Abstract
In the Prostate Cancer Prevention Trial (PCPT), genotypes that may modify the effect of finasteride on the risk of prostate cancer have not been identified. Germline genetic data from 1,157 prostate cancer cases in PCPT were analyzed by case-only methods. Genotypes included 357 SNPs from 83 candidate genes in androgen metabolism, inflammation, circadian rhythm, and other pathways. Univariate case-only analysis was conducted to evaluate whether individual SNPs modified the finasteride effect on the risk of high-grade and low-grade prostate cancer. Case-only classification trees and random forests, which are powerful machine learning methods with resampling-based controls for model complexity, were employed to identify a predictive signature for genotype-specific treatment effects. Accounting for multiple testing, a single SNP in SRD5A1 gene (rs472402) significantly modified the finasteride effect on high-grade prostate cancer (Gleason score > 6) in PCPT (family-wise error rate < 0.05). Men carrying GG genotype at this locus had a 55% reduction of the risk in developing high-grade cancer when assigned to finasteride (RR = 0.45; 95% confidence interval, 0.27-0.75). Additional effect-modifying SNPs with moderate statistical significance were identified by case-only trees and random forests. A prediction model built by the case-only random forest method with 28 selected SNPs classified 37% of PCPT men to have reduced risk of high-grade prostate cancer when taking finasteride, while the others have increased risk. In conclusion, case-only methods identified SNPs that modified the effect of finasteride on the risk of high-grade prostate cancer and predicted a subgroup of men who had reduced cancer risk by finasteride.