Unify learns cellular evolution with universal multimodal embeddings Authors: Huawen Zhong, Wenkai Han, Guoxin Cui, David Gomez-Cabrero, Jesper Tegner, Xin Gao, Manuel Aranda Published: 2026 ABSTRACT AbstractIntegrating single-cell RNA-sequencing (scRNA-seq) data across species is hindered by evolutionary divergence, technical batch effects, and the reliance on one-to-one orthologs. Here, we present Unify, a transfer learning methodology that learns universal cell embeddings by defining functionally coherent, multi-modal macrogenes. This is achieved by combining RNA expression with embeddings from protein language models and general-purpose language models. Unify transcends species boundaries, enabling cross-species comparisons beyond strict gene-level homology. Unify corrects batch effects while preserving conserved biological signals across vast evolutionary distances and enables more accurate prediction of perturbation responses across species, such as from mouse to human. Applied to species separated by over 700 million years, Unify reconstructs more accurate multi-species cell-type evolutionary trees and uncovers convergent gene programs. Together, these results establish Unify as a powerful method for comparative single-cell genomics and evolutionary biology.Subject terms: Machine learning, Data integration, Evolutionary biology, Gene expression analysis Introduction The advent of single-cell RNA-sequencing (scRNA-seq) has revolutionized our ability to map the landscape of cellular identity and function within individual organisms1–3. An exciting and critical new frontier is comparative single-cell genomics, which seeks to extend these insights across the tree of life. By comparing cell atlases from different species, we can uncover the fundamental principles of gene regulation that are conserved through evolution, trace the origins and diversification of cell types, and, critically, enhance the translation of biological findings from model organisms to human health3–6. However, to achieve meaningful cross-species integration of single-cell data is a significant challenge. The task is profoundly challenging due to the need to distinguish conserved biological signals from the confounding effects of technical variability between experiments and, more significantly, the deep evolutionary divergence that has reshaped genes and genomes over millions of years7–9. There is a growing interest in cross-species data integration, and several methods have been developed to address this challenge6,10–13. Early computational efforts relied heavily on one-to-one orthologs as anchors to map transcriptomes across species14–20. More recently, methods like SAMap6 and CAME13 have advanced this by utilizing BLAST-based homology to construct cross-species graphs. Furthermore, a newer generation of approaches, including SATURN10, UCE11, and TranscriptFormer12, leverage embeddings from protein language models like ESM-221 to capture functional relationships between both orthologous and non-orthologous genes. However, their reliance on a single modality, i.e., protein sequence, provides an incomplete picture of gene function, neglecting the rich, contextual information captured in curated, expert-annotated biological knowledge bases. While text-based foundation models, such as GPT-422 and Gemini-2.523, have shown promise in encoding this knowledge for gene annotation, they have not yet been integrated into cross-species cell embedding frameworks. Furthermore, a difficult trade-off exists for current methods. Aggressive batch correction often comes at the cost of losing subtle but critical biological information across species7–9. Additionally, existing integration approaches typically produce either corrected expression profiles restricted to one-to-one orthologous genes14–17 or integrated embeddings6,10,18–20 that lack direct interpretability. Currently, no method excels at both robust batch correction and faithful preservation of conserved biology, thereby limiting sensitive downstream applications, such as cross-species perturbation prediction or the reconstruction of cell-type phylogenies8,9. Here, we introduce Unify (Universal Non-orthologous Integration via Functional Yield), a transfer learning framework designed to resolve the core challenges of cross-species single-cell analysis. Unify builds on the concept of functionally coherent macrogenes that transcend species boundaries, extending it with a multimodal representation. While conceptually related to approaches using protein embeddings (e.g., SATURN), Unify integrates structural information from protein language models with functional context derived from general-purpose large language models interpreting biological knowledge. To achieve this, it constructs a multimodal gene representation that integrates RNA expression with two complementary data types: structural information from protein language models and functional context derived from general-purpose large language models that interpret biological knowledge. By learning a shared cell embedding space based on these macrogenes, Unify operates beyond the limitations of strict orthology. Its adversarial architecture, guided by a Pareto-optimal multi-task learning strategy, is explicitly designed to resolve the critical trade-off in comparative genomics: achieving robust batch-effect correction while simultaneously preserving the conserved biological signals essential for evolutionary discovery. We systematically benchmarked Unify against state-of-the-art methods and demonstrated its superior ability to balance batch correction and biological conservation, even across vast evolutionary distances of over 700 million years. We show that this integrated representation enables a range of powerful downstream applications. For instance, Unify uncovers functionally convergent gene programs that reflect critical evolutionary cellular functions, significantly improving the accuracy of predicting cellular perturbation responses between species, such as from mouse to human. By applying this framework to diverse organisms, we show that it can construct more accurate multi-species cell-type evolutionary trees, offering a powerful lens for comparative single-cell genomics. Overall, Unify provides a versatile and robust solution for exploring cellular function and evolutionary dynamics across the tree of life. Results Overview of Unify To overcome the dual challenges of evolutionary divergence and technical batch effects in cross-species analysis, we developed Unify, a transfer learning framework that integrates scRNA-seq data through a multi-modal gene representation (Fig. 1 top panel). At its core, Unify moves beyond sequence homology by defining macrogenes: functionally coherent gene modules that are consistent across species. To build these, Unify constructs a rich, multi-modal feature space for every gene by integrating two complementary sources of information: (1) structural properties, captured by embeddings from a pre-trained protein language model (ESM-221), and (2) functional context, derived from a general-purpose large language model (LLaMA-224) interpreting curated biological knowledge. Genes with similar structural or functional embeddings are grouped into macrogenes, creating an interpretable, functional unit for integration. Unify then transfers this pre-existing, universal knowledge and applies it to the specific task of aligning cells across species. The Unify framework for balanced cross-species integration. The Unify workflow consists of two main stages. (Top Panel) First, Unify constructs multi-modal macrogenes by integrating gene representations from two complementary sources: (1) structural properties, derived from protein language models (e.g., ESM-2), and (2) functional context, derived from large language models (e.g., LLaMA-2). (Bottom Panel) These macrogenes then serve as input for an adversarial autoencoder. A Pareto-optimal multi-task learning strategy guides the training process to systematically balance three competing objectives: robust batch correction (mixing species), the preservation of conserved biological signals (aligning cell types), and the reconstruction of input data (capturing the cellular state). The final output is a unified cell embedding space that enables accurate and robust cross-species comparisons. Before integration, Unify projects the scRNA-seq expression profiles into a macrogene expression space. This is achieved by calculating each macrogene’s expression as a learnable, weighted combination of its constituent genes (see Methods). This resulting macrogene-by-cell