HESpotEx Revolutionizes Gene Activity Prediction Without Transcriptomics Data

A new deep learning tool called HESpotEx is changing how researchers study gene activity in tissue samples. The framework uses standard histopathological images to predict gene expression with high accuracy. Its performance surpasses existing models, even across complex datasets involving both cancerous and healthy tissues. HESpotEx works by combining two key approaches: image analysis and graph-based data representation. This dual-stream design allows it to capture both molecular patterns and tissue structure from whole-slide images (WSIs). Unlike previous methods, it does not rely on spatial transcriptomics (ST) data, making predictions faster and more accessible.

The model can estimate expression levels for up to 5,457 genes within individual spots on a tissue sample. Tests on large datasets, including The Cancer Genome Atlas (TCGA), confirm its reliability across different tissue types. Importantly, its predictions remain interpretable—researchers can overlay gene activity heatmaps directly onto tissue images. Beyond accuracy, HESpotEx improves consistency in spatial transcriptomic studies. It helps identify diagnostic features in WSIs, offering pathologists new quantitative tools. This could speed up molecular profiling, aiding earlier detection of cancerous changes and supporting tailored treatment plans.

HESpotEx provides a faster, more precise way to predict gene expression from standard tissue images. Its ability to work without ST data while maintaining interpretability could streamline cancer research and clinical diagnostics. The framework’s cross-sectional reliability also strengthens the quality of spatial transcriptomic analyses.