch effects, including scran,12 Scanorama,13 and FastMNN.12 Seurat (v.3 and above)14 uses canonical correlation analysis (CCA) and MNN to identify anchor cells between datasets and perform joint dimension reduction on both datasets. Deep learning techniques have become prominent in data integration tasks. For instance, scVI15 employs deep generative modeling for single-cell transcriptomics. Recent scRNA-seq data integration models leverage graph neural networks to learn higher-order representations and topological relationships in cell-cell graphs rather than solely extracting representations from individual cell expression data.16,17,18 Advancements in large language models, such as the transformer model, have accelerated single-cell omics research by incorporating semantic attention information into various data analysis tasks, including cell-type annotation.19 CIForm20 and TOSICA21 are recent approaches that implement multi-head self-attention deep learning models based on the transformer architecture. They enable interpretable cell-type annotation using biologically meaningful entities such as the Gene Ontology,22 although they did not account for the complexity of cross-species scenarios. In particular, a strong “species effect” often emerges when combining scRNA-seq data from different species: cells cluster by species rather than by true cell identity.23 We propose that it is crucial to consider the complex homologous relationships that arise from gene duplication events before or after speciation. These relationships include one-to-one orthologs with a single copy in each species and one-to-many/many-to-many orthologs, which can arise from species-specific gene duplications.24 When integrating cross-species scRNA-seq data, the focus is usually on one-to-one orthologs, while other relationships are either overlooked or simplified.23 Recent approaches have integrated additional orthologs for cross-species datasets, preserving information from species with distant evolutionary relationships. SAMap25 builds a gene-gene graph using BLAST for similarity, projects single-cell expressions into a common principal component space, and merges datasets through iterative updates to cell-cell and gene-gene graphs. CAME implements a recurrent relational graph neural network with homologous relations and gene expressions,26 while TACTiCS employs relationships of homologs through protein sequence data.27 However, integrating datasets from evolutionarily distant sources remains challenging, as even high-performing methods need to balance combining shared cell types with maintaining species-specific features. Additionally, previous methods have not evaluated their effectiveness in pathological contexts, such as when integrating scRNA-seq data from animal models and human samples. To address these challenges, we introduce TACMAN (transformer-based alignment of cross-species metapath aggregation network), a method to integrate cross-species scRNA-seq data. TACMAN employs a heterogeneous graph neural network and an encoder-only transformer model to capture both high-order cross-species semantic representations. This architecture captures higher-order relationships between cells using graph metapaths, resulting in a unified embedding in which cells from different species with shared metapath semantics cluster together. We benchmark TACMAN against existing integration methods (using only one-to-one orthologs or including duplicates) on datasets of human and model organisms under healthy conditions. We then apply TACMAN to scRNA-seq data from several diseases, focusing on immune, tumor, and stromal cells in patient samples and their corresponding animal models. TACMAN successfully identifies conserved disease-associated cell states across species. More importantly, to assess TACMAN’s ability to integrate scRNA-seq data from distantly related species, we generated the first single-cell transcriptomic dataset from a C. elegans germline tumor model. We then used TACMAN to perform cross-species integration with human testicular germ cell tumor (TGCT) to investigate phenotypic similarities among tumor cells of these two distant species. This creates a unique resource for studying cross-species molecular mechanisms in pathological conditions. Together, these results demonstrate TACMAN’s potential to align single-cell atlases from pathological samples and improve the translation of insights from models to human disease. Results The architecture of TACMAN To capture cross-species information, TACMAN uses scRNA-seq expression matrices from both the reference and query species, along with homologous relationships, as input (Figure 1A). TACMAN then adopts a heterogeneous graph approach to represent relationships among node categories, such as genes and cells across species. After preprocessing the expression matrix, cells and genes with related information (see STAR Methods) are used as input features to create heterogeneous graphs, including connections between cells and genes (cell-gene graph, Figure 1B) and between homologous genes (gene bipartite graph, Figure 1B). The connections among cells within the same species' datasets are established using the K-nearest neighbors (KNN) method. Subsequently, a hybrid graph is formed (Figure 1B). In this way, TACMAN generates a cross-species single-cell heterogeneous graph (Figure 1C) that includes non-one-to-one homologous genes as input features. Additionally, TACMAN employs an MNN method to construct an inter-species cell graph among homologous cells from scRNA-seq datasets.28 The architecture of TACMAN (A) TACMAN integrates cross-species single-cell RNA sequencing (scRNA-seq) data. The inputs include (1) a well-annotated reference dataset, such as human clinical samples spanning healthy and diseased states; (2) a query dataset, typically from animal models, to align cell types with the reference; and (3) gene homology information, provided as cross-species gene adjacency matrices or homologous gene mappings. (B) Highly variable genes and differentially expressed genes are identified within the datasets. Shared genes are filtered using homologous gene mappings to construct a gene-gene graph. A cell-gene graph is built based on the gene expression matrix, and cell-cell connections are established in a low-dimensional space using K-nearest neighbors (KNN) and mutual nearest neighbors (MNN). (C) A heterogeneous graph is constructed using predefined cross-species metapaths, such as Cell-Cell-Gene-Gene-Cell (CCGGC) and Cell-Gene-Gene-Cell (CGGC), to capture biologically meaningful relationships. Mean aggregation consolidates information from metapath-based neighbors to generate semantic feature vectors for each cell. (D) During each training epoch, semantic feature vectors undergo projection through transformer layers to produce classification outputs. The model is optimized using a composite loss function that combines cross-entropy (CE) for classification accuracy and mean-squared error (MSE) for cross-species cell similarity, as measured by MNN. Multi-stage training further refines the model after several epochs. (E) TACMAN outputs include low-dimensional embeddings of cross-species-aligned cells and posterior probabilities for cell type predictions in the query species. The figures were created with the help of BioRender.com. TACMAN leverages a metapath-based approach to enhance the interpretability and efficiency of the heterogeneous graph representations in cross-species scRNA-seq data integration. A metapath is defined as a sequence of node and edge types that encapsulate specific semantic relationships between entities. Metapath-based models first capture structural information through defined sequences and aggregate semantic information.29 Here, we define cross-species metapaths to generate species-level semantic information for individual cells. This involves predefining several metapaths grounded in biologically reasonable contexts. For example, CGC (Cell-Gene-Cell) captures the co-expression relationship between different cell types. CGGC (Cell-Gene-Gene-Cell) extends this to include gene homology and describes the co-expression relationship between cells, irrespective of their cell type or species origin (Figure 1C). Other metapaths, such as CC (Cell-Cell), CGGCC (Cell-Gene-Gene-Cell-Cell), and CCGGCC (Cell-Cell-Gene-Gene-Cell-Cell), are employed to represent various inner and cross-species semantic relationships. Then, TACMAN generates a cell feature matrix using a metapath-based neighbor-mean aggregation during the pre-processing phase (Figure 1C). Following pre-processing, TACMAN generates unique, semantically