matrix then serves as the input for an adversarial autoencoder, which learns the final unified latent space. This architecture is explicitly trained to disentangle biology from technology by optimizing three competing objectives simultaneously. A reconstruction loss ensures the latent space retains cellular information, while two classifiers drive the core integration: a cell-type classifier preserves conserved biological signals (bio-conservation), and a species adversarial discriminator removes technical, species-specific artifacts (batch correction). Another key innovation in Unify is its use of Pareto-optimal multi-task learning to navigate the inherent trade-off between these objectives (Fig. 1 bottom panel). While Pareto optimality often defines a final set of non-comparable solutions, our framework employs this principle dynamically to guide the training of a single model. Our approach, based on the Multi-Gradient Descent Algorithm-Upper Bound (MGDA-UB) algorithm25, forgoes fixed manual weighting. Instead, at each training step, it solves for a single gradient update direction that represents the optimal compromise—a direction that improves all objectives when possible, or provides the most balanced descent when they conflict. By iteratively applying these locally optimal updates, the model follows a single, Pareto-guided training trajectory. This process yields one robust, final model, eliminating the need for any manual selection from a set of possible solutions. This principled optimization ensures that Unify achieves robust and balanced integration. The resulting framework generates unified cell embeddings and supports conditional reconstruction of gene expression, providing representations that can be used for large-scale data integration and gene-level downstream analyses. To validate its performance, we systematically benchmarked Unify against state-of-the-art methods. Unify robustly balances batch correction and biological conservation across vast evolutionary distances To rigorously evaluate its performance, Unify was benchmarked against a comprehensive suite of competing methods. This included the current state-of-the-art in many-to-many ortholog integration, SATURN10 and SAMap6, alongside seven widely used methods14–20 restricted to one-to-one orthologs. To ensure a fair and transparent comparison, we explicitly distinguish between “label-informed” methods, which utilize cell-type annotations to guide model fitting (e.g., Unify, SATURN10, scGen26), and fully “unsupervised” methods (e.g., SAMap6, scVI18) (Supplementary Data 1). Our evaluation framework was designed to assess the two fundamental, and often competing, goals of integration: the removal of species-specific technical artifacts (batch correction) and the preservation of conserved cell-type identity (bio-conservation). Using a comprehensive suite of metrics for each aspect (Methods), we computed an overall score (40% batch correction, 60% bio-conservation) to provide a single, robust measure of integration quality, based on previous benchmark studies7–9. This weighting acknowledges that while removing technical artifacts is essential, preserving shared biological structure remains the paramount objective of cross-species comparative analysis. Across the tested phylogenetic distances, from closely related species to those separated by hundreds of millions of years, Unify achieved the highest overall score (Fig. 2a and Supplementary Data 2 and 3). The results indicate that Unify maintained a favorable balance between batch correction and biological conservation in these benchmarks, as reflected by the integration metrics shown in Fig. 2b, c and Supplementary Figs. 1–16. Crucially, while the performance of competing methods degraded significantly with increasing evolutionary distance, Unify’s performance remained stable, highlighting its robustness for challenging comparative genomics tasks. Interestingly, Unify exhibited comparatively lower trajectory conservation scores in integrations of closely related species, highlighting a trade-off inherent to supervised cross-species alignment. Although cell-type labels are not enforced as strict matching constraints, the classification objective still promotes compact, label-consistent organization in the latent space. This favors accurate cross-species alignment and strong cluster-level biological conservation, but may compress transitional intermediate states and thereby reduce trajectory conservation. In these settings, Unify therefore favors robust alignment of annotated cell identities over strict preservation of fine-scale developmental continuity. To assess the statistical robustness of the benchmark results, we performed pairwise comparisons across integration tasks. Unify achieved significantly higher overall scores than baseline methods (Supplementary Figs. 17 and 18a). This result remained robust across alternative weighting schemes, with Unify retaining the highest average overall score and remaining significantly better than most competing methods (Supplementary Fig. 18b, f). Overall performance of Unify in different cross-species integration tasks. a Overall performance of all the benchmarked methods across different categories of cross-species integration tasks (n = 10 tasks) based on the phylogenetic distance between the species. Label-informed indicates methods that use cell type information during training, whereas unsupervised indicates methods that do not use cell type information during training. b Average bio-conservation and batch correction scores with standard error of all methods in all the integration category tasks (n = 10 tasks). The vertical dashed line represents the average batch correction score for all methods. The horizontal dashed line represents the average bio-conservation score for all methods. c Integration performance of each method for different phylogenetic distance species integration tasks. d, e UMAP plot of the selected top four methods’ integration results of integrating Felis catus and Panthera tigris altaica (cross-genus species integration task). Source data are provided as a Source data file. This performance pattern is consistent with the characteristic failure modes of the competing methods and helps explain the advantages of Unify’s design. Ortholog-reliant methods like scGen predictably suffered from poor bio-conservation as the number of shared orthologs decreased significantly between distant species. Conversely, methods that incorporate all genes faced the opposite challenge. SATURN, for example, used protein embeddings to maintain strong bio-conservation but exhibited poor batch correction, failing to merge cell populations effectively at the local level (e.g., kBET, graph iLISI scores; Fig. 2d, e and Supplementary Figs. 1–10). SAMap occupied a middle ground, but its performance was inconsistent; its reliance on mutual nearest-neighbor stitching, a process sensitive to expression levels and sequencing depth9, resulted in suboptimal bio-conservation, particularly between closely related species. Beyond its strong integration performance, Unify also showed favorable computational scalability across datasets ranging from 70 K to 1.1 M cells. It scaled more favorably than several competing cross-species integration approaches, including SATURN, SAMap, Seurat v4 CCA, and BBKNN, particularly at larger dataset sizes (Supplementary Fig. 19 and Supplementary Data 4). This indicates that Unify combines strong integration accuracy with practical scalability for large comparative atlases. Ablation studies validate the key design choices of Unify To assess the individual contributions of Unify’s core design components to overall integration performance, we conducted a comprehensive series of ablation studies evaluating the multimodal representation, optimization strategy, foundation-model choice, cell-type supervision, and input gene space (Supplementary Figs. 20–24 and Supplementary Data 5–9). Across ten integration tasks spanning cross-genus to cross-phylum comparisons, the full multimodal model, initialized with both ESM-derived protein embeddings and LLaMA-derived functional text embeddings, outperformed the single-modality variants. The full model achieved significantly higher bio-conservation and overall scores (Supplementary Fig. 20 and Supplementary Data 5). This indicates that structural and functional priors provide complementary information for macrogene construction, enabling the latent space to better preserve shared cellular programs across vast evolutionary distances. The choice of optimization strategy proved equally critical