Explainable Prototype Booster: Enhancing Latent Representations of Foundation Models for Gene Expression Prediction

Spatial transcriptomics (ST) is a cutting-edge technology that measures gene expression while preserving spatial context and generating pathology-grade tissue images. Although ST has enabled numerous discoveries and demonstrated a huge application potential in pathological diagnosis and prognosis, the technology remains time-consuming and costly. The ability to predict gene markers of cancer from histological H&E-stained tissue images can overcome these technological barriers to open new horizons for precision and personalised pathology. Recently, foundation models have demonstrated improvements in generating general-purpose embeddings of H&E-images. However, these improved representations are not optimized for gene expression prediction and lack task-specific adaptability. To address this limitation, we propose the Explainable Prototype Booster (EP-Booster), which incorporates biological prior knowledge to guide the construction and training of learnable prototypes for embedding refinement, thereby improving gene expression prediction. Importantly, model predictions are inherently interpretable through pathway-level attributions associated with the prototypes. Extensive experiments across multiple datasets, cancer types, and spatial transcriptomics platforms demonstrate that EP-Booster consistently outperforms existing methods. Moreover, EP-Booster can be integrated with diverse foundation models to enhance task-specific representations, thereby improving predictive performance and biological interpretability in clinically relevant applications, including cancer biomarker prediction, survival analysis, and drug response prediction.