Abstract 7286: Optimizing prediction of metastatic castration-resistant prostate cancer: Multimodal fusion of clinicopathological and genomic features with machine learning

Abstract Background: Transitioning from metastatic castration-sensitive prostate cancer (mCSPC) to its resistant variant poses risks of intensified medication, toxic side effects, and financial burdens for men undergoing treatment. Our study proposes a multimodal approach that integrates clinical, pathological, and genomic features to optimize prediction of metastatic castration-resistant prostate cancer (mCRPC) events. By applying machine learning (MML) techniques to multimodal data, we aim to stratify these patients based on the risk of developing mCRPC, contributing to the development of safer and personalized treatment strategies. Methods: We analyzed a large cohort of patients with mCSPC (n=399) with a minimum of 36 month follow-up, annotated with clinical, pathological, and genomic features, encompassing relevant variables such as disease volume, PSA, metastases timing, variant classification, Grade Group, fraction of genome altered, MSI score, mutation counts, karyotypes, tumor mutational burden, and purity. Using a multimodal machine learning framework, we extracted informative features and employed eight distinct machine learning algorithms: K-Nearest Neighbor, Support Vector Machine, Decision Tree, and Logistic Regression, Support Vector Machine with RBF Kernel, Random Forest, and Gradient Boosting, in addition to a Deep Neural Networks architecture for building predictive models in supervised binary classification experiments. The models' performance was evaluated on separate training (80%) and testing sets (20%), and standard evaluation metrics, such as the area under the receiver operating characteristic curve (AUROC), and accuracy, were computed. Correlations were utilized to explore the relationships between features, and permutation analysis was employed to determine the importance of predictors. All raw data included in the analysis are available on cBioPortal database. The study is supported by the NIH-NCI-T32 for Next Generation Pathologists Program at our institution. Results: Of the 399 mCSPC patients, 56% developed castration resistance, 50% had Grade Group 5 disease, and 69% exhibited missense mutations. In a series of 8 experiments, the Support Vector Machine with RBF Kernel ensemble model and SVC exhibited the highest performance with an AUROC score of 0.77, while the remaining models tested in the study demonstrated AUROC scores ranging from 0.62 to 0.75. Notably, the fraction of altered genome emerged as the top-ranked predictor, displaying the highest importance score. Conclusions: Our study demonstrates the strong predictive performance of MML models in accurately identifying mCRPC events by leveraging routine clinical, pathological, and genomic data, offering valuable guidance for optimizing patient stratification and tailoring personalized therapeutic interventions. Citation Format: Mohammad K. Alexanderani, Francesca Khani, Matthew Greenblatt, Brian Robinson, Massimo Loda, Luigi Marchionni. Optimizing prediction of metastatic castration-resistant prostate cancer: Multimodal fusion of clinicopathological and genomic features with machine learning [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2024; Part 1 (Regular Abstracts); 2024 Apr 5-10; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2024;84(6_Suppl):Abstract nr 7286.