to model performance. Training with MGDA-UB yielded significantly higher bio-conservation and overall scores compared to fixed loss weighting (Supplementary Fig. 21 and Supplementary Data 6). These findings suggest that the inherent trade-off between species mixing and preservation of cell identity is better resolved through Pareto-optimal multi-objective optimization than through manually specified, static weights. While performance remained broadly robust across different foundation-model combinations, distinct trade-offs between protein and text encoders were observed. Multiple pairings of protein language models (ESM27, ProstT528, and ProtT529) and text models (LLaMA24, BLOOM30, and PubMedBERT31) showed competitive performance, demonstrating that Unify is not dependent on a specific embedding source (Supplementary Fig. 22 and Supplementary Data 7). We assessed the contribution of biological supervision by testing a semi-supervised variant, where unsupervised cluster identities replaced one species’ cell-type labels, and a fully unsupervised variant lacking the cell-type classifier entirely (Supplementary Fig. 23 and Supplementary Data 8). Unify (semi-supervised) retained strong bio-conservation performance, performing comparably to SAMap and becoming increasingly competitive as phylogenetic distance increased. These results indicate that while explicit cross-species supervision maximizes alignment, it is not the sole source of biological structure. These results suggest that the multimodal macrogene priors capture substantial conserved biological signals, allowing Unify to remain highly competitive even when labels are missing or unaligned. Finally, restricting the input to one-to-one orthologs (Unify (one-to-one)) reduced performance relative to the full model using all genes, but Unify (one-to-one) still significantly outperformed most baseline methods (Supplementary Fig. 24 and Supplementary Data 9). This underscores that the Unify framework remains effective and robust even under a restricted gene space. These results offer direct support for Unify’s core architectural choices and help explain its strong performance relative to other methods. By constructing multi-modal macrogenes that integrate large language model based functional annotations, Unify appears to build a more robust “functional anchor”, which likely contributes to its ability to maintain high bio-conservation. Similarly, the performance trade-offs observed across most competing methods highlight the inherent difficulty of balancing competing losses with heuristic or fixed-weight approaches. Unify’s use of Pareto-optimal adversarial training offers a more systematic approach to this problem, aiming to find a principled balance rather than a static compromise. The combination of this richer gene representation and a more principled optimization strategy likely explains Unify’s capacity to achieve a better balance between batch correction and biological discovery in cross-species integration. Unify enables multiple-species integration and meaningful cross-species comparison of differentially expressed macrogenes A central goal of comparative genomics is to identify gene programs that define cell identity and function across species. Unify moves beyond simple data integration to become an engine for this discovery through differential macrogene analysis. A key advantage of Unify is that the multi-modal macrogene approach enables the decoupling of structural and functional gene relationships, revealing evolutionary dynamics. By comparing the similarity scores from protein embeddings (structure) and text embeddings (function), we can directly identify cases of evolutionary divergence and convergence that are often missed by single-modality analysis. This allows us to systematically categorize gene pairs into groups such as: (1) conserved (high structural and functional similarity), (2) functionally divergent (high structure, low function), and (3) functionally convergent (low structure, high function). Finally, Unify overcomes the residual batch effects that confound downstream comparisons in other methods (e.g., Supplementary Fig. 25, which shows incomplete species mixing by SATURN). To systematically assess whether Unify captures functional divergence and convergence, an annotation-supported human-mouse benchmark was constructed. Functionally divergent pairs were defined as reciprocal best-hit (RBH) orthologs, in which each gene is the best cross-species match of the other in BLAST32, with low GO Lin semantic similarity33, a measure of functional similarity derived from Gene Ontology (GO) structure. Functionally convergent pairs were defined as non-homologous gene pairs with high GO Lin semantic similarity (Methods). Under these definitions, Unify macrogenes were significantly enriched for both divergent and convergent gene pairs relative to permutation-based null models, indicating that the learned macrogenes capture biologically meaningful complex evolutionary relationships rather than random gene co-assignment. Consistent with this, area under the receiver operating characteristic curve (AUROC) analysis showed that neither structural similarity nor functional text similarity alone was sufficient to recover both patterns, supporting the value of Unify’s multimodal design (Supplementary Fig. 26). To showcase this capability, we integrated a five-species atlas of the eye’s aqueous humor outflow structures (AH Atlas) (Fig. 3b). We then performed differential analysis at the macrogene level to identify key programs defining macrophage identity. This analysis successfully pinpointed macrogenes enriched with well-established macrophage markers along with their homologous genes (e.g., C1QA34–36, C1QB35,36, C1QC35,36, CD7437, CCL336,38/436,39), validating that our approach captures cross-species conserved biology (Fig. 3c and Supplementary Table 1). This finding is further substantiated by GO and Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis, which show that the component genes of these macrogenes are highly enriched in pathways characteristic of macrophage function (Supplementary Figs. 27–31). For instance, macrogene 857, which was highly expressed in macrophages, contained genes significantly enriched for “antigen processing and presentation,” a core macrophage function40 (Fig. 3d). Unify identifies conserved, divergent, and convergent macrogenes in the Cell Atlas of Human Trabecular Meshwork and Aqueous Outflow Structures (AH atlas). a Schematic of Unify’s macrogene analysis. b UMAP plot of the macrogene space and the basal space of the AH atlas. c Dotplot of the expression levels of the differentially expressed macrogenes in macrophages. d Top enriched KEGG pathways of the component genes of differentially expressed macrogene 857 across species in macrophage. KEGG enrichment was performed separately for each species. Significance was assessed using a one-sided hypergeometric test for pathway over-representation, with Benjamini–Hochberg adjustment for multiple comparisons. Pathways with BH-adjusted P < 0.05 were retained, and the ten pathways with the lowest adjusted P-values are shown for each species. Dot size represents the gene ratio, and dot color represents −log10(BH-adjusted P-value). e Comparison of ESM and LLaMA similarity scores for gene pairs within the component genes of macrogene 838, with each score type max–min scaled separately. f Comparison of ESM and LLaMA similarity scores for gene pairs within the component genes of the differentially expressed macrogene 2077 across species of macrophages, with each score type max–min scaled separately. hs: human (Homo sapiens); ss: pig (Sus scrofa); mu: monkey (Macaca mulatta); mm (Mus musculus). Source data are provided as a Source data file. A clear example of functional divergence was found in macrogene 838. This analysis revealed high structural similarity but low functional similarity between paralogous genes from the C1QTNF family, such as human C1QTNF1 and human C1QTNF6. This pattern persists across species (e.g., human C1QTNF1 vs pig C1QTNF6), where the paralogs share a conserved protein domain but have different biological roles—one pro-inflammatory and immune-related41, the other anti-inflammatory and metabolic42 (Fig. 3e and Supplementary Figs. 32–34). This demonstrates Unify’s ability to avoid confounding functionally distinct paralogs, a critical flaw in SATURN and SAMap, which rely on sequence similarity alone. Conversely, we found compelling evidence of functional convergence. Macrogene 2077 contained genes like human HP and mouse H2-Ab1 that share relatively