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

Androgen deprivation therapy (ADT) is a backbone therapy for advanced prostate cancer, but prognostic estimation using clinical parameters such as PSA, Gleason score, and TNM is suboptimal (often C-index < 0.7). With the emergence of intensified regimens (radiation, docetaxel, and androgen receptor signaling inhibitors, including triplet therapy) for metastatic disease, improved prediction of ADT response is clinically important to guide treatment intensity. Genetic background likely influences ADT outcomes, with differences observed across ethnicities and within families. Prior GWAS by the authors identified two SNPs associated with prognosis during ADT in Japanese patients, but predictive performance was limited. This study aimed to integrate genetic (SNP) and clinical data using machine learning to predict castration resistance at 2 years in patients receiving primary ADT for advanced prostate cancer.

Previous risk models based on clinical variables have shown modest predictive power for ADT outcomes, with AUCs typically less than 0.7. Ethnic differences in ADT outcomes have been documented, with Asians often experiencing better survival than Caucasians and African Americans. The authors’ prior GWAS in Japanese patients undergoing primary ADT identified SNPs rs76237622 (PRR27) and rs117573572 (MTAP) associated with prognosis, but with insufficient predictive capability. GWAS can identify associations between SNPs and traits, but translating these into accurate predictive models is challenging due to high dimensionality; machine learning approaches can help model complex genomic data and potentially improve prediction.

Design and cohorts: Data were drawn from KYUCOG-1401-A (UMIN000022852), linked to the prospective multi-institutional KYUCOG-1401 clinical trial (UMIN000014243, JRCTs071180035). Japanese patients with de novo advanced prostate cancer (TanyN1M0 or TanyNanyM1) receiving primary ADT (GnRH antagonist degarelix or GnRH agonist leuprorelin/goserelin plus bicalutamide) were eligible. The study followed the Declaration of Helsinki and Japanese Ethical Guidelines; informed consent obtained; Kyushu University review board approval (23087-00). Patients censored before 2 years (n=8) were excluded. A total of 119 patients were included and randomly split 7:3 into discovery (n=82) and validation (n=37) cohorts. Clinical endpoints and definitions: Progression was defined as PSA progression (PSA ≥2.0 ng/mL, ≥50% rise from nadir, and three consecutive increases at least one week apart) or radiographic progression. For survival analyses: PFS event = progression or death; CSS event = death from prostate cancer; OS event = death from any cause; others censored at last follow-up. For 2-year castration resistance prediction, patients who progressed to castration resistance within 2 years were non-responders; those who did not were responders. J-CAPRA risk stratification was used for comparison. Genotyping: Genomic DNA was genotyped using the Japonica Array v2 (Axiom, Thermo Fisher Scientific), optimized for the Japanese genome. Genotype calling used Genotyping Console v4.2. SNP sets were selected from prior GWAS: 2 SNPs associated with PSA-PFS at 2 years at p<1.0×10⁻⁵, and 46 SNPs at p<1.0×10⁻⁴. Preprocessing and features: Variables were binary (1/−1) or quantitative. Quantitative variables were standardized (mean subtraction, division by standard deviation); missing values were set to 0. Three datasets were built: (1) clinical only; (2) clinical + 2 SNPs; (3) clinical + 46 SNPs. Feature importance was assessed via model-specific weights. Machine learning models: Three algorithms were trained: (a) Point-wise linear (PWL) algorithm, a deep learning-based approach that outputs a custom logistic regression model (sample-wise weight vector) per sample, implemented with deep unified networks (PyTorch 1.5.1, Python 3.7.4). (b) Logistic regression with elastic-net regularization (LR) (scikit-learn v0.24.2). (c) eXtreme Gradient Boosting (XGBoost) (xgboost 1.0.2). Hyperparameters were tuned via 5-fold cross-validation using the discovery cohort; best models were evaluated on the validation cohort. Performance metrics included AUC (for 2-year castration resistance), sensitivity, specificity; and Harrell’s C-index for PFS, CSS, and OS. Ethnic effect estimation: Minor allele frequencies (MAF) from the 1000 Genomes Project were used to estimate aggregate SNP effects across populations. Estimated effect per SNP = coefficient × 2×MAF×(1−MAF) + 2×coefficient×(MAF)²; summed over SNPs. Statistics: Continuous and categorical data reported as median (IQR) and n (%), respectively. Associations among categorical variables used chi-square. Survival analyses used Kaplan-Meier and log-rank. C-indices computed in Stata v18. Two-sided P<0.05 considered significant.