specific transformation matrices for each training epoch and projects the data into low-dimensional embeddings (Figure 1D). These embeddings are then fed into a transformer, which uses attention mechanisms to evaluate the relevance of different semantic information for each metapath.30 The final embedding vector is generated by concatenating all output vectors, with an additional multilayer perceptron used for downstream tasks to infer and integrate cell labels in the query data with the reference data (Figure 1D). Furthermore, TACMAN employs multi-stage training techniques to incorporate data from the query set and avoid overfitting reference data. To enhance single-cell integration across species, TACMAN’s loss function calculates cross-entropy loss between predicted cell type probabilities and reference labels (Figure 1D). It also incorporates similarity loss in low-dimensional embeddings between nearest neighbor cells across species28 (Figure 1D). After each training epoch, TACMAN can extract low-dimensional features for the next stage of training. The final output comprises integrated cross-species cell embeddings, which can be visualized using uniform manifold approximation and projection (UMAP), and posterior probabilities indicating the alignment of query cells with reference cell types (Figure 1E). TACMAN enables stable cross-species alignment of homologous cell types under normal physiological conditions We first tested TACMAN on scRNA-seq data from eight tissues across multiple datasets deposited in the Human Cell Landscape (HCL)31 and the Mouse Cell Atlas (MCA)32 data portals. The eight tissues include essential organs like the brain, heart, and kidney, along with the spleen, which exhibits structural and functional differences between human and mouse33 (Table S1). We evaluated TACMAN’s performance in integrating cross-species scRNA-seq data against five established methods. These include SAMap and CAME, which use one-to-many and many-to-many orthologs, and Seurat, TACTiCS, and TOSICA, which depend on one-to-one orthologs. TACMAN consistently ranks among the top methods for integrating human and mouse spleen, as shown by the weighted F1 score and accuracy (Figures 2A and 2B; Table S2). And TACMAN generally ranks in the top three methods for integrating other tissues (Figures S1A and S1B; Table S2). All six integration methods exhibit good performance when integrating human and mouse brains, likely due to conserved expression patterns in mammalian brains,34 as shown in our previous comparative analysis of bulk RNA-seq data.35,36 Further examination of the brain UMAP plots of cross-species embeddings reveals that integrations by Seurat, TOSICA, SAMap, and CAME retained dataset-specific neural cell clusters (Figures S1C and S1D), suggesting unresolved batch effects. TACMAN effectively mitigated these batch effects, enabling clustering of major cell types across species into distinct, biologically meaningful groups. Moreover, TACMAN’s integrated embedding preserved the differential trajectory from stem cells to neural cells in the fetal human and mouse brains (Figure 2C). Only a small fraction of stem cells was assigned to adult human samples, consistent with the expectation that stem cells are predominantly found in fetal brain tissue. TACMAN demonstrates robust performance in aligning homologous cell types (A and B) We evaluated six computational methods, Seurat, TACTiCS, TOSICA, SAMap, CAME, and TACMAN, for cross-species cell type assignment in the normal spleen, using data from the Human Cell Landscape (HCL) and Mouse Cell Atlas (MCA). Performance was measured with the weighted F1 score (A) and accuracy (B). For each method, the mean and 95% confidence interval for the metric were calculated based on six repeated runs. The error bars represent the 95% confidence interval. Colored values above the error bars are the means. (C) UMAP visualizations of TACMAN-generated cell embeddings are presented: the top panel is colored by developmental stage, and the bottom panel shows the stage composition of stem and neural cells. (D and E) We assessed the impact of including non-one-to-one orthologs on cross-species cell type assignment using a three-pair scRNA-seq retina dataset. Zebrafish served as the reference species, with chicken, mouse, and human as queries. TACMAN, SAMap, and CAME were evaluated with (+) and without (−) non-one-to-one orthologs. Performance metrics included the weighted F1 score (D) and accuracy (E), with values reported for TACMAN. Red highlights indicate TACMAN’s best performance, while orange highlights denote second or third place rankings. To further assess the performance of TACMAN in integrating distant animal-model species, we used retinal scRNA-seq datasets from four species: zebrafish,37 mouse,38 human,39 and a non-model species, chicken.37 We applied all six methods using adult zebrafish retina scRNA-seq data as the reference and chicken, mouse, and human as queries. We compared the results of using only one-to-one homology with those that included one-to-many and many-to-many orthologous information (Table S3). Tests demonstrated that TACMAN ranked highest among the six methods when evaluated using weighted F1 scores and accuracy metrics (Figures 2D and 2E; Table S2). Additionally, incorporating one-to-many and many-to-many orthologs significantly enhanced the performance of TACMAN, SAMap, and CAME. These methods outperform Seurat, TACTiCS, and TOSICA (Figures 2D and 2E), which rely solely on one-to-one homology. We conducted experiments to assess how variations in gene selection, such as highly variable genes and differentially expressed genes, and network architecture, including the number of features and hidden layers, influence TACMAN’s performance (see STAR Methods). Results showed that the choice of input genes or hyperparameters had a limited impact on TACMAN’s ability to align cell types across species (Figures S2A and S2B; Table S4). Increasing the number of hidden layers could modestly improve alignment accuracy but may be more computationally expensive. We further evaluated the effects of data size on performance in the retina datasets. We randomly selected a gradient of data proportions (25%, 50%, 75%, and 100%), with each proportion repeated three times. TACMAN’s performance remained stable across these data proportions (Figures S2C and S2D). TACMAN integration preserves phylogenetic signals Homology between distant species is more complex because of increased one-to-many and many-to-many orthologs (Figure 3A; Table S3). To determine if these metapaths capture phylogenetic signals, we performed ablation experiments by selectively removing individual metapath structures from the heterogeneous graph. A key architectural feature of TACMAN, compared to other neural network-based integration methods, is its inclusion of phylogenetic-aware metapaths. We found that removing metapaths decreased zebrafish-to-human alignment performance more significantly than in other species (Figures 3B and 3C). These results suggest that homology-aware metapaths help preserve evolutionarily conserved transcriptional semantics. TACMAN integration preserves phylogenetic signals (A) Bar plots depict the number and percentage of orthologous genes shared among four species—human, mouse, chicken, and zebrafish—categorized into one-to-one, one-to-many, and many-to-many orthologous relationships. (B and C) Ablation experiments show the performance of TACMAN alignment between zebrafish retina and three other species (chicken, human, and mouse) when removing individual phylogenetic-aware metapath structures such as Cell-Cell-Gene-Gene-Cell (CCGGC) and Cell-Gene-Gene-Cell (CGGC). Performance metrics include accuracy (B) and weighted F1 score (C). (D) The neighbor-joining phylogenetic tree was constructed from a pairwise Euclidean distance matrix among spleen macrophages from six species: human, macaque, rat, mouse, pig, and zebrafish, projected into a shared 64-dimensional space. Bootstrap values for internal nodes were estimated by resampling macrophage populations 1,000 times. Research indicates that transcriptome evolutionary distance closely reflects species divergence, although gene expression evolution rates differ among organs.35,40 To evaluate whether the embeddings generated by TACMAN preserve evolutionary distance across species, we analyzed an scRNA-seq dataset comprising spleen immune cells from six species. By projecting cell embeddings into a shared space and computing pairwise Euclidean distances, we constructed neighbor-joining (NJ) phylogenetic trees for the macrophages and their subsets from the resulting