very low sequence similarity but have highly similar functional embeddings, linking them to share roles in defense and antioxidant activity (Fig. 3f and Supplementary Data 10 for the enriched GO term description of these genes). This demonstrates Unify’s ability to identify functionally analogous genes that have evolved independently—a critical blind spot for all ortholog-based methods. We further confirmed these findings by integrating a human and mouse immune dataset. In neutrophils, Unify successfully identified macrogenes containing neutrophil marker genes along with their orthologous genes (such as CAMP/Camp43, S100A8/S100a844,45, S100A9/S100a944,45, Supplementary Fig. 35a, b). Moreover, Unify uncovered functionally coherent gene programs without direct orthologs. Macrogene 2225, for instance, contained no homologous gene pairs. However, GO enrichment analysis revealed that its component genes in both species converge on core neutrophil functions, including phosphatase activity and host-mediated cell killing46 (Supplementary Fig. 35c). We further identified functionally convergent pairs, such as human PRNP and mouse Calm3. Despite sequence dissimilarity, both are implicated in calcium signaling pathways essential for neutrophil function47, highlighting functionally conserved programs that would otherwise be invisible, which is further illustrated by word clouds of the biological description (Supplementary Fig. 36). Collectively, these results show that Unify’s differential macrogene analysis provides a powerful framework for dissecting the complex evolutionary trajectories of gene function, uncovering both conserved programs and unexpected patterns of divergence and convergence that shape cellular identity across the tree of life. Unify accurately predicts cross-species perturbation effects A critical application for cross-species integration is in translational research, where predicting a human cellular response from model organism data can accelerate drug discovery and improve our understanding of disease. However, this task is severely hampered by methods that rely on one-to-one orthologs, which discard a vast amount of genetic information. We hypothesized that Unify’s ability to learn a unified embedding space using all genes would enable far more accurate cross-species perturbation prediction. We then evaluated the accuracy of this prediction against the actual, experimentally measured data in three key ways: (1) overall transcriptome similarity (R²): we calculated the R² score between the predicted and actual gene expression values across all genes. A high R² means the overall shape of our predicted transcriptome closely matches the real one. (2) identification of key genes (differentially expressed genes (DEGs) hit rate): of the genes that truly changed expression in the target species, we measured the percentage that our model correctly identified. (3) correctness of response (directional accuracy): among all genes predicted as DEGs, we measured the proportion that were not only true differentially expressed genes but also had the correctly predicted direction of change (up-regulation or down-regulation). To test this, we designed a highly challenging translational scenario: predicting the transcriptomic response of human PBMCs to IFN-β stimulation based on data from stimulated mouse lymph nodes—a setup involving different species, tissues, and experimental conditions (Fig. 4a, Methods). This prediction was performed in a zero-shot setting, where the target species perturbation condition was entirely excluded during training and inferred solely from perturbation representations learned from the source species. When restricted to predicting changes based on one-to-one orthologs alone, Unify variants (Unify-basal and Unify-macrogene) performed comparably to scPARM48,49, a state-of-the-art method specifically developed for perturbation effect prediction (Fig. 4b and Supplementary Data 11). However, the advantage of Unify framework becomes more evident when the predictions were expanded to the full gene set. By predicting across all genes, which capture the most dynamic aspects of the cellular response, both Unify variants achieved significantly higher accuracy (R²) in predicting the final expression state compared to all baseline methods and showed strong performance in DEG hit rate and directional accuracy (Fig. 4b–e). The robustness of this framework was further assessed by performing additional zero-shot cross-species predictions, using data from lipopolysaccharide (LPS) stimulated mouse macrophages to predict the responses in rat, rabbit, and pig macrophages. Across these transfers, both Unify variants again achieved significantly higher R2 in the all-gene setting and remained highly competitive with scPARM in DEGs hit rate and directional accuracy (Fig. 4f–h and Supplementary Data 12). Unify accurately predicts cross-species perturbation effects. a Schematic of the two proposed approaches for predicting perturbation effects across species using the Unify framework. b Average R2 scores (with standard error) for predicted human IFN-β perturbation expression data across nine cell types (n = 9 tasks). c Average hitting rate (with standard error) for differentially expressed genes predicted human in the IFN-β stimulated dataset (n = 9 tasks). d Average direction accuracy (with standard error) for differentially expressed genes predicted human in the IFN-β stimulated dataset (n = 9 tasks). e Average hitting rate and direction accuracy (with standard error) of predicted differentially expressed genes by baseline methods, evaluated across different top DEGs categories in one-to-one orthologous and all-genes scenarios (n = 9 tasks). f Average R2 (with standard error) for predicted LPS perturbation expression data of rat, rabbit, and pig macrophages based on mouse data (n = 6 tasks). g Average hitting rate (with standard error) for differentially expressed genes predicted in the LPS-stimulated rat, rabbit and pig datasets (n = 6 tasks). h Average hitting rate and direction accuracy (with standard error) for differentially expressed genes predicted in the LPS-stimulated rat, rabbit and pig datasets (n = 6 tasks). Statistical significance between methods was assessed using paired two-sided t-tests. P-values with Benjamini–Hochberg correction are shown (*** adjusted P < 0.001, ** adjusted P < 0.01, * adjusted P < 0.05). Source data are provided as a Source data file. Direct comparison of the two Unify variants revealed distinct strengths. In the more challenging IFN-β task, Unify-macrogene achieved significantly higher R² than Unify-basal in both the one-to-one and all-gene settings, whereas Unify-basal showed a modest advantage in DEG hit rate. This pattern suggests that the macrogene framework tends to better capture broad transcriptional programs, whereas the basal variant is more sensitive to gene-specific changes that cross differential-expression thresholds. In the less complex LPS tasks, where perturbation signals were stronger and more conserved, no significant differences were observed between the two variants. Together, these results show that Unify can be applied to zero-shot cross-species perturbation prediction beyond one-to-one ortholog-restricted comparisons. By incorporating all genes through sequence- and function-informed representations, the framework improved transcriptome-level prediction accuracy in the all-gene setting and supported the comparison of perturbation responses across model organisms and human cells. Unify enables higher resolution cell type tree construction A major goal of cross-species integration methods is to reconstruct the evolutionary history of cell types across vast phylogenetic distances. A robust cell-type tree of life not only reveals the origins and diversification of cellular functions but also serves as a foundational step toward building comprehensive, predictive virtual cells. To demonstrate Unify’s power in this domain, we applied it to one of the most challenging integration tasks: constructing a cell-type tree of life from seven species spanning over 700 million years of evolution, including vertebrates, tunicates, insects, nematodes, and flatworms (Fig. 5a). To evaluate the quality of this tree, we used two key metrics: (1) technical repeatability (jumble score): this is a consistency check that measures whether the tree topology is reproducible when the starting position of the maximum likelihood search is varied, ensuring it’s not a random artifact. (2) biological repeatability (scjackknife score): this is a more stringent test of