Developing deep learning model for predicting treatable molecular alterations from H&E-stained histological section slides

Determining biomarkers—such as microsatellite instability (MSI), TP53 mutations, and other molecular alterations—is crucial for guiding effective cancer treatment strategies. However, the most reliable way to identify these genomic features currently requires whole-genome sequencing, a procedure that is costly for widespread clinical use. As a result, treatment decisions often depend mainly on individual pathologists who analyze whole slide images (WSIs) based on cell architecture, staining patterns, and other histopathological cues and molecular analysis are only used when treatments have not been effective. Deep neural networks have proven highly effective in learning complex data representations for image classification, offering a potential solution to this problem. In this work, we present a deep learning framework for prostate cancer that classifies WSIs according to their likelihood of having specific genomic alterations—such as MSI or TP53 mutations—using histopathological features alone. To accomplish this, we leverage deep foundational models, such as UNI, to extract feature embeddings from pathology pictures and incorporate data from multiple cancer types, thereby enhancing the network’s generalization capabilities. By employing this model, clinicians can identify patients at higher risk for alterations without relying on costly whole-genome sequencing for every patient. Our research will stratify patients into low risk and high-risk groups for these biomarkers, allowing for a cost-effective introduction of molecular diagnostics tools with more precision, to become part of the routine clinical procedures.