Prediction of molecular subtypes from histology: AI-driven analysis of prostate cancer morphological patterns and therapeutic implications

Abstract Molecular subtypes in prostate cancer significantly influence disease characteristics and treatment outcomes, yet obtaining this information requires specialized molecular testing. In this study, we develop and validate artificial intelligence models that can predict PAM50 and Prostate Subtyping Classifier (PSC) molecular classifications directly from standard hematoxylin and eosin (H&E)-stained biopsy slides. Using a cohort of 903 biopsy slides from 424 patients with matched molecular data, we demonstrate that our novel UNIv2-MIL framework, which fine-tunes a pre-trained pathology foundational model (UNIv2) using a multiple instance learning (MIL) strategy, achieves AUCs of 0.863 and 0.81 for PAM50 and PSC subtyping, respectively. Through computational clustering of high-attention regions, we identify histological patterns associated with molecular subtypes. Furthermore, in an independent validation cohort of 131 patients, we observed that patients with UNIv2-MIL predicted luminal subtypes tended to be more responsive to hormone therapy (HT) ( p < 0.03). We also examined another independent cohort of 122 patients who transitioned from active surveillance to radical prostatectomy (RP) and observed that model-predicted PAM50 luminal B and PSC luminal proliferating scores showed significant correlation with adverse pathologic features (APFs) ( p < 0.001 and p = 0.003, respectively). In conclusion, while further validation is needed, we have developed an AI-driven approach based on routine histopathology that has the potential to inform prostate cancer risk stratification and treatment response.