Identification of a Predictive Genomic Signature for NADT Response in High-Risk Prostate Cancer Using Real-World Clinical Trial Data Analyzed by the KEM® Explainable Artificial Intelligence Platform
Prostate cancer is one of the most common malignancies worldwide, and treatment decisions for localized high-risk disease increasingly rely on molecular predictors to guide therapy selection. Neoadjuvant androgen deprivation therapy (NADT) can improve outcomes in selected patients, but to date, clinicians lack validated genomic indicators to distinguish exceptional responders from poor responders.
To address this need, Ariana Pharmaceuticals, in collaboration with Philips, conducted an Explainable AI–driven study to identify genomic features predictive of NADT response. The results were presented as a scientific poster at ESMO-AI 2025 in Berlin.
A total of 58 patients from two clinical studies—NCT02430480 (Wilkinson et al., 2021) and NCT02268175 (Tewari et al., 2021)—were analyzed using the KEM® (Knowledge Extraction and Management) platform. KEM® is an Explainable Artificial Intelligence (eXp.AI) technology based on Formal Concept Analysis (FCA) that systematically evaluates all relationships within a dataset to reveal transparent, interpretable biomarker signatures.
A curated panel of 100 gene symbols was selected based on biological relevance, including genes involved in viral–bacterial defense, T-cell receptor signaling, and PDE4D7-associated androgen response pathways. Gene expression values were discretized into tertiles using an unsupervised method designed to support logical reasoning and reduce noise.
Using the KEM® Biomarker methodology, predictive signatures were generated from association rules acting as voters in a majority-vote system. Models were optimized via 3-fold cross-validation, and 50 repeated train–test splits were performed to strengthen robustness and rank recurrent genomic features. This approach acted as an effective regularization method for smaller clinical datasets.
The resulting KEM® signature identified four key genes — GUCY1A1, PLS3, ARSD, and CUX2 – whose high expression predicted favorable response to NADT. The signature also provided patient-level interpretability via a rule-based coverage matrix, offering clinicians insight into why a patient was classified as a likely responder or non-responder.
When benchmarked against Immunoscore and conventional multivariate approaches, the KEM® signature demonstrated superior predictive performance and enhanced transparency, highlighting the promise of explainable AI in biomarker discovery and oncology decision support.
This preliminary work lays the foundation for further validation as additional prospective and real-world patient data become available. The study underscores the potential of Explainable AI to enable precision stratification and predictive biomarker development in prostate cancer treatment.
