Predicting gene expression from whole slide images in prostate cancer using deep learning

Prostate cancer exhibits complex heterogeneity, requiring resource-intensive sequencing to characterize its diversity for precision medicine. While image-based transcriptomic prediction models offer a promising alternative, current approaches lack comprehensive validation and downstream interpretation. Here, we propose ProGENIE, a novel multi-head attention-pooling framework to predict transcriptomics from whole slide images (WSIs) for prostate cancer research and clinical utility. Trained on The Cancer Genome Atlas (TCGA), ProGENIE demonstrates strong generalizability on an independent South Australian hospitals (SAH) cohort, achieving a median Pearson correlation coefficient (PCC) close to 0.6 for the top 1,000 genes expression and accurately predicting 3,167 genes expression with PCC > 0.4. ProGENIE accurately predicts gene expression associated with prostate cancer development and reliably characterizes the tumor microenvironment. Furthermore, the predicted transcriptomic profiles are correlated with drug sensitivity and immunotherapy response. This cost-effective approach links tissue morphology to molecular profiles, supporting personalized treatment in prostate cancer.