distance matrix (see STAR Methods). The NJ trees for all macrophages and their subsets are annotated based on their normal physiological functions related to inflammation, iron recycling, and phagocytosis and align well with established taxonomic relationships (Figures 3D, S3A, and S3B), indicating that TACMAN effectively preserves phylogenetic signals in its integrated cell embeddings. TACMAN integration reconstructs conserved infiltrated myeloid cells between lung cancer patients and murine tumors To evaluate TACMAN’s effectiveness in integrating cross-species scRNA-seq data in a pathological context, we first applied TACMAN to a well-analyzed dataset from a recent study of human lung cancer (LC) patients and murine tumors.6 The study focused on tumor-infiltrating myeloid cells (TIMs) and demonstrated structural similarities in TIM populations through clustering analysis and cross-species mapping of gene expression markers. We aim to test whether TACMAN’s de novo integration can reconstruct TIM structures. We integrated LC patients and murine tumor-infiltrated myeloid cells using TACMAN with human data as a reference (Figures S4A and S4B; Table S1). The co-embeddings captured biologically meaningful patterns. For example, red blood cells (RBCs) were clearly separated from white blood cells in the UMAP space, with no mouse cells misclassified as RBCs. Plasma cells, which were absent in the mouse data, showed minimal mislabeling, with only six cells being incorrectly assigned. TACMAN accurately aligned major categories of mouse immune cells with their human counterparts, achieving a weighted F1 score and accuracy of approximately 0.96. This performance exceeds that of four other evaluated methods and is comparable to CAME (Figures S4C and S4D). Human dendritic cells (DCs) are originally classified into five subclusters: DC1, DC2, DC3, plasmacytoid DCs, and monocyte-derived DCs (monoDCs). TACMAN de novo integration using a human reference aligns with the original mouse DC annotations, demonstrating conserved gene expression patterns (Figures 4A and 4B). We further verified the markers of DC populations (Figure S4E). MonoDCs exhibited high CLEC10A and CD74 expression, with low expression of cDC1 and cDC2 markers.6 Both human and mouse DC1 expressed XCR1, while the CCR7+ DC3 subgroup, previously shown to be located near tumor blood vessels, interacted with CXCR6+ tumor-infiltrating T cells, enhancing their cytotoxicity.41 Plasmacytoid DCs expressing IRF7 and TCF4 were distinctly separated from other DCs on UMAP (Figure 4A). For comparison, we used five other approaches to integrate cross-species DC populations. To evaluate integration accuracy, we used a proxy where cell annotations consistently identified by multiple methods were considered more likely to be correct, which does not constitute definitive ground truth. Under this assumption, TACMAN identified the highest number of DCs cross-verified by at least two other methods (Figure S4F). TACMAN reconstructs conserved infiltrated myeloid cells in lung cancers between patients’ samples and murine models (A) UMAP plots show TACMAN-generated embeddings of dendritic cells (DCs) from lung cancer (LC) human patients (reference, R) and murine tumors (query, Q), with colors indicating cell types of human and mouse DC subsets. The cellular composition of DCs is shown at the bottom. (B) A dot plot compares original mouse DC annotations (mouse O, y axis) with query mouse annotations (mouse Q, x axis) transferred via TACMAN from human-integrated labels. Dot size represents the proportion of cells with a given original label assigned to a transferred label, while color reflects the median posterior probability of assignment. (C) UMAP plots illustrate TACMAN-derived embeddings of macrophages (Mac) from LC human patients (R) and murine tumors (Q), colored by cell type. The associated cellular compositions of human and mouse macrophage subsets are provided. (D) A dot plot presents the original mouse macrophage annotations (mouse O, y axis) alongside the query annotations (mouse Q, x axis) transferred from TACMAN-integrated human labels. (E) M1 and M2 phenotype scores for macrophage subsets in both LC patients and murine tumors. Macrophages are classified into nine subsets (Mac 1–9) in LC patients and four (Mac 1–4) in murine tumors. Consistent with the original analysis, TACMAN integration aligned mouse macrophage subsets Mac1 and Mac2 with a coherent gradient of human macrophages (Figures 4C and 4D). Mouse Mac4, annotated as alveolar macrophages (AMs), aligns with human Mac7 by expressing canonical tissue-resident AM markers, such as MARCO/Marco and Siglecf/SIGLEC542 (Figure S4G). This Mac7 population could also be cross-verified by other methods (Figure 4C). Compared with other methods, TACMAN identified the highest number of cross-verified macrophages (Figure S4H). As anticipated, tumor-infiltrating macrophages predominantly exhibit an M2 phenotype associated with tumor promotion. The macrophage polarization signature is largely consistent between LC patients and murine tumors (Figure 4E). To further investigate whether cross-species co-embedding reflects biological conservation, we designed a concise conservation checklist evaluation framework to assess transcriptional similarity using differentially expressed genes (see STAR Methods). This framework helps determine which cells are relatively more conserved in a disease context. Using it, we compared two macrophage subsets, Mac7 and Mac8. Mac7 is generally more co-embedded between human and mouse (Figure S4I; Table S5) and shows higher concordance at both the gene expression and pathway levels. Conversely, Mac8 is less co-embedded and shows fewer similarities (Figures S4J–S4L). TACMAN integration identifies conserved proliferative inflammatory fibroblast subsets with extracellular matrix remodeling capacity in rheumatoid arthritis Fibroblasts and macrophages are present in all tissues and likely communicate directly, affecting the microenvironment and disease progression.43 Recent scRNA-seq data indicate that proliferative inflammatory fibroblasts, identified as HLA-DRAhi sublining fibroblasts, may act as key disease mediators in the synovial tissue of rheumatoid arthritis (RA) patients.44 Another study revealed that NOTCH3 signaling is essential for the proliferation of pathological fibroblasts, primarily sublining and perivascular fibroblasts, in RA mouse models.45 To examine whether the proliferative inflammatory fibroblast subtypes are conserved, we used TACMAN to integrate fibroblasts from the synovial tissue of RA patients and mouse models (Figures 5A, 5B, and S5A). TACMAN integrates the lining fibroblasts of RA patients and mice into a single cluster (Figure 5A). Other sublining and perivascular mouse fibroblasts were aligned to sublining human fibroblasts with their marker genes matched (Figures 5B and 5C). The HLA-DRAhi sublining fibroblasts are a group of CD34− fibroblasts that express genes related to inflammation and interferon, as confirmed in RA patients. The degradation of the extracellular matrix, including collagen formation and IL-17A signaling, was significantly enriched in the marker genes of HLA-DRAhi sublining fibroblasts in mice (Figure S5B). This suggests that murine cells categorized as HLA-DRAhi sublining fibroblasts by TACMAN possess extracellular matrix remodeling and inflammatory response functions consistent with those of HLA-DRAhi sublining fibroblasts in RA patients. Additional evaluation shows that they are relatively conserved, with concordance in antigen processing pathways (Table S5). TACMAN integration identifies conserved proliferative inflammatory fibroblast subsets with extracellular matrix remodeling capacity in rheumatoid arthritis (A) Cell embeddings generated by TACMAN from scRNA-seq data of synovial fibroblasts in human rheumatoid arthritis (RA) patients (reference, R) and RA mouse models (query, Q), with cells colored by fibroblast subtype. (B) A dot plot illustrates the posterior probabilities of mouse fibroblast annotations transferred via TACMAN integration. Dot size represents the proportion of cells from each original mouse label (mouse O) assigned to a corresponding query label (mouse Q), while color intensity indicates the median posterior probability. (C) Expression patterns of fibroblast subtype marker genes are shown in a dot plot for both human and mouse datasets. Dot size reflects the proportion of cells within each subtype expressing the marker, and color denotes the Z score of counts per million (CPM) expression values. (D) The distribution of HLA-DRAhi sublining fibroblasts, identified in