Multi-cohort validation of machine learning radiogenomic models (ML-RGx) to predict late toxicity in prostate cancer

PROJECT SUMMARY Radiotherapy is a cornerstone of treatment for prostate cancer, but radiation-related genitourinary (GU) and gastrointestinal (GI) toxicities can negatively impact quality of life among survivors. Radiotherapy can damage the bladder and rectum leading to gross bleeding, inflammation, pain, fibrosis, and when severe, life- threatening complications. Up to 20% of men treated with radiotherapy for prostate cancer develop mild to moderate late GU and/or GI toxicities that are often permanent and negatively impact quality of life; up to 5% develop severe or life-threatening effects requiring medical or surgical intervention. Radiation exposure drives risk of late toxicity, but genetic predisposition is a significant contributor and can explain why some patients develop toxicity while others no not despite identical treatment plans. Our prior work shows that late GU and GI toxicities are polygenic in etiology, with risk modified by the combined effects of many low-penetrance single nucleotide polymorphisms (SNPs), raising the attractive possibility of using a polygenic risk score to identify susceptible patients prior to starting radiotherapy. Towards this goal, we developed a novel machine learning approach to combine information from many risk SNPs and dose-volume parameters into a radiogenomic (ML- RGx) risk score. Our preliminary data shows that this modelling approach out-performs existing methods and shows promise for use in the clinic. The proposed project will apply this method to a large training dataset from the NCI-supported International Radiogenomics Consortium to build ML-RGx models of GU and GI toxicity that will then be externally validated using data and biospecimens from two large phase III radiotherapy trials completed through the NRG Oncology cooperative group. Bioinformatic approaches will be applied to prioritize SNPs for inclusion in the modelling and to uncover biologic pathways underlying genetic predisposition to normal tissue injury. The study has three aims: (1) to train ML-RGx models for each of radiation-induced GU and GI toxicity and define a threshold for low and high risk; (2) to validate ML-RGx models in two independent datasets from NRG Oncology cooperative group trials; and (3) to assess feasibility and impact of ML-RGx models on treatment planning workflow in a Radiation Oncology clinic. This work will bring personalized medicine to the field of radiation oncology and improve prostate cancer care. Our innovative modelling approaches will also uncover important molecular pathways that could be targeted with interventions to prevent and/or mitigate toxicities.