HRDPath: An Explainable Multi-Model Deep Learning Architecture for Predicting Homologous Recombination Deficiency from Histopathology Images

Abstract

Homologous recombination deficiency (HRD) is a critical biomarker for guiding treatment decisions in high-grade serous tubo-ovarian carcinoma (HGSOC), a cancer with few reliable biomarkers. However, existing genomic-based tests for HRD are variable, expensive, and time-consuming. To this end, we developed HRDPath, a novel patient-level deep learning architecture that combines the strengths of two complementary models with a multi-task design, to predict genomically derived HRD status from whole slide images in HGSOC. HRDPath was comprehensively validated across three datasets and benchmarked against leading deep learning models. It achieved an AUC of 0.846, surpassing previously reported H&E-based HRD prediction results for HGSOC images by 0.09, and for the first time, reporting a specificity of 0.938, where accuracy significantly increased when multiple slides per patient were used. Our proposed patient-level approach and interpretability pipeline enhance model trustworthiness and reveal important clinical and biological insights into HRD-positive cancers, highlighting the associated morphological and pathological changes at the cellular and tissue levels. HRDPath is a potentially accessible and scalable digital biomarker that could improve ovarian cancer diagnosis and therapy selection.