the human reference via TACMAN integration, is compared across three mouse experimental groups: RA mice treated with an anti-Notch3 antibody versus isotype control (RA WT anti-NRR3 vs. RA WT ISO), Notch3 knockout RA mice versus wild-type controls (RA KO-Notch3 vs. RA WT), and Notch3 knockout mice versus wild-type controls under healthy conditions (Healthy KO-Notch3 vs. Healthy WT). Differences in cell proportions were tested using a chi-square test. (E) A comparative analysis of six integration approaches evaluates their effectiveness in recapitulating the dependency of annotated HLA-DRAhi sublining fibroblasts on upstream perturbation of NOTCH3 signaling pathways. Differences in cell proportions were tested using a chi-square test. To further validate whether the proliferation of TACMAN-annotated HLA-DRAhi sublining fibroblasts depends on upstream NOTCH3 signaling, we found that the proportion of HLA-DRAhi sublining fibroblasts in the synovial tissue of healthy mice decreased after Notch3 knockout and increased after RA induction, while it decreased following treatment with anti-NRR3 antibodies to inhibit NOTCH3 signaling (Figure 5D). TACMAN integration for replicating perturbations of fibroblast subsets is robust across different input feature selections (Figure S5C) and hyperparameter settings (Figure S5D). We tested other approaches to restore the patterns under perturbation. However, TOSICA and SAMap did not replicate the decrease in HLA-DRAhi sublining after antibody treatment (Figure 5E). TACMAN integration identifies primate-conserved macrophage subsets linked to sex differences in COVID-19 severity Primate models, especially macaques, are considered more physiologically, immunologically, and genetically comparable to humans than mouse models, making them better suited to modeling human disease pathogenesis.46 We integrated data from COVID-19 patients with a primate (macaque) infection model using TACMAN. The macaque dataset includes well-annotated macrophage populations across different stages of COVID-19 infection (Table S1). This integration allows us to evaluate TACMAN’s ability to align corresponding cell populations across species in a clinical context. We integrated time-stamped bronchoalveolar lavage (BAL) samples collected from macaque models of three stages of COVID-19 infection (before infection, acute infection, and endpoint) as a reference and severe COVID-19 patients as a query using TACMAN (Figure S6A). TACMAN aligned human immune and non-immune cell types with macaque references (Figure S6B). Macrophages were the largest immune group (Figure S6A). Mac, Mac2, and Mac3 macaque subsets appear in pre-infected samples and reappear as inflammation declines at necropsy. The proportion of Mac2 and Mac3 lacking the IFN signature is significantly depleted during the acute infection stage, when macrophage subsets with high IFN signatures, such as Mac_IFN_1, are populated. Although the integration shows that the predominant Mac_IFN_1 population from macaques and humans do not cluster together, it suggests that they may not be largely conserved (Figure 6A). TACMAN integration identifies primate-conserved macrophage subsets linked to sex differences in COVID-19 severity (A) UMAP plots display cell embeddings generated by TACMAN using scRNA-seq data from bronchoalveolar lavage (BAL) macrophages of COVID-19-infected macaques (reference, R) across three disease stages (pre-infection, acute, endpoint), and human patients (query, Q). Cells are colored by cell type and subtype in macaques and humans. Four distinct macrophage subsets, Mac, Mac2, Mac_IFN_1, and Mac_TREM2_IFN, are highlighted with circles. (B) UMAP plots show the expression of marker genes for Mac_TREM2 and Mac_TREM2_IFN subsets in both macaques and humans. Colors represent log-transformed counts per million (CPM) expression values. (C) Boxplots illustrate gender-associated differences in the proportions of Mac2 and Mac_TREM2_IFN subsets among patients. Boxes indicate the median and interquartile range. Statistical significance was determined using a two-sided Mann-Whitney U test (∗p < 0.05). Two subsets of macrophages with elevated TREM2 (triggering receptor expressed on myeloid cells 2) expression have been identified (Figures 6A and 6B). The Mac_TREM2 subset was abundant in the BAL of healthy macaques. During infection, TREM2+ macrophages transitioned to an interferon-responsive phenotype, termed Mac_TREM2_IFN, before returning to a low-interferon state by the endpoint (Figure 6A). A separate study on COVID-19 infection in macaques examined the TREM2+ macrophage populations, the main contributors to inflammatory cytokine production.47 Another study on humans indicated that the TREM2 protein could recognize SARS-CoV-2 membrane protein and activate the T helper cell 1 response.48 Notably, the TACMAN integration shows that this Mac_TREM2_IFN subset is conserved in both infected macaques and severe COVID-19 patients (Figures 6A and 6B). The Mac_TREM2_IFN exhibits an M1-biased status in infected macaques and COVID-19 patients (Figure S6C). It also expresses monocyte markers, such as CD14 and CD16 (encoded by the FCGR3 gene in macaques and the FCGR3A gene in humans), indicating that they are primarily monocyte derived in both macaques and humans (Figure S6D). We further validate that Mac_TREM2_IFN is conserved at the pathway expression level (Figure S6E), and is also detected in another COVID-19 patient cohort when aligning with macaque data (Figures S6F and S6G). Mac_IFN_1 is depleted in this cohort, which further suggests that Mac_IFN_1 is more likely a macaque-specific subset (Figure S6F). Further evaluation shows that Mac_TREM2_IFN exhibits higher concordance in ortholog expression, especially in the interferon program, while Mac_IFN_1 shows much less concordance in ortholog expression (Table S5). Another observation is that the proportions of Mac_TREM2_IFN and Mac2 subpopulations differ notably between male and female patients with severe infection. Specifically, Mac2 does not exhibit an IFN response with a lower expression of IFI27, whereas Mac_TREM2_IFN displays a strong IFN response (Figures 6A–6C). Mac2 cells in macaques are depleted during the acute infection stage, but they reappear at the pre-infection and endpoint stages. In contrast, Mac2 cells are present in some patients with severe COVID-19, especially females. This indicates that male patients with severe COVID-19 exhibit higher IFN reactivity in certain macrophages than female patients. Research shows that male patients are more likely to experience severe disease and have a higher mortality risk than female patients,49 and IFN treatment in the later stages of COVID-19 infections has been linked to increased mortality.50 TACMAN’s COVID-19 data integration from macaques and humans shows a sex-based difference in conserved TREM2+ macrophages with higher interferon signatures, indicating a possible link to sex-biased clinical outcomes. TACMAN integration reveals conserved ER-stress tumor phenotype shared by C. elegans germline tumor model and human TGCT cells To further evaluate TACMAN’s overall adaptability, we tested its performance in a particularly extreme scenario—aligning across distant evolutionary lineages possible within our current dataset. Specifically, we aimed to see if TACMAN could effectively integrate a canonical “Notch-activated” C. elegans germline tumor model51 with human TGCTs.52 This scenario serves as an ultimate test of algorithmic flexibility, examining whether TACMAN can account for extensive evolutionary divergence and still uncover biologically relevant links among pathological cell states. To date, no study has reported single-cell transcriptomic profiling of a C. elegans germline tumor model. In this study, we present the first such dataset: an in-house scRNA-seq atlas derived from the GC833 tumorous strain. This strain carries the gain-of-function (gf) allele ar202 of the Notch receptor glp-1, which develops germline tumors where germ cells remain in the mitotic cycle.51,53 For comparison, we also include data from the wild-type N2 strain. This dataset enables cellular-level analysis of germline tumor heterogeneity in worms (Figures 7A–7C). The C. elegans data were filtered using standard quality control, retaining samples with at least 200 genes and 500 reads and less than 20% mitochondrial gene expression. Consistent with previous reports, the glp-1(ar202gf) mutant strain shows upregulation of glp-1 expression and corresponding transcriptional changes of its downstream target genes in germline cells (Figures 7D and 7E).53,54,55,56 The clustering analysis identified two unique