Predictive model of castration resistance in advanced prostate cancer by machine learning using genetic and clinical data: KYUCOG-1401-A study

BACKGROUND: The predictive power of treatment efficacy and prognosis in primary androgen deprivation therapy (ADT) for advanced prostate cancer is not satisfactory. The objective was to integrate genetic and clinical data to predict castration resistance in primary ADT for advanced prostate cancer by machine learning (ML). METHODS: Clinical and single nucleotide polymorphisms (SNP) data from the KYUCOG-1401-A study (UMIN000022852) of Japanese patients with advanced prostate cancer treated with primary ADT were used. Three ML algorithms—point-wise linear (PWL), logistic regression with elastic-net regularization, and eXtreme Gradient Boosting—were evaluated. Area under the curve (AUC) for predicting castration resistance and C-index for prognoses were calculated. RESULTS: Among the algorithms, the PWL algorithm yielded the highest AUCs to predict castration resistance at 2 years across datasets. Three PWL models were created: a clinical model (clinical variables only) and two SNP-augmented models (clinical + 2 SNPs; clinical + 46 SNPs). C-indices for overall survival were 0.636 (clinical), 0.621 (small SNPs), and 0.703 (large SNPs). CONCLUSION: SNP-augmented ML models produced excellent prediction of castration resistance and prognosis in primary ADT for advanced prostate cancer and may aid treatment selection.

Masaki Shiota, Shota Nemoto, Ryo Ikegami, Shuichi Tatarano, Toshiyuki Kamoto, Keita Kobayashi, Hideki Sakai, Tsukasa Igawa, Tomomi Kamba, Naohiro Fujimoto, Akira Yokomizo, Seiji Naito, Masatoshi Eto