EACR 2025 Congress: Innovative Cancer Science: Deep learning in digital histopathology for clinically applicable biomarker discovery in prostate cancer

IntroductionProstate cancer is a leading cause of cancer-relatedmortality, with disease heterogeneity complicating riskstratification and treatment decision-making. AI modelsdeveloped using gene expression or digital histopath-ology have demonstrated significant prognostic value forcancer patients. However, the biological interpretation ofthese AI models, and the integration of these multi-modaldata types, could enhance clinical applicability andprognostic value, respectively. This research aims to usefunctional genomics to improve AI model explainability,and deep learning to refine prostate cancer riskstratification.Material and methodThis study will utilise two in-house prostate cancercohorts: the FASTMAN cohort, which includes 466locally advanced patients, and the FIR cohort, whichconsists of 159 intermediate-risk patients. This study willalso utilise the 499 patients within the TCGA-PRADcohort. Self-supervised learning (SSL) will be used ondigital prostate histopathology whole-slide images(WSIs) to extract morphological features without relianceon extensive annotations. Functional genomics throughgene expression will be used to biologically interpret theArteraAI Prostate Test, and hypertension status. SSL-derived histopathology features, gene expression, andclinicopathological factors will be integrated using deeplearning autoencoders for patient subgroup discovery.Molecular Oncology 19 (Suppl. 1) (2025) 1-940© 2025 The Authors. Molecular Oncology is published by John Wiley & Sons Ltd on behalf of Federation of European Biochemical Societies59Result and discussionThrough the exploration of different methods within theanalysis pipeline of SSL methods in digital histopath-ology, we will provide a benchmark for other researchersto avail to, as well as developing SSL-derived histopath-ology features for our curated prostate cancer patientcohorts. The biological interpretation of the ArteraAIProstate Test will enhance its explainability for greatertrust towards future use in clinical practice. Hypertensionstatus was significantly associated with better outcomesin disease progression and overall survival for locallyadvanced prostate cancer patients treated with radio-therapy and ADT. We anticipate that the integration ofthe described multimodal data types will reveal novelpatient subgroups with improved patient stratification.ConclusionThrough SSL digital histopathology analysis, functionalgenomics, and multimodal data integration, this researchaims to enhance risk stratification and characterize theunderlying biology associated with prognostic AI modelsand hypertension status in prostate cancer patients