germline subsets in the glp-1(ar202gf) mutant strain (Figures 7B, 7C, and S7A), one of which exhibits an enhanced secretory and apoptotic phenotype (secretory-apoptotic germline tumor). Another germline subset exhibits activation of endoplasmic reticulum stress (ER-stress) and the unfolded protein response (UPR) program, characterized by higher expression of the ER marker sec-16a.2 and stress markers adm-2/adm-4 (Figures 7F and S7A). This conserved protective process suppresses the glp-1-driven tumor phenotype in C. elegans.57 Additionally, ongoing ER-stress and UPR signaling are common features across human cancers, promoting survival, plasticity, immune evasion, and resistance to therapy.58,59 TACMAN integration reveals conserved ER-stress tumor phenotype shared by C. elegans germline tumor model and human TGCT cells (A) Grayscale image showing DAPI staining of wild-type and the glp-1(ar202gf) mutant adults of C. elegans. Scale bar, 100 μm. (B) UMAP of the scRNA-seq atlas generated from wild-type and glp-1(ar202gf) strains, colored by major cell populations. (C) Per-category cell-type fractions (percentage of total cells) for the two C. elegans strains. (D) UMAP colored by glp-1 expression for wild-type and glp-1(ar202gf) strains. (E) Comparison of average log-normalized expression of glp-1 downstream genes (sygl-1, lst-1, fbf-2, and gld-1) across conditions using a two-sided Mann-Whitney U test with significance indicated as ∗∗∗∗p < 0.0001. Bars indicate mean expression values, and error bars represent the standard errors of the mean. (F) Functional enrichment of differentially expressed genes between ER-stress germ tumors and germline. (G) UMAP illustrating cell embeddings integrated by TACMAN using all cells of C. elegans glp-1(ar202gf) strain as the reference and human TGCT as the query, based on all orthologs. (H) Joint UMAP illustrating TACMAN integration, annotated by cell type. Outlined regions highlight the C. elegans endoplasmic reticulum stress (ER-stress) and secretory-apoptotic germline tumor subsets and two human germline tumor subsets. (I) Cross-species similarity between C. elegans germline subsets and human germline tumor subsets. The size of each dot represents the proportion of human labels assigned to each C. elegans cell category, while the color intensity indicates the median posterior probability, reflecting the confidence of annotation transfer. We further used TACMAN to integrate this C. elegans atlas with human TGCT clinical data (Figure S7A), in which tumor cells from TGCT subtypes (seminoma, embryonal carcinoma, and mixed) cluster closely (Figure S7B). TACMAN effectively aligns human germline tumor cells with C. elegans ER-stress glp-1(ar202gf) germline tumor cells despite the significant evolutionary distance (Figures 7G–7I). Additionally, C. elegans muscle cells can be aligned with human fibroblasts/peritubular myoid cells, indicating that features of homologous cell types are well captured in the TACMAN integration. For comparison, we found that using only one-to-one orthologs in the integration is insufficient to align germline tumor cells across humans and C. elegans (Figures S7C–S7F). Human tumors group with germline cells rather than with more cancerous ER-stress germline subsets. Additional evaluations show that UPR response programs are consistent in human germline tumors and C. elegans (Table S5), although in general, correlation between differentially expressed orthologs is limited given the large species gap. Discussion In this paper, we introduce TACMAN, a framework for integrating scRNA-seq data across species. TACMAN employs a metapath-based heterogeneous graph neural network to model complex biological relationships among cells and genes across species. The mean neighbor aggregator in the preprocessing phase is more efficient at reducing training time, particularly when applied to significantly larger datasets than those currently used. By leveraging biologically meaningful metapaths, such as the cross-species “cell-gene-gene-cell” pathway, TACMAN effectively retains crucial cross-species information, as shown in our ablation experiments. Further investigation of metapath feature weights could improve TACMAN’s biological interpretability and deepen understanding of cross-species cellular relationships. Benchmarking analyses show that TACMAN outperforms current methods for aligning cross-species scRNA-seq datasets under physiological conditions, particularly between evolutionarily distant species such as humans and zebrafish. Its ability to incorporate non-one-to-one orthologs into the model makes TACMAN a strong choice for analyzing scRNA-seq data across distantly related species, thereby enhancing coverage and biological relevance in multi-species integration. Nevertheless, the metapath structure employed includes gene-to-gene edges, and TACMAN’s performance depends on the accuracy and completeness of existing homology annotations, which vary across species. Errors or gaps in gene ortholog databases can cause false alignments or information loss, especially in non-model organisms with poorly annotated genomes.23 In instances where cell types are absent in the reference species, they may be assigned to other closely related cell types and subjected to additional scrutiny through posterior probability calibration or other supplementary analyses. These limitations emphasized the importance of continued improvements in comparative genomics and the need to include robust uncertainty modeling into future versions of TACMAN. Beyond normal conditions, TACMAN has proven useful in various diseases. TACMAN effectively reconstructs the population structure of infiltrated myeloid cells in LC samples and models. Its application in infectious and autoimmune diseases links clinically relevant macrophage subpopulations in humans and animals. In TACMAN-aligned COVID-19 data, differences in TREM2+ macrophages with high interferon signatures between sexes may relate to sex-biased outcomes. TACMAN also identifies conserved fibroblast subpopulations in RA and mouse models that rely on NOTCH3 signaling, indicating potential therapeutic targets. Applying TACMAN shows reliable alignment between C. elegans germline tumor cells and human TGCT tumor cells, despite significant evolutionary differences. It confirms TACMAN’s cross-species effectiveness and opens new avenues for studying conserved transcriptional programs with distant models. Using C. elegans’ transparent and genetically tractable system, featuring tumorigenic mechanisms like aberrant proliferation, dedifferentiation, and meiotic entry,51,53 our integration underscores TACMAN’s potential to discover therapeutic targets by identifying shared molecular vulnerabilities across distant tumor models. Given the dynamic and context-dependent nature of the transcriptome, TACMAN could identify species-specific subpopulations. For instance, a macrophage subset (Mac_IFN_1) did not cluster with the primate model, hinting at species-specific cells. While these differences might reflect immune responses, they do not co-embed and are supported by an independent dataset. Our evaluation shows relatively small concordance in their cross-species transcriptional programs. Still, more analyses are needed to determine if these differences are species-driven or due to experimental artifact. In summary, TACMAN represents progress in cross-species single-cell data integration. While it has some limitations, it provides a strong foundation for upcoming multimodal and translational studies. By enabling more accurate comparisons of scRNA-seq data between humans and animal models, TACMAN can improve experimental planning and speed up the discovery of conserved cellular programs in disease. Limitations of the study The TACMAN integration depends on public gene homology resources, which are incomplete for some non-model taxa and may impact mappings over large evolutionary distances. Our benchmarks focus on specific tissues and cohorts, so testing across more organs, developmental stages, and disease contexts will better show its overall applicability. Functional perturbations were outside the scope but could offer additional validation of conserved states predicted by TACMAN. Resource availability Lead contact Requests for further information, resources, and reagents should be directed to and will be fulfilled by the lead contact, Hang Ruan (hangruan@suda.edu.cn). Materials availability This study did not generate new unique reagents. Data and code availability •Code and test data have been deposited at https://github.com/hr1912/TACMAN.60 Code and test data have been deposited at