COAST Predicts Gene Expression from Histology Images

A new framework named COAST (Context-Aware Differential Learning) has been developed to predict spatial gene expression directly from standard H&E histopathology images. This innovation addresses the high cost and low throughput limitations of current spatial transcriptomics technologies. Existing methods primarily focus on predicting absolute gene expression levels, but COAST introduces a novel approach by explicitly leveraging the relative expression relationships between different spatial spots. COAST integrates local and global contextual features, modulated by cell type, and uses a Transformer encoder to capture both fine-grained local patterns and broader slide-level structures. Its training objective combines traditional absolute expression regression with a unique signed differential regression, which accounts for the relative differences in gene expression. Experiments across various spatial transcriptomics datasets demonstrate that COAST consistently improves prediction accuracy, as measured by both correlation and distribution-based metrics, highlighting the effectiveness of its context-aware differential learning strategy.