Deep learning on genome-wide association studies to predict the patient-specific risk of radiation-induced erectile dysfunction.
Academic Article
Overview
abstract
OBJECTIVE: Radiation-induced erectile dysfunction (RIED) is a frequent, unpredictable complication following radiotherapy for prostate cancer. We hypothesized that genome wide genetic variants and clinical data could be analyzed using a novel deep learning model that incorporates known biological components to improve the patient-specific prediction of RIED risk. MATERIALS AND METHODS: Germline DNA from blood samples of 668 prostate cancer patients treated with radiotherapy was genotyped through two-phase genome-wide association studies (GWAS) in the GenePARE study. Evaluable participants (N = 387) did not have erectile dysfunction (ED) prior to radiotherapy and were categorized into RIED cases (N = 221) or non-RIED controls (N = 166). To predict the risk of RIED, we developed a biologically informed deep learning model called BioDeepGWAS. This model was trained on both genetic and clinical data. The genetic input included single nucleotide polymorphisms (SNPs) with a univariate association p-value of less than 0.001. Prior to modeling, the entire dataset was stratified into training (70%), validation (10%), and test (20%) sets. RESULTS: In total, 810 lead SNPs along with two clinical variables including age and use of androgen deprivation therapy were incorporated into the deep learning model. The final model applied to the test data yielded an area under the curve (AUC) of 0.75. Calibration analysis demonstrated no statistical difference between observed and predicted incidence rates in six predicted risk bins (p = 0.9531), indicating good calibration agreement. The odds ratio between the highest 1/3 and lowest 1/3 risk groups was 11.8 (p = 0.0002), demonstrating robust predictive power. A post-modeling bioinformatic analysis revealed biological pathways driving the prediction of RIED, including neurophysiological processes, reproduction-gonadotropin regulation, and blood vessel morphogenesis. CONCLUSION: The BioDeepGWAS model provided a high level of RIED risk discrimination and identified related molecular and physiological pathways. These strong results support the model's potential for risk stratification in clinical settings.