https://github.com/hr1912/TACMAN.60 •In-house nematode scRNA-seq data have been deposited in the Genome Sequence Archive (GSA: CRA028482). Other datasets are obtained from public data portals (detailed in Table S1). In-house nematode scRNA-seq data have been deposited in the Genome Sequence Archive (GSA: CRA028482). Other datasets are obtained from public data portals (detailed in Table S1). •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request. Acknowledgments This research was supported by the National Natural Science Foundation of China (32100523 to H.R. and 32170690 to Huimin Zhang), the Suzhou Gusu Innovation Leading Talent Program (ZXL2022483 to H.R.), and the Priority Academic Program Development of Jiangsu Higher Education Institutions. The authors thank Dr. Leng Han at the Indiana University School of Medicine for his advice on improving the structure of our manuscript. Author contributions Conceptualization, H.R.; methodology, H.S., H.R., C.L., D.Y., and X.G.; investigation, C.L., H.R., H.S., D.Y., R.F., K.Y., Hongwei Zhang, and Z.H.; writing – original draft, H.R., C.L., H.S., and D.Y.; writing – review & editing, H.R., Z.H., R.F., and Huimin Zhang; funding acquisition, H.R., S.X., and Huimin Zhang; resources, H.R. and Huimin Zhang; supervision, H.R. and S.X. All authors have read and approved the final manuscript. Declaration of interests The authors declare no competing interests. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used Grammarly to improve the manuscript’s readability and clarity. After using this tool or service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERBacterial and virus strainsEscherichia coli OP50Caenorhabditis Genetics CenterRRID:WB-STRAIN:WBStrain00041969Chemicals, peptides, and recombinant proteinsDAPISolarbioCat#C0060Deposited dataRaw reads of scRNA-seq data of N2/GC833This paperGSA:CRA028482Experimental models: Organisms/strainsC. elegans: Strain N2: WildtypeCaenorhabditis Genetics CenterRRID:WB-STRAIN:WBStrain00000001C. elegans: Strain GC833: glp-1(ar202gf)Caenorhabditis Genetics CenterN/ASoftware and algorithmsTACMAN (Python package)This paperhttps://github.com/hr1912/TACMANSeurat (R package)Satija Labhttps://satijalab.org/seurat/Harmony (R package)Korsunsky Lab/Immunogenomicshttps://github.com/immunogenomics/harmonyScanpy (Python package)Theis Labhttps://scanpy.readthedocs.ioSciPy (Python package)SciPy Communityhttps://scipy.orgclusterProfiler (R/Bioconductor package)Bioconductorhttps://bioconductor.org/packages/clusterProfilerMSigDB (Database)Broad Institutehttps://www.gsea-msigdb.org/gsea/msigdbBiopython (Python package)Biopython Projecthttps://biopython.orgggtree (R/Bioconductor package)Bioconductorhttps://bioconductor.org/packages/ggtreeNCBI BLAST+ (Software)NCBIhttps://blast.ncbi.nlm.nih.gov/BioMart (Database)Ensemblhttps://www.ensembl.org/biomart/UniProt (Database)UniProt Consortiumhttps://www.uniprot.orgMobiVision (Software)OE Biotech (Shanghai)N/ASAMap (Python package)Tarashansky Labhttps://github.com/atarashansky/SAMapCAME (Python package)Liu Labhttps://github.com/XingyanLiu/CAMEProtBert model (Python package)Rostlab/HuggingFacehttps://huggingface.co/Rostlab/prot_bert Experimental model and study participant details C. elegans germline tumor model C. elegans hermaphrodites were used and maintained according to standard procedures.61 Worms were grown on Nematode Growth Medium (NGM) agar seeded with Escherichia coli OP50 (source: Caenorhabditis Genetics Center). The wild-type N2 (Bristol) strain and the GC833 strain carrying the glp-1(ar202gf) gain-of-function allele were obtained from the Caenorhabditis Genetics Center (CGC). All strains were incubated at 15°C for maintenance and at 25°C to induce the formation of germline tumors for experiments. Day 1 adults were used for DAPI staining and single-cell RNA-seq. Method details Description of single-cell RNA-seq datasets Datasets were obtained from various data resources and public repositories, as detailed in Table S1. For the HCL and MCA datasets, the processed matrices from different batches were merged using Seurat (V5.0.0) and Harmony (V1.2.0). Marker genes for each cluster were identified using the FindAllMarkers function, and each cluster’s cell types were determined based on their markers (Table S6). Cell types from other reference datasets were provided using labels from corresponding research. The dataset “bronchoalveolar lavage fluid of macaca” (GSE190659) was randomly sampled at 20% per individual to reduce memory usage. We further performed label-granularity harmonization so that cell labels are symmetrically matched between the reference and query. Single-cell RNA-seq data preprocessing for TACMAN Generally, for each scRNA-seq dataset, we denote the dataset with known cell labels as reference data XR∈Rnr×mr and the dataset to be integrated as query data XQ∈Rnq×mq , where mr and mq are the numbers of genes in reference dataXR and query data XQ. nr and nq are the numbers of cells in reference dataXR and query dataXQ (Figure 1A). First, each cell’s read count is normalized by its library size, which is its total count, and then multiplied by a scale factor. The scale factor is calculated as the median of library sizes divided by the read count or normalized expression value. After this, the result is log-transformed with a pseudo-count added for downstream analysis. We then used the SCANPY62 built-in function to identify highly variable genes (HVGs) and differentially expressed genes (DEGs). To identify HVGs, we calculated the dispersion (variance-to-mean ratio) for all genes in each dataset and selected the top 2,000 genes with the highest dispersion from each dataset. These genes were then considered as HVGs for that specific dataset. For DEGs, we conducted a Wilcoxon rank-sum test or Student’s t test separately for the reference and query datasets using the “rank_genes_groups” function from the SCANPY package. In the reference dataset, cells were grouped by their cell-type labels, while in the query dataset, cells were grouped by their Leiden cluster labels. By default, we identify the top 50 DEGs for reference and query datasets. Cross-species cell gene nodes heterogeneous graph The TACMAN’s cell-gene heterogeneous graph G consists of three types of subgraphs (Figure 1B): cell-cell subgraphs Gcc, cell-gene subgraphs Gcg, and gene-gene subgraphs Ggg, featuring two types of nodes and three types of relationships (Figure 1C). Selection of genes for graph nodes and training features To establish the complex relationships between gene expressions in different species for constructing a heterogeneous graph, we use tools such as BioMart63 and BLAST64 to identify homologous genes as connections. Subsequently, we use BLASTP and UniProt to evaluate the similarity between protein-coding genes as weight information. Genes with these connections from reference and query datasets are selected as node genes. These relationships between genes are generally assumed to be one-to-one orthologs by default. At the same time, we also consider non-one-to-one orthologous relationships, especially for species with significant evolutionary distances. A gene must meet two criteria to act as a node on a diverse map and as a cell feature. First, the gene must be expressed (passing a threshold) by the cell in the specific species dataset. Second, the gene must have a homologous gene from another species. We typically choose highly variable or differentially expressed genes as node genes among those that meet the conditions above.65 Construction of the cell-cell subset graph We construct the reference internal graph GccRR for cells in the reference dataset using the KNN approach with PCs = 30 and k = 5. Similarly, we construct the query graph GccQQ for cells in the query dataset. By default, we identify shared DEGs with one-to-one homology between species as cell features.65 We then use the mutual nearest neighbor (MNN) approach to find the closest neighbors between reference and query datasets after simultaneously reducing their dimensionality using PCs = 30 and k = 10. The reference-to-query graph GccRQ is constructed with the MNN approach based on the projected reference and query datasets. For a cell i in the reference data and another cell j in the query data, Gcc(ij)RQ=1 if and only if cell i belongs to the nearest neighbors of cell j, and at the same time cell j is also one of the nearest neighbors of cell i (Figure 1B right panel). Construction of the cell-gene and gene-gene subset graph Gcg(ij) represents an edge between gene j in cell i. In the expression matrix XR(ij), if read count of a gene XR(ij)in a cell is non-zero, it indicates