Standardizing Lightweight Neural Adaptation of Pathology Foundation Models for Gene Mutation Prediction

Cancer care increasingly relies on molecular biomarkers and oncogenic mutations to guide therapy, but genomic testing remains costly and time-consuming. This motivates AI-based prediction of tumor genotype directly from routine H&E histopathology. We ask how large pretrained pathology foundation models (PFMs) can be adapted in a lightweight, standardized way for slide-level mutation prediction in colorectal cancer (CRC). We propose a simple adaptation branch that attaches to frozen PFM encoders such as TITAN and Prov-GigaPath. The branch combines a shallow feedforward neural architecture (FNA) augmented with calibration, thresholding, stabilization (CTS), and configurable multi-head adaptation (MHA). Using the TCGA-CRC cohort of approximately 600 COAD and READ patients, we systematically evaluated hundreds of gene targets through full-set screening, broad-threshold inclusion, and clinically significant gene panels. Logistic regression on PFM features was used as the baseline. The adapted models consistently outperformed baselines. Across approximately 790 retained TCGA-CRC genes, TITAN-FNA-CTS-MHA, increased mean AUROC from about 0.62 to 0.71. In curated 13-gene clinical and 20-gene CRC-focused panels, AUROC improved from 0.58–0.60 under logistic regression to 0.64. Adaptation also greatly increased the number of well-predicted genes, with 443/560 genes reaching AUROC ≥ 0.70 compared with 158/620 for logistic regression. Finally, experiments on six additional TCGA cohorts showed broader applicability; e.g., in UCEC, mean AUROC improved from 0.617 to 0.708 across 7106 retained genes. Overall, this work presents a rigorous, reproducible framework for PFM adaptation in histopathology-based biomarker prediction.