Transfer Learning Of Gene Expression Using Reactome

AO_SCPLOWBSTRACTC_SCPLOWIn clinical research, translating findings from model organisms to human applications remains challenging due to biological differences between species as well as limitations of orthologous, and homologous, gene comparisons, which is fraugt with information loss as well as many-to-many mapping. To address these issues, we introduce a novel Universal Gene Embedding (UGE) model that leverages transformer-based few-shot learning for species-agnostic transfer learning with heterogeneous domain adaptation. The UGE model, trained on a dataset of gene expression from ten organs across rats and mice, establishes a unified biological latent space that effectively represents genes from any organ or species. By focusing on reactomes--comprehensive profiles of gene expression responses to drugs--the UGE model enables functional gene mapping across species based on the similarities of these profiles. Our contributions include a gene reactome vector prediction module, a robust framework for mapping drug-induced gene expression patterns across species, strategies for optimizing experimental design, and enhanced gene mapping precision. These advancements provide a new tool for genetic research and a new paradigm for cross-species insights, potentially revolutionizing our understanding of gene function, drug responses, and the translation of findings from model organisms to human clinical applications.