a connecting edge between cell i and gene j Gcg(ij) = 1 (If using other data processing method, a cut-off value can be set) (Figure 1B upper-left panel). Ggg(jj′) indicates a cross-species gene homology relationship between gene j of reference species and gene j′ of query species, where Ggg(jj′)=1 represents a homologous gene relationship, and Ggg(jj′) could be set as the homologous similarity value between genes (Figure 1B down-left panel). Metapath-based neighbor aggregation TACMAN adopts a Simple and Efficient Heterogeneous Neural Network (SeHGNN) approach,30 which is an improved metapath integration method for heterogeneous graphs. The simplified neighbor mean aggregation is executed once during the preprocessing stage and produces a list of feature matrices M={XP:P∈Φ} for all given metapaths Φ. Compared to other heterogeneous graph neural networks, which usually repeat aggregation in each training epoch, metapath neighbor aggregation lowers complexity to linear level in the preprocessing stage. These matrices are used as inputs of the network during later training. In particular, for each node vi in the heterogeneous graph, the mean aggregator is used to combine features from the set of metapath-based neighbors and then create a semantic feature vector ziP for each given metapath P, where mi is a semantic feature vector set of nodes vi, pi, …,j is a metapath instance of metapath P pointing from node j to node i. We use the multiplication of adjacency matrices to calculate the semantic feature matrices. For a particular metapath CGGC (Cell-Gene-Gene-Cell) for node type n, let Xn={x0n;x1n;…;x∥Vc∥−1n}∈R∥Vn∥×dn be the raw feature matrix of all nodes belonging to type c, where each row in this matrix represents the raw feature vector of the corresponding node and dn is its feature dimension. Then the simplified neighbor aggregation can be expressed as where P=CGGC is the metapath, and AˆG,C is the row-normalized form of the adjacent matrixAG,C between node type G and C. In addition, if XP′ has been calculated given the metapath P′=CGGCC, then XP can be calculated with XP′ as the intermediate result, i.e., XP=Aˆn,n1XP′. For instance, in the cell-gene data, we have two pre-defined metapaths: CGGC and CCGGC(Cell-Cell-Gene-Gene-Cell), we can first calculate P=CGGC and then calculate P′=CGGCC as XCCGGC=AˆC,CXCGGC. Transformer-based semantic aggregation Before integrating inter-species and intra-species metapath information using the transformer module,66 we start by projecting different metapath features with feature projection layers, each layer is defined as: where WPil is a semantic specific transformation matrix for each metapath Pi in l layer, HPil is the input (HPi0 = XPi)and HPil+1 is the output. To aggregate various metapath features and incorporate cross-species path semantics, we use a transformer to combine vectors passing through the feature projection layer. The prior feature projection step generates a list of semantic vectors {h′P1,...,h′Pk} for each node for pre-defined metapath list Φ={P1,...,Pk}. The transformer-based semantic aggregator learns attention for each pair of semantic vectors based on the similarity of the key vectors to the query vector. The attention output of this semantic PI is the weighted sum of all value vectors vPj with their corresponding attention values α(Pi,Pj). A residual layer with β learnable weight parameter is then added. The process of semantic aggregation is presented as: with Where WQ, WK, WV are value, query and key projection matrices, WQ, WK, WV, β are trainable parameters shared for all metapaths. Then the final embedding vector (hP1‖hP2‖...‖hP|Φ|) of each node is the concatenation of all those output vectors, and softmax function is used to calculate the probabilities for cell node classification. where Woutput is the project matrix for classifier, y′ is the predicted probabilities over the T cell types for a cell. The dimensions of query and key vectors are 1/4 of the hidden vectors, and the dimensions of value vectors are equal to those of the hidden vectors. The default activation function is ReLU, it can be replaced with leakyReLU or sigmoid. The default dropout rates are 0.5 for activations in the hidden layers and 0.1 for initial input features to handle noise. No dropout is applied to attention weight matrices by default. Loss function The classification loss for cells in the reference dataset is calculated by the weighted cross-entropy loss between the ground-truth label Y and the predicted probabilities Y′, combined with the L2 regularization as: where wt is the class weight for cell type t, and NR is the total cell number of the reference dataset. To prevent the model from being dominated by the major cell populations and ignoring those rare cell types, we set wt∝1Nt Nt is the number of cells of cell type t in the reference dataset. To enhance cross-species cell information aggregation, we defined loss Ls as the mean of the Euclidean distances between the hidden vectors of cell i and the hidden vectors of cell j (Xhidden,i and Xhidden,j): where NMNN is the number of cross-species mutual nearest neighbors. Finally, the overall loss of TACMAN is defined as the combination of a classification loss and a cross-species cell similarity loss with the L2 regularization as below: where β (default value 0.1) is the weight of similarity loss, and λ (default value 0.01) is the penalization coefficient that controls the power of L2 regularization, and θ is the parameter of TACMAN. Multi-stage training In cross-species single-cell annotation, using the reference species dataset alone may lead to insufficient labeled data or overfitting. The multi-stage training method67,68 may serve as a data augmentation to improve the situation. Multi-stage training selects query cell nodes based on prediction probabilities at the end of each training stage and adds these nodes to the training set for the next step. The model parameters are reset, and new training is performed. In each training stage k>1, the entire process of adding training data is presented as: where Lk is the k stage training set and L1 is the initial training set, maxPredi is the maximum prediction probability of cell i in the multiple-classification task, and δk (default 0.8) is the threshold considered confident. By default, TACMAN has three training stages of total 500 epochs. Cell-type integration benchmark We compared the performance of TACMAN to five recent methods: SAMap, CAME, Seurat, TACTiCS, and TOSICA, for cross-species integration. For SAMap, we generated the gene maps file using NCBI-blast (V2.15.0) by running the bash script provided by the SAMap project (https://github.com/atarashansky/SAMap). We then ran SAMap with default parameters. For CAME, we obtained gene homology information from BioMart and ran CAME using its default pipeline. For Seurat, we input the raw data and use the default normalization process through the “NormalizeData” function. We then extracted the top 2,000 highly variable genes using the “FindVariableFeatures” function for reference and query datasets. We further annotated the data using the “FindTransferAnchors” and “TransferData” functions. For TACTiCS, we sourced Swiss-Prot protein sequences from UniProt and created ProtBert embeddings for zebrafish and chicken using a Hugging Face model (https://huggingface.co/Rostlab/prot_bert). Human and mouse embeddings were already included in the TACTiCS package. We calculated protein embedding distances across all species and ran the algorithm using the expression matrix, focusing on one-to-one orthologs. For TOSICA, we utilized Gene Ontology (GO) biological process annotations in GMT format. The GMT files for human and mouse were provided in advance, while the zebrafish file was sourced from GO. We followed the TOSICA demo to process the expression matrix, again using only one-to-one orthologs. The evaluation metric of accuracy is calculated as: while the F1 score and weighted F1 score are calculated as: where TP, TN, FP, and FN represent true positives, true negatives, false positives, and false negatives of integrations, and weighted F1 is considered the proportion of each cell type t in the dataset. Refer to the benchmark folder in the GitHub repository for repeat testing of all six methods for cross-species alignment of human and mouse spleen. The impact of different parameters on model performance In this experiment, we evaluated how varying the number of highly variable genes (HVGs) and differentially expressed genes (DEGs) during preprocessing, as well as adjusting the input and hidden layer dimensions of the TACMAN model, affects its performance. For