robustness. It repeatedly re-builds the tree on subsets of the data to see if the core biological relationships (e.g., all vertebrate neuron types clustering together) remain stable. A high scjackknife score indicates a tree that reflects a true, reproducible biological signal50. Unify enables better cell type tree construction. a UMAP plot of the integrated single-cell datasets from seven species across the tree of life (Schmidtea mediterranea, Danio rerio, Ciona intestinalis, Mus musculus, Homo sapiens, Drosophila melanogaster, and Caenorhabditis elegans). b Unrooted cell type tree for seven phylogenetically distant species constructed from the integrated basal embedding generated by Unify. Species and cell types are labeled at the tips. Node support values are printed as “jumble score/scjackknife score”. MYA: million years ago. Boxplot shows the distribution of the scjackknife score/jumble score for each node in the cell type tree calculated from Unify and SATURN (n = 45 species cell types). The center line denotes the median, the lower and upper bounds of the box denote the 25th and 75th percentiles, respectively, and whiskers denote the minimum and maximum values. The P-value was calculated using a two-sided Mann–Whitney U test. Source data are provided as a Source data file. We compared the cell-type tree generated by Unify to that of SATURN, evaluating their technical and biological repeatability. While both methods showed similar technical consistency, Unify’s tree demonstrated greater biological coherence. It achieved a significantly higher biological repeatability score, indicating that the relationships it uncovered were more consistent with known biology across independent samples (Fig. 5b and Supplementary Fig. 37). When normalized for technical performance, Unify captured substantially more meaningful biological information, demonstrating its superior ability to distill true evolutionary signals from complex, multi-species data. In addition to recovering expected broad evolutionary relationships, Unify also revealed finer-grained higher-order connections among conserved cell lineages that were less clearly resolved by SATURN. For example, neuronal populations from multiple species formed a tighter shared clade in the Unify-derived tree, consistent with the deep conservation of neuronal regulatory programs across metazoans3. Similarly, epithelial cell types from vertebrates and invertebrates formed a coherent branch whose internal ordering broadly reflected known phylogenetic relationships, with mammals (Homo sapiens and Mus musculus) clustering as sister groups, followed by teleost fish (Danio rerio), and then invertebrates including Drosophila melanogaster and Caenorhabditis elegans. This pattern is consistent with the possibility that core epithelial regulatory programs were established early in animal evolution and have remained relatively stable despite extensive lineage diversification51,52. By successfully organizing 125 distinct cell types from deeply divergent phyla into a biologically plausible evolutionary structure, Unify demonstrates the advantages of its design. Its multi-modal macrogenes and Pareto-optimized integration approach provide the robustness needed to bridge large evolutionary distances, moving beyond simple data alignment to deliver genuine biological insight. This high-resolution reconstruction of cell-type evolution provides a powerful resource for understanding the fundamental principles that govern cellular identity across the animal kingdom. Discussion Unify enhances cross-species single-cell analysis by tackling its two most fundamental challenges: the limitations of sequence homology and the trade-off between batch correction and biological conservation. By introducing multi-modal macrogenes that integrate structural information from protein models with functional context from large language models, Unify learns a universal cellular language that captures both functionally conserved and convergent evolution. Crucially, we pair this with a Pareto-optimal multi-task learning strategy that systematically finds the optimal balance between removing technical noise and preserving true biological signals—a principled solution to a problem that has long plagued the field. Although Unify leverages cell-type annotations during training, these labels are used as soft biological priors rather than rigid alignment constraints. This is particularly important in cross-species settings, where annotations, especially in non-model organisms, often represent approximate rather than strict one-to-one correspondences. In practice, such labels are frequently inferred from canonical marker genes or expression similarity within a limited ortholog space, and therefore may not define fully equivalent cellular states across species. Accordingly, the main value of Unify lies not in de novo cell-type discovery, but in enabling comparative analysis in a shared embedding space that goes beyond discrete label assignments. Rather than forcing identically labeled populations into exact overlap, Unify supports the analysis of comparative relationships among partially overlapping cell states, thereby providing a basis for downstream analyses such as cell-type tree construction, evolutionary inference, perturbation prediction, and the identification of conserved and species-specific regulatory programs. A similar principle of balance applies to Unify’s dual representation spaces. The distinction between Unify-basal and Unify-macrogene in the perturbation prediction task highlights the complementary value of gene-level sensitivity versus biologically structured abstraction. While the basal variant provides strong predictive performance by maintaining fine-grained gene resolution, the macrogene framework introduces a higher-level, evolutionarily informed representation that better captures broad transcriptional programs. This structured representation enhances interpretability and provides a natural framework for gene-program-level comparison across species that are not accessible in the basal embedding alone. Together, the two variants illustrate how predictive accuracy and biological interpretability can be integrated within a unified cross-species framework. This added interpretability enables analyses that would be difficult to achieve under strict one-to-one orthology. Unify uncovered functionally convergent gene programs that reinforce cell-type identity across species—illuminating mechanisms of evolutionary conservation in how cell types form—and simultaneously identified divergent genes whose variation underlies species-specific adaptations within the same cell type, shedding light on cellular diversification. These dual insights not only boost the accuracy of cross-species perturbation prediction—a critical step for translating model-organism findings to human disease—but also generate concrete hypotheses about the origins and evolution of cell types. A further key outcome is our 700-million-year, high-resolution cell-type tree of life, which provides a study framework for tracing both convergent and divergent evolutionary trajectories. Together, these achievements establish Unify not merely as an integration tool but as a true engine of biological discovery, capable of revealing how cell types emerge, adapt, and diversify across the tree of life. Despite its advances, Unify has limitations that open avenues for future work. Its performance is inherently tied to the quality of its underlying foundation models; as protein and text-based models become more powerful and biologically nuanced, Unify’s precision will continue to improve. While our macrogenes offer enhanced interpretability, developing automated methods for their functional annotation and characterization remains an important next step. Furthermore, extending the Unify framework to integrate other data modalities, such as epigenomic (scATAC-seq) or spatial transcriptomic data, represents an exciting frontier for building even more comprehensive cross-species atlases. In conclusion, Unify represents a significant conceptual and technical step forward for comparative single-cell genomics. It moves the field beyond the limitations of sequence homology to a more holistic, function-centric paradigm. By providing a robust and versatile framework for integrating, interpreting, and predicting cellular behavior across the tree of life, Unify provides a framework for unraveling the fundamental design principles of cell types and ultimately contributes to the ambitious goal of constructing high-resolution, predictive virtual cells53. Methods