Systematic benchmarking and optimal strategy selection of cross-species integration methods

Single-cell RNA sequencing provides an unprecedented resolution for cellular heterogeneity and gene regulation, fostering cross-species comparative analyses with increasing interspecies data. However, integrating single-cell transcriptomic data faces challenges, including gene selection, evolutionary distance, and batch effects, with varying method performances. We utilized single-cell transcriptomic data from hippocampal tissues of seven mammals (e.g. mouse, human), evaluating 13 mainstream integration methods across 27 tasks with 11 metrics. To compare the performance of different methods, we developed a machine learning-based scoring model that assesses the contribution of each metric in a data-driven manner, thereby addressing the oversimplified assumptions of traditional manual weighting approaches. Our findings show that selecting highly variable one-to-one orthologous genes best balances species differences and commonalities. Most methods integrated closely related species, whereas scVI, a probabilistic model with distributions specified by deep neural networks, and its semi‑supervised extension scANVI, as well as the Seurat v5 method, which uses reciprocal principal component analysis (RPCAv5), effectively mapped distantly related species. Increased species numbers reduce gene overlap and heighten heterogeneity, increasing integration difficulty. The scANVI best maintained quality by balancing the biological signals and batch effect removal. Furthermore, we established an evaluation website to guide researchers in selecting the optimal integration methods for cross-species single-cell transcriptomic data analysis. Collectively, our findings provide a systematic, evidence-based framework that can assist researchers in rapidly selecting appropriate integration methods for cross-species single-cell transcriptomic studies.