Cohorts: 119 patients were analyzed (discovery n=82; validation n=37). Several clinical parameters (e.g., Gleason score, extent of disease grade, PSA, hemoglobin) differed between responders and non-responders in the discovery cohort. Model performance (Table 1): - Clinical data only: Discovery AUCs 0.710–0.785; validation AUCs 0.720–0.786 across algorithms. PWL validation performance: AUC 0.786, sensitivity 0.812, specificity 0.684. - Clinical + 2 SNPs (p<1.0×10⁻⁵): Discovery AUCs 0.796–0.810; validation AUCs 0.701–0.878. PWL validation: AUC 0.878, sensitivity 0.750, specificity 0.789. - Clinical + 46 SNPs (p<1.0×10⁻⁴): Discovery AUCs 0.962–0.988; validation AUCs 0.984–1.000. PWL discovery/validation: AUC 0.988/1.000, sensitivity 0.864/0.938, specificity 0.978/1.000. Overall, PWL most often achieved the highest AUCs. Final PWL prediction models and AUCs (figure-based models using fixed formulas): - Clinical model: AUC 0.730 (95% CI 0.610–0.849) discovery; 0.585 (95% CI 0.383–0.787) validation. Top features included hypertension, total testosterone, total cholesterol, lymphocyte %, Gleason score, albumin, BUN, N-category, age, dyslipidemia, AST, PSA. - Small SNPs model (clinical + 2 SNPs): AUC 0.857 (95% CI 0.756–0.959) discovery; 0.852 (95% CI 0.706–0.998) validation. Important factors: diabetes mellitus, Gleason score, rs11231949 (LOC101928443), extent of disease grade, rs2035081 (PRIM1), CK, total P1NP, lymphocyte %. - Large SNPs model (clinical + 46 SNPs): AUC 0.920 (95% CI 0.854–0.986) discovery; 0.978 (95% CI 0.932–1.000) validation. Important features: rs12979986 (ZNF702P), rs9625031 (SRRD), M-category, rs1931229 (VGLL2), rs1660281, rs10064620, Gleason score, rs10860210 (RMST), rs941207 (BAZ2A), rs9868579 (RPN1), rs62174680, rs11232056 (LOC101928443), rs8124833 (C20orf78), rs74522810, rs74369678 (LRRN1), chr3:74213366, rs2035081 (PRIM1), total bilirubin, glucose, rs79404120, rs11672661 (FAM19A5), rs28625772 (COL5A3), rs9298681. Prognosis stratification and C-indices: - PFS significantly stratified by all three models; C-index: clinical 0.617 (95% CI 0.556–0.678), small SNPs 0.727 (95% CI 0.681–0.774), large SNPs 0.730 (95% CI 0.667–0.793). - CSS significantly stratified by all; C-index: clinical 0.678 (95% CI 0.546–0.809), small SNPs 0.670 (95% CI 0.551–0.790), large SNPs 0.781 (95% CI 0.671–0.890). - OS significantly stratified only by large SNPs model; C-index: clinical 0.636 (95% CI 0.520–0.753), small SNPs 0.621 (95% CI 0.512–0.731), large SNPs 0.703 (95% CI 0.583–0.822). J-CAPRA stratified PFS but not CSS/OS (C-indices ≤0.602). Ethnic allele frequency effects: - Among 16 key SNPs with available MAFs, the estimated aggregate effect (higher value indicates higher responder probability) was 0.94 in East Asians and −1.20 in Europeans, consistent with better ADT outcomes reported in Asian populations. Biological insights: - SNPs linked to androgen metabolism pathways were implicated: rs1931229 associated with TSPYL1 (regulates CYP17A1/CYP3A4) and rs941207/rs2035081 associated with HSD17B6 expression. Metabolic comorbidities showed associations: hypertension and higher total cholesterol with favorable response; diabetes and higher glucose with unfavorable response.

Integrating SNPs with clinical variables via machine learning markedly improved prediction of 2-year castration resistance compared with clinical variables alone, directly addressing the need for better prognostic tools in primary ADT. The large SNPs model achieved high AUCs and C-indices exceeding 0.70 for PFS, CSS, and OS, surpassing traditional clinical risk models (e.g., J-CAPRA). These models can inform treatment selection, potentially guiding intensification for predicted non-responders and de-escalation for predicted responders. The allele frequency analysis suggests genetic contributions to observed ethnic differences in ADT outcomes, with East Asian populations showing a higher predicted responder probability. The identification of SNPs tied to androgen synthesis and metabolism (e.g., TSPYL1, HSD17B6 pathways) provides biological plausibility for the predictive signals. Additionally, clinical and metabolic factors (hypertension, diabetes, lipid parameters) correlated with response, aligning with prior observations and underscoring the interplay between systemic metabolism and ADT efficacy.

SNP-augmented machine learning models, particularly those built with the PWL algorithm, substantially improved prediction of castration resistance at 2 years and stratified prognoses (PFS, CSS, OS) in men receiving primary ADT for advanced prostate cancer. Measuring a focused set of SNPs alongside select clinical features can achieve clinically meaningful predictive performance and may aid individualized treatment decisions. Future work should validate these models in larger, multi-ethnic cohorts and in contemporary treatment settings that include combination therapies with androgen receptor pathway inhibitors or chemotherapy.

- Sample size was relatively small (n=119), limiting statistical power and increasing the risk of overfitting, particularly for high-dimensional SNP models. - Generalizability is uncertain; the cohort comprised Japanese patients, and performance in other populations requires external validation. - Primary ADT alone is no longer the universal standard; current practice often includes combination therapies, so applicability to modern regimens needs assessment. - Some SNP frequency data were unavailable for certain variants, and estimated ethnic effects were based on aggregate calculations using public MAFs rather than direct validation. - Differences between cross-validated metrics and fixed-formula model performance suggest potential model optimism; independent validation cohorts would strengthen conclusions.