Overview of Unify Unify takes multiple annotated single-cell RNA expression count datasets generated from S species XS1, XS2,…, XSs, where XSi belongs to N+Csi×∣Gsi∣, Csi is the number of cells in species Si and Gsi is the set of genes in species Si. We used cell type annotation information sourced from the original datasets. Unify also uses two additional types of information: (1) the p-dimensional protein embeddings P∈R∣G∣×p generated from large protein language models; (2) the l-dimensional gene functional description embeddings L∈R∣G∣×l generated from the general-purpose large language models, LLaMA-2, where G=∪i=1sGsi. Unify initializes Mprot protein-macrogenes and Mfunc functional description macrogenes by conducting the k-means clustering on the protein embedding spaces and gene functional description embedding spaces separately across all the species. Unify links the genes from multi-species to the protein-macrogenes and functional description macrogenes with weights WMprot\inR+∣G∣×Mprot,WMfunc\inR+∣G∣×Mfunc, where Wg,m,Wg,n\inR+, representing the weight from a gene g∈G to macrogenes m∈Mprot,n∈Mfunc. Further, for each species Si, Unify maps its expression data to a joint low-dimensional macrogene space by combining the protein-macrogene and functional-description macrogene together through: Mfinal=XMsiprot,XMsifunc where XMsiprot=Xsi×WMprot,XMsifunc=Xsi×WMfunc. Unify further maps the multispecies’ macrogene dataset to a joint d-dimensional latent space Z using an adversarial autoencoder framework guided by two auxiliary classifiers: a cell type classifier to promote the retention of cell type identity, while the other removes species-specific information through adversarial training. The resulting latent embedding captures conserved cellular features across species. We further applied conditional decoding based on given species labels to accurately reconstruct either the original gene expression profiles or the macrogene dataset. To balance competing objectives: preserving biological signals, removing technical batch effects, and ensuring faithful reconstruction, we used the MGDA-UB multi-task learning algorithm25, aiming to identify model parameters that achieve an optimal trade-off across these goals. Generation of protein embeddings using a protein language model Protein embeddings were generated using a pretrained protein embedding language model applied to the reference proteome of each species (Supplementary Data 13). The protein embedding language model takes the protein’s amino acid sequence as input and outputs a dimensional vector representing the embedding of the protein. In this study, ESM was used to generate protein-based embeddings because its suitability is supported by previous work, which showed that ESM-based embeddings are effective for constructing biologically meaningful macrogene representations in cross-species integration10,11. For benchmarking, protein embeddings were generated using the ESM-1b27 model; in all other analyses, we employed the updated ESM-221 model for embedding generation. Specifically, the ESM-1b model outputs a dimensional vector of 1280 to represent the protein, while ESM-2 model outputs a dimensional vector of 5120 to represent the protein. Any other protein language models that can reflect the sequence characteristics of the coding gene can be used as an input to Unify. Generation of gene function description embeddings from general-purpose large language model Gene functional description embeddings were generated from the general-purpose pretrained large language model using each species’ Gene Ontology functional description for each gene sequence. The large language model takes the text description as input and outputs a dimensional vector representation of the embedding of the genes’ function. We used the open-source LLaMA-224 model to generate the functional description embedding of each gene in the species. LLaMA-2 was chosen because it provides a strong, fully open-weight framework for representing semantic information from functional descriptions. As gene functions are often described in diverse and complex natural-language formats across databases, a general-purpose language model with robust semantic understanding offers an effective way to encode functional context. The gene functional description can either be downloaded from the NCBI website54, Gene Ontology website55, Uniprot database56, or generated de novo based on the protein sequences (Supplementary Note 1). Apart from the Gene Ontology description, any other term that describes the gene function, such as KEGG pathways or Reactome pathways, could be used to generate the gene functional description embeddings. We envision that any other general-purpose language model could be used to generate embeddings that work as inputs to Unify. Macrogene initialization To capture gene-level similarity from both structural and functional perspectives, we constructed two sets of macrogenes: one based on protein embeddings and the other based on functional description embeddings. These macrogenes serve as interpretable meta-units representing groups of genes with shared sequence or functional properties. Protein-based macrogene construction For each gene gsi from species Si, we obtained its protein embeddings egsiprot∈Rdp, derived from pretrained protein language models. For genes encoding multiple protein isoforms, protein embeddings were generated for each isoform sequence independently and subsequently averaged to represent the corresponding coding gene. All embeddings across species are concatenated into a joint matrix: We applied Kp-means clustering to this matrix to identify Kp protein-based macrogenes, where each cluster centroids Mkprot∈Rdp. Functional-based macrogene construction Similarly, we generated functional description embeddings egsifunc∈Rdf using the pretrained large language model based on gene function descriptions. They are aggregated as: We performed Kf-means clustering on Efunc to generate Kf functional macrogenes, where each cluster centroid Mkfunc∈Rdl. Gene-to-macrogene weight initialization For both protein and functional macrogenes, we computed a soft assignment weight matrix W∈RN×K, where Wgk represents the contribution of gene g to macrogene k. These weights are initialized using a ranked Euclidean distance between each gene and all the macrogene centroids. Let rdg,k∈{1, 2,…,R} denotes the rank of all the genes g for a specific macrogene k. rdg,k=1 indicates the nearest gene to macrogene k. The initial weight is defined as: where ⋅∈{prot,func}. Macrogene expression initialization Given the raw gene expression matrix Xsi∈RCsi×∣G∣ for species Si, where Csi is the number of cells, and ∣G∣ is the number of genes, the macrogene expression matrix M∈RCsi×K is computed as: for protein-based and functional-based macrogenes, respectively. Combined macrogene representation The two macrogenes’ expression matrices are concatenated to form the combined macrogenes expression representation: This combined representation serves as input to the downstream component of Unify. Adversarial learning and Pareto optimization during training with an autoencoder After obtaining the combined macrogene representation for each species, Unify employs an adversarial autoencoder to learn species-invariant but biologically informative embeddings of single cell transcriptomes. The training process is designed to ensure that the latent representation retains meaningful cell-type information while discarding species-specific variation. This is achieved through a combination of three loss terms: reconstruction loss, cell-type classification loss, and adversarial species classification loss, which are balanced through Pareto optimization. Autoencoder framework Given a transcriptome expression profile Xc for a cell c, we computed a macrogene activation vector ec by applying a transformation based on gene-to-macrogene weights: Here, W∈R∣G∣×∣M∣ is the gene-to-macrogene weight matrix, and LayerNorm refers to standard layer normalization: The macrogene representation ec∈R∣M∣ is then passed through a four-layer neural network encoder with ReLU activations, layer normalizations, and dropout to produce a low-dimensional basal embedding: Further, the decoder De reconstructs either the original macrogene representation or the expression profile from the basal embedding, conditioned on the species label sc: The reconstruction loss is defined as: Adversarial training To enforce biological relevance and species-invariance in the embedding space, we introduced two auxiliary classifiers: A cell type classifier that