CSLAN: Cross-Species Latent Alignment Network for Trauma-Related Cell-Type Classification

Understanding cellular identity and function across species is fundamental to deciphering conserved biological principles and translating insights from model organisms to human health. However, integrating single-cell transcriptomic data, particularly from resource-limited human studies, with comprehensive datasets from model organisms like mice remains a significant challenge due to evolutionary divergence and technical variability. Here, we introduce the Cross-Species Latent Alignment Network (CSLAN), a transfer learning framework that effectively bridges this gap for robust cell type classification. CSLAN employs a novel strategy involving focused feature selection on source (mouse) data followed by pretraining an encoder-decoder architecture. Crucially, for transfer to human data, only the encoder is fine-tuned while the latent space processor and the decoder, which encapsulate learned, conserved cell type signatures, remains fixed. This targeted adaptation mechanism allows CSLAN to achieve exceptional accuracy (e.g., 96.67% for human immune cells in a trauma context) by efficiently aligning human gene expression profiles to a biologically informed latent space derived from mouse data. Our findings establish a powerful and broadly applicable paradigm for cross-species knowledge transfer, enabling deeper biological insights from limited human scRNA-seq data by leveraging the wealth of information from model systems, with significant implications for translational research and the construction of comprehensive, cross-species cell atlases.