Androgen deprivation therapy (ADT) -- treatment that lowers testosterone to suppress prostate cancer growth -- is the standard backbone of care for advanced prostate cancer. However, most patients eventually stop responding as their tumors adapt to grow despite low testosterone levels. This progression is called castration-resistant prostate cancer (CRPC), and it marks a critical turning point where more aggressive treatment is required.
Modern prostate cancer treatment has become considerably more complex. In addition to ADT, patients may receive radiation, docetaxel chemotherapy, and a class of drugs called novel androgen receptor signaling inhibitors -- including abiraterone, enzalutamide, darolutamide, and apalutamide. More recently, triplet combination therapies combining these approaches have been shown to extend survival in metastatic disease. Knowing in advance which patients are most likely to develop castration resistance could help oncologists decide upfront whether a patient needs intensive combination therapy or might be safely managed with less aggressive treatment.
Current tools for predicting ADT response rely on clinical measurements such as PSA level, Gleason score, and TNM staging, but these produce only modest prediction accuracy -- generating C-index values below 0.7, a commonly used threshold for clinical usefulness in survival prediction. C-index, also known as the concordance index, ranges from 0.5 (no better than chance) to 1.0 (perfect prediction), and prior clinical-only models have consistently fallen short of the 0.7 bar.
Single nucleotide polymorphisms (SNPs) -- tiny inherited variations in the DNA sequence at specific positions in the genome -- have emerged as a promising additional source of predictive information. Differences in ADT response between ethnic groups, and the clustering of outcomes within families, suggest that genetic background shapes how patients respond to hormone deprivation therapy. This study tests whether integrating SNP data with clinical variables using machine learning can break through the performance ceiling of clinical-only prediction models.
Data came from the KYUCOG-1401-A study, a companion genetic study to a prospective multi-institutional randomized clinical trial conducted at multiple hospitals in the Kyushu region of Japan. The 119 patients included all had de novo advanced prostate cancer -- meaning their cancer had already spread at the time of diagnosis -- and all were treated with primary ADT. Patients were classified as non-responders if they progressed to CRPC within two years, and as responders if they did not.
Genetic data were obtained by genotyping each patient's DNA using the Japonica Array, a SNP microarray platform customized for the Japanese genome. From a prior genome-wide association study (GWAS) of the same cohort, the researchers identified SNPs associated with PSA-based progression at two years at two significance thresholds: 2 highly significant SNPs (p less than 1.0 x 10-5) and 46 moderately significant SNPs (p less than 1.0 x 10-4). These SNP sets were used as genetic inputs alongside clinical variables.
Three machine learning algorithms were compared: the point-wise linear (PWL) algorithm, a deep learning-based approach that generates a customized logistic regression model for each individual patient rather than a single universal model; logistic regression with elastic-net regularization (LR), which applies statistical penalties to prevent overfitting; and eXtreme Gradient Boosting (XGBoost), a widely used ensemble method known for high performance on tabular data. All models were trained on a discovery cohort (82 patients) and evaluated on an independent validation cohort (37 patients).
The 119 patients were divided randomly in a 7:3 ratio into discovery and validation cohorts. Importantly, the final predictive scores were simplified into a formula using the most important features identified by each model, enabling the creation of interpretable simple prediction scores that could realistically be applied in clinical practice without requiring the full machine learning infrastructure.
Using clinical data alone, the three ML algorithms produced AUC values for predicting castration resistance at two years of 0.720 to 0.786 in the validation cohort -- modest improvement over the clinical tools described in prior literature but still limited. Adding the two highly significant SNPs (p less than 1.0 x 10-5) improved validation AUC to 0.701 to 0.878 across the three algorithms. Adding all 46 moderately significant SNPs produced a dramatic leap: validation AUC values reached 0.984 to 1.000 -- near-perfect discrimination between patients who would and would not develop castration resistance within two years.
The PWL algorithm consistently achieved the highest AUC among the three methods when genetic data were included. Using the large SNPs model, the PWL algorithm achieved a validation AUC of 1.000 with a sensitivity of 93.8% and a specificity of 100% -- meaning it correctly identified every non-responder in the validation cohort and misclassified no responders as non-responders. These values are exceptionally high for a clinical prediction problem of this complexity.
When patients were divided into four equal-sized groups (Q1 through Q4) by their predicted risk scores, the large SNPs model correctly classified 90.5% of Q1 patients as non-responders and 95% of Q4 patients as responders, with statistical significance (p less than 0.0001). The small SNPs model performed similarly impressively, achieving 96.4% correct classification in both the lowest and highest risk groups. In contrast, the clinical-only model showed much weaker separation, correctly classifying 77.8% in