A Deep Learning Framework Integrating Tumor Microenvironmental Features Accurately Predicts Multiple Driver Gene Mutations in Lung Cancer Pathology Images

Abstract Deep learning (DL) has the potential to enable the prediction of gene mutations directly from routine histopathology slides in lung cancer. However, existing approaches have largely been limited to mutation-level prediction and have not achieved precise identification of driver mutation subtypes or exonic variants, constraining the translation of DL into targeted therapy. In this study, we assembled a large multicenter dataset of paired pathology images and next-generation sequencing from 2573 lung cancer patients from four hospitals in China. Development of NAVF-Bio, an adaptive multi-view feature fusion framework based on multiple-instance learning, enabled integration of tumor microenvironment (TME) features from whole-slide images (WSIs) to predict driver mutations and tumor mutational burden (TMB). Benchmarking against 11 state-of-the-art DL methods indicated that NAVF-Bio consistently outperformed existing models in predicting driver mutations (TP53, EGFR, KRAS, ALK) and TMB status, achieving clinically relevant performance in external multicenter validation. Notably, NAVF-Bio accurately predicted the mutated driver gene exons across centers, while interpretability analyses using WSI visualization and TME quantification further demonstrated the ability of NAVF-Bio to elucidate pathologically relevant tumor features. Finally, a multi-gene mutation prediction platform for lung cancer was generated to facilitate the screening of driver gene mutations. Overall, NAVF-Bio mimics the workflow of pathologists when examining slides by observing multi-scale features of WSIs and TME characteristics to predict driver gene mutations in lung cancer, which could guide the selection of targeted therapies for patients.