Cross-species integration of single-cell data reveals conserved pathology-associated cell populations across animal models and human samples.

Single-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.