encourages cell identity by minimizing the cell-type classification loss: A species classifier trained adversarially to remove species-specific signals by maximizing species prediction error: The overall loss during training is expressed in the equation below: A naive approach could be optimizing a fixed set of alpha, beta, and gamma. However, manually tuning these weights is subjective and fails to capture the dynamic trade-offs during training. To resolve this, we framed the problem as multi-objective optimization and employed the Multiple Gradient Descent Algorithm-Upper Bound (MGDA-UB)25 to find a Pareto-optimal solution. Instead of using static weights, MGDA-UB treats each loss as a distinct objective. At each training step, it computes the gradients of each loss with respect to the shared latent cell embeddings. It then solves for a common descent direction that represents the best trade-off among the competing objectives, effectively finding weights that steer the model towards the Pareto front—the set of solutions where one objective cannot be improved without degrading another. This principled approach removes the need for arbitrary loss weighting and ensures that Unify robustly and automatically balances the competing goals of bio-conservation and batch correction while retaining its ability to reconstruct the input data. Perturbation prediction using Unify Unify supports two strategies (Unify-macrogene and Unify-basal approach) for predicting perturbation responses from a source species to a target species. Unify-macrogene is an approach based on macrogene spaces, which encodes abstracted gene modules from both structural and functional features learned from both species. Our perturbation transfer strategy follows the linear-manifold assumption introduced in scGen26 for cross-species perturbation modeling. We assumed that perturbation effects can be represented as a direction (or shift vector) in the shared macrogene embedding space, and that this perturbation-induced shift is approximately conserved across species. We used the below function to predict the target species’ macrogene under perturbation: where Macro^targetperturb is the predicted macrogenes of the target species under perturbation; Macrotargetcontrol is the macrogenes of the target species under control condition. Macrosourceperturb is the macrogenes of the source species under perturbation. Macrosourcecontrol is the macrogenes of the source species under control condition. Once Macro^targetperturb is obtained, it is then passed through Unify’s decoder to generate the corresponding perturbed gene expression profile of the target species. In this approach, perturbation prediction was carried out in the integrated embedding space by jointly integrating source-species control and perturbation data with target-species control data. We assumed that the cross-species mapping learned in the control state is preserved well enough under perturbation to support transfer to the unseen target state. A linear transformation F is trained to align the basal embeddings of the control condition from the source species to the target species. After the model was trained, we used the model to predict the basal embedding under the perturbed condition of the target species from the source species. The transformed embeddings Z^targetperturb are directly input into Unify’s decoder to obtain the predicted perturbation-response expression profile in the target species. Both strategies leverage Unify’s adversarially learned, species-invariant decoder, yielding accurate, high-resolution predictions of how perturbations translate across evolutionary distances. Differentially expressed genes Datasets were first filtered to retain cells with at least 200 detected genes and genes expressed in at least 10 cells. The filtered data were then normalized and log1p-transformed. Differentially expressed genes (DEGs) were identified using the “rank_genes_groups” function in SCANPY57 with multiple-testing correction applied to obtain adjusted p-values. Genes with adjusted p-value < 0.05 were defined as significant DEGs. For top-N DEG evaluation, significant genes were ranked by descending absolute log fold change, and the highest-ranked genes were used for hit-rate and directional accuracy analyses. Statistics We used SCANPY57 to perform the Wilcoxon rank-sum test for calculating the differentially expressed macrogenes cross species. Specifically, we used the “rank_genes_groups” function with default parameters. Macrogenes with adjusted p-value < 0.05 were considered as the significant differentially expressed macrogenes. The R2 between the ground truth gene expression profile and the predicted one is calculated using the function “r2_score” from sklearn.metrics58. Statistical comparisons of method performance in perturbation prediction were performed in R59 (stats package) using paired two-sided t-tests across cell types/conditions, comparing Unify-macrogene and Unify-basal against each other and all competing methods. Resulting p-values were adjusted for multiple testing using the Benjamini-Hochberg procedure. Structural similarity between genes was calculated as the cosine similarity between their ESM-derived protein embeddings, and functional similarity was calculated as the cosine similarity between their LLaMA-derived functional description embeddings. Because these two similarity distributions are derived from different embedding spaces, ESM- and LLaMA-based similarities were min–max scaled separately to [0,1] before comparison. This normalization enabled relative structural and functional similarity to be compared on a common scale. A two-sided Mann–Whitney U test was conducted for comparing the biological reproducibility scores (scjackknife scores) per technical reproducibility scores (jumble scores) extracted from the unrooted cell type tree for seven phylogenetically distant species of SATURN and Unify. We used the function “mannwhitneyu” from scipy.stats60. Determining gene homologs Gene ortholog information was retrieved using the Ensembl BioMart tool61 via the Ensembl genome browser62. Benchmark evaluations We evaluated Unify, SATURN10, SAMap6, scGen26, scVI18, Harmony19, fastMNN15, BBKNN20, Scanorama16 and Seurat v4 CCA14 in the different phylogenetic distance species integration tasks. Among these methods, Unify, SATURN and SAMap accept all genes from the integrated species as input, while the other methods only take the one-to-one orthologous genes between the integrated species as input. For more details about each selected method, see Supplementary Note 2. For cross-species dataset integration purposes, we evaluate the methods’ performance in batch-correction and bio-conservation scores using scIB packages7. Batch-correction scores are calculated based on the average of below sub-metrics: batch ARI, batch ASW, graph iLISI, batch NMI, kBET, PCR batch, and graph connectivity. Bio-conservation scores are calculated based on the average of the sub-metrics: cell type ARI, cell type ASW, graph cLISI, cell type NMI, HVG conservation and trajectory conservation. We further combined batch-correction scores with bio-conservation scores in a ratio of 4:6 to form an overall score for direct comparison (Supplementary Note 3). Ablation analyses For the multimodality ablation, three model variants were compared: Unify, using both ESM-derived protein embeddings and LLaMA-derived functional text embeddings; Unify (ESM-only), using only protein embeddings; and Unify (LLaMA-only), using only text embeddings. For the optimization ablation, Unify (MGDA-UB) was compared with fixed-weight variants in which the relative contributions of the reconstruction loss (α), species-classifier loss (β), and cell-type-classifier loss (γ) were manually specified. Four representative weight settings were tested to span reconstruction-dominant (α = 0.8,β = 0.1,γ = 0.1), batch-correction-dominant (α = 0.1,β = 0.8,γ = 0.1), balanced (α = 0.33,β = 0.33,γ = 0.33), and bio-conservation-dominant (α = 0.1,β = 0.1,γ = 0.8) regimes. To assess dependence on foundation-model choice, alternative combinations of protein language models (ESM, ProstT5, and ProtT5) and text models (LLaMA, BLOOM, and PubMedBERT) were benchmarked. For the supervision ablation, three variants were evaluated: Unify, the original model uses cell type labels from both input species; Unify (semi-supervised), in which one species retained curated cell-type labels while the other used unsupervised cluster identities; and Unify (no cell-type classifier), in which the cell-type classifier was removed entirely. For the input-gene-space ablation, Unify, which uses all available genes as input, was compared with Unify(one-to-one), in which the input was restricted to one-to-one orthologs. Multimodal-initialization and optimization ablations were evaluated across ten integration tasks spanning cross-genus to cross-phylum comparisons. Foundation-model, supervision, and input-gene-space ablations were evaluated across five representative tasks spanning increasing phylogenetic distance (Supplementary Data 14). In all ablation analyses, model parameters and training settings were kept identical to those of the Unify model, and only the component under investigation was modified. Performance was assessed using the same integration metrics as in the main benchmark. Hyperparameter configuration and benchmarking fairness To ensure a fair and rigorous evaluation, we benchmarked Unify and all competing methods using consistent, default hyperparameter strategies, explicitly avoiding dataset-specific fine-tuning. For all baseline methods (e.g., SATURN, scGen, SAMap, scVI), we utilized the default parameters, standard preprocessing pipelines, and recommended best practices provided in their official software implementations. For Unify, we established a single, universal set of default hyperparameters that was applied consistently across all integration tasks. This global configuration is available in Supplementary Data 15. Finally, in all ablation studies, these exact network hyperparameters were held strictly constant. This ensures that any observed performance degradation in an ablated model (e.g., Unify-ESM only) is directly attributable to the removal of specific biological priors or architectural components, rather than confounding tuning artifacts. Protein sequences comparison To analyze and compare gene sequence conservation, we employed a multi-step approach. First, protein sequences of the target genes were aligned using CLUSTALW63 to perform multiple sequence alignment (MSA), enabling the identification of conserved and variable regions across sequences. The aligned sequences were then input into the SWISS-MODEL64 homology modeling server to predict the three-dimensional (3D) protein structures, which were returned in PDB format. Finally, we used ESPript65 (Easy Sequencing in PostScript) to integrate the MSA results with the predicted 3D structures. This tool enabled the visualization of secondary structural elements (e.g., α-helices, β-strands) and facilitated a detailed comparison of both sequence and structural differences among the selected proteins. GO term and KEGG enrichment analysis To investigate the biological functions and pathways associated with the component genes of the differentially expressed macrogenes, we performed Gene Ontology (GO) term and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analyses. Differentially expressed macrogenes were identified based on predefined threshold-adjusted p-values less than 0.05. For GO enrichment, we focused on the Biological Process (BP) category to capture functional aspects of gene regulation. Enrichment analysis was conducted using the clusterProfiler66 R package (version 4.10.1). Background gene sets were defined based on all expressed genes in the corresponding dataset. The enrichment results were calculated using a hypergeometric test, and multiple testing correction was applied using the Benjamini-Hochberg method. Terms with adjusted p-values (false discovery rate, FDR) below 0.05 were considered significantly enriched. For KEGG analysis, we used gene identifiers mapped to KEGG orthology via the org.X.eg.db annotation package (where “X” corresponds to the organism, specifically, Hs for human or Mm for mouse, Ss for pig, Mmu for Macaca mulatta, Mmf for Macaca fascicularis). KEGG pathway enrichment was performed using the same statistical framework as GO analysis. The enriched GO terms and KEGG pathways were visualized using dot plots for interpretability. Systematic evaluation of functionally divergent and convergent gene pairs Because no curated benchmark explicitly labels cross-species gene pairs as functionally divergent or convergent, these categories were defined using sequence relationships and GO-based functional similarity. To ensure high-quality annotation support, this analysis was restricted to the human–mouse comparison. Functional similarity between gene pairs was quantified using GO Lin semantic similarity33, an information-content-based measure derived from GO structure and annotation. Functional divergence was defined as reciprocal best-hit (RBH) ortholog pairs with low functional similarity. RBH orthologs were identified by BLAST32 using default parameters, and pairs with GO Lin similarity ≤ 0.3 were classified as functionally divergent. Functional convergence was defined as non-RBH gene pairs lacking significant sequence homology but showing high functional similarity. Specifically, gene pairs with BLAST E-value > 0.05 and GO Lin similarity ≥ 0.7 were classified as functionally convergent. Only gene pairs within the BLAST search space were considered in the convergence analysis. For each macrogene, the top K human and mouse genes were selected according to the learned gene-to-macrogene weight matrix produced by Unify. All possible cross-species gene pairs were then generated as the Cartesian product of the two sets. A macrogene was considered to contain functional divergence or convergence if at least one qualifying gene pair was identified among these cross-species combinations. To test whether such relationships were captured more often than expected by chance, permutation testing was performed using 1000 randomizations of gene-to-macrogene assignments while preserving the number of genes assigned to each macrogene in each species. Enrichment was summarized as the fraction of macrogenes containing at least one divergent or convergent pair. To assess the contribution of the individual modalities, structural similarity and functional text similarity were evaluated separately for their ability to distinguish divergent and convergent gene pairs. Area under the receiver operating characteristic curve (AUROC) values were computed through the function “roc_auc_score” in sklearn.metrics by using ESM-derived structural similarity and LLaMA-derived functional similarity against the benchmark definitions described above, allowing assessment of whether either modality alone was sufficient to recover these relationships. Cell type tree construction To construct the cell type phylogenetic tree spanning seven species (Schmidtea mediterranea, Danio rerio, Ciona intestinalis, Mus musculus, Homo sapiens, Drosophila melanogaster, and Caenorhabditis elegans), we followed the approach described by Jasmine et al.50. First, we generated integrated cross-species embeddings. Cell types represented by fewer than 100 cells were excluded to ensure robustness. We then performed principal component analysis (PCA) on filtered data and retained the top 20 principal components for downstream analysis. To reduce computational complexity, one cell per cell type per species was selected. Specifically, for each species-cell-type, cells were clustered by k-means in the learned embedding space, and the resulting cluster centroids were averaged to obtain a meta-centroid. The cell nearest to this meta-centroid was selected as the representative cell (“centroid cell”) for tree construction. We constructed the phylogenetic tree using the contml program from the PHYLIP package (version 3.698), applying a Brownian motion model. Finally, to evaluate both technical and biological reproducibility of the resulting tree, we calculated jumble scores and scjackknife scores for each tip by randomly repeating the tree construction process and randomly selecting the starting tip of the tree for 500 times separately (Supplementary Note 4). To assess robustness to representation choice, an alternative tree was constructed directly by averaging the cell embeddings within each species-cell-type and compared with the centroid-cell-based tree using the weighted Robinson–Foulds distance67 (Supplementary Fig. 38 and Supplementary Note 4). Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Peer Review file Description of Additional Supplementary Files Supplementary Data 1 Supplementary Data 2 Supplementary Data 3 Supplementary Data 4 Supplementary Data 5 Supplementary Data 6 Supplementary Data 7 Supplementary Data 8 Supplementary Data 9 Supplementary Data 10 Supplementary Data 11 Supplementary Data 12 Supplementary Data 13 Supplementary Data 14 Supplementary Data 15 Reporting Summary Source data Source Data