the impact of gene selection, we tested six combinations of HVG and DEG counts: [1000, 50], [1000, 100], [2000, 50], [2000, 100], [3000, 50], and [3000, 100]. The ranges for HVGs and DEGs were chosen based on commonly used settings in scRNA-seq transcriptomics workflows and applied in prior integration methods such as CAME. By altering one parameter while keeping the other constant, we assessed their individual effects on model accuracy and weighted F1 scores across different species annotations. For the impact of the model architecture, we experimented with input feature sizes and hidden layer dimensions set to [32, 64], [64, 32], [64, 64], [64, 128], and [128, 64] to examine the model’s stability and generalizability. All other parameters remained unchanged during these tests. The combination of parameter selection and performance is detailed in Table S4. Construction of phylogeny based on cross-species embeddings We use Correlation Alignment (CORAL) to align cell embeddings across species into a shared 64-dimensional space.69 We designate the human–macaque feature space as the target domain. A transformation matrix is first computed using the human cell features from two embeddings produced by TACMAN—one that combines human with zebrafish, pig, rat, and mouse, and the other that combines human with macaque. This matrix is then used to project cells from zebrafish, pig, rat, and mouse into the human–macaque feature space, unifying cells from all these species, including macaque and human. Next, for each species, we compute the median value for each feature to obtain a species-wise feature vector, from which we derive a distance matrix. After applying a logarithmic transformation to the distance matrix, we construct a neighbor-joining (NJ) phylogenetic tree using the DistanceTreeConstructor function in Biopython and visualize it with the ggtree R package.70 Gene sets enrichment analysis We use clusterProfiler (V4.10.0) for enrichment analysis. Gene sets and Gene Ontology were obtained from the MSigDB database. IFN response scores were calculated based on the union of “response to interferon-alpha (GO:0035455)” and “response to interferon-beta (GO:0035456)” with tl.score_genes of scanpy. Human and mouse IFN gene sets and their orthologous relationship are provided in Table S7. Conservation checklist evaluation framework To evaluate each highlighted cross-species cell subset, four complementary metrics are used: (1) neighborhood purity, (2) posterior probability, (3) DEG concordance, and (4) pathway concordance. These metrics help distinguish subsets with strong evidence of conservation from those with support for alignment but limited biological evidence, rather than simply treating co-embedding as evidence of conservation. Neighborhood purity reflects confidence in alignment after TACMAN integration, defined here as the fraction of cells among the k-nearest neighbors (default k = 30) in the query-species manifold that belong to the mapped counterpart population. Posterior probability offers an additional measure of confidence for label transfer. Further assessments of the concordance between differentially expressed orthologs and pathways are used to evaluate biological conservation more directly, although they are less reliable indicators for distant species with fewer shared orthologs. DEG concordance captures the cross-species consistency of differential expression patterns across aligned subsets, while pathway concordance evaluates whether the aligned populations share higher-order functional programs based on cell annotations, marker expression, and pathway enrichment derived from differential expression analysis. Worm preparation Synchronized L1 worms were obtained by bleaching and seeded onto 90 mm NGM plates with OP50 to grow for 66 h into adulthood at 25°C. Worms were washed off the plates using S buffer and were allowed to settle in 15 mL tubes. When most adults settled to the bottom of tubes, the supernatant containing the larvae and eggs was discarded. This wash was repeated five times to isolate the adult worms. About 300 adult worms per strain were fixed in 95% ethanol for DAPI staining. The remaining adult worms (approximately 10,000 per strain) were collected for scRNA-seq. DAPI staining Worms were fixed in 95% ethanol for 10 min at room temperature. Ethanol was then removed and 20 μL of DAPI was added at a concentration of 1 μg/mL. Worms were transferred to the agarose pad and covered for imaging using LEICA DM2500 upright epifluorescence microscope. ScRNA-seq of C. elegans samples The MobiCube High-throughput Single Cell 3′ Transcriptome Set V2.1 (PN-S050200301) and MobiNova-100 microfluidic platform were used for scRNA-seq. The single-cell suspension was adjusted to an appropriate concentration (700–1200 cells/μL) and loaded immediately onto a chip to run on the MobiNova-100 for micro-droplet formation. 11,012 cells from N2 and 11,031 cells from GC833 were yielded for scRNA-seq. The FASTQ files were processed and aligned to the C. elegans reference genome (GCF_000002985.6_WBcel235) using MobiVision software (version 3.2) with unique molecular identifier (UMI) counts summarized for each barcode. The sequencing and raw data processing bioinformatics analysis were conducted by OE Biotech Co., Ltd. (Shanghai, China). Cells in the C. elegans dataset were filtered with standard QC, keeping those with at least 200 genes and 500 reads, and less than 20% mitochondrial gene expression. Doublets were identified and removed using DoubletFinder. Human and C. elegans homologous genes used in ablation experiments are in Table S7. Quantification and statistical analysis We performed all analyses using SciPy (V1.12.0) with statistical significance set at p < 0.05 adjusted for multiple testing. The Mann-Whitney U rank test was used to evaluate associations with continuous variables, and the Spearman rank-order correlation test was used to evaluate the correlation between two continuous variables. The chi-square test was also used to analyze the cell count contingency table of status and cell subtypes. All multiple hypothesis tests were corrected using the Benjamini-Hochberg (BH) method. Published: June 3, 2026 Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.crmeth.2026.101469. Contributor Information Huimin Zhang, Email: zhanghuimin@suda.edu.cn. Sidong Xiong, Email: sdxiong@suda.edu.cn. Hang Ruan, Email: hangruan@suda.edu.cn. Supplemental information Document S1. Figures S1–S7 Table S1. Metadata, related to STAR Methods Table S2. Benchmarks of six methods for integrating scRNA-seq data, related to Figure 2 and STAR Methods Table S3. Summary of gene counts by category across benchmark datasets, related to Figure 3 and STAR Methods Table S4. Impact of input gene selection and hyperparameters on TACMAN performance, related to STAR Methods Table S5. Conservation checklist in four disease datasets, related to STAR Methods Table S6. Cell markers for cell type annotation, related to Figures 2, 3, 4, 5, and 6 and STAR Methods Table S7. Orthologous relationship for mapping specific functional pathways, related to Figures 6 and 7 and STAR Methods Document S2. Article plus supplemental information References Associated Data Supplementary Materials Document S1. Figures S1–S7 Table S1. Metadata, related to STAR Methods Table S2. Benchmarks of six methods for integrating scRNA-seq data, related to Figure 2 and STAR Methods Table S3. Summary of gene counts by category across benchmark datasets, related to Figure 3 and STAR Methods Table S4. Impact of input gene selection and hyperparameters on TACMAN performance, related to STAR Methods Table S5. Conservation checklist in four disease datasets, related to STAR Methods Table S6. Cell markers for cell type annotation, related to Figures 2, 3, 4, 5, and 6 and STAR Methods Table S7. Orthologous relationship for mapping specific functional pathways, related to Figures 6 and 7 and STAR Methods Document S2. Article plus supplemental information Data Availability Statement •Code and test data have been deposited at https://github.com/hr1912/TACMAN.60 Code and test data have been deposited at https://github.com/hr1912/TACMAN.60 •In-house nematode scRNA-seq data have been deposited in the Genome Sequence Archive (GSA: CRA028482). Other datasets are obtained from public data portals (detailed in Table S1). In-house nematode scRNA-seq data have been deposited in the Genome Sequence Archive (GSA: CRA028482). Other datasets are obtained from public data portals (detailed in Table S1). •Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this work paper is available from the lead contact upon request.