AI-based Molecular Alteration Prediction from Histopathological Images in Prostate Cancer

Background: Identifying molecular alterations at cancer diagnosis can guide treatment selection, but comprehensive molecular testing is expensive and is not routinely performed for all patients. In this study, we aim to assess whether artificial intelligence (AI)-based models trained on routinely collected hematoxylin and eosin (H&E)-stained slides can infer molecular characteristics. Here we piloted a framework to predict six clinically actionable molecular alterations (AR, HRD, PTEN, TP53, RB1, and MSI-H/dMMR) in prostate cancer (PC).

Methods: A total of 924 samples (n = 875 patients) from three diverse clinical PC cohorts were included. The model was trained and internally validated on the TCGA-PRAD and University of Washington (UW) cohorts and externally validated on a clinical trial cohort. Our framework included four steps: 1) Whole slide images (WSIs) were divided into smaller tiles and stain-normalized; 2) Cancerous tiles identified by our in-house model were retained; 3) tile-level features were extracted using the foundation model UNI2; and 4) these features were aggregated using the multiple-instance learning ACMIL for slide-level binary classification of six molecular alterations.

Results: Our method achieved the highest area under the receiver operating characteristic curve (AUROC) for MSI-H/dMMR in hold-out validation (0.80+-0.06), followed by AR (0.76+-0.04), TP53 (0.74+-0.00), and PTEN (0.72+-0.03), while HRD (0.60+-0.04) and RB1 (0.56+-0.05) were lowest. Five-fold cross-validation showed similar trends. In external validation, AUROCs ranged from 0.58 to 0.73, with RB1 highest (0.73+-0.06) and HRD lowest (0.58+-0.07); other alterations showed intermediate performance (0.61 to 0.68) (Fig.1).

Conclusion: Our study demonstrates the feasibility of predicting molecular alterations from WSIs using deep learning in PC. Some alterations are more easily reflected in histomorphology than others, and more challenging ones may require specialized modeling strategies.

Learning Objective:

1. Understand the potential of AI-based digital pathology to infer molecular alterations from H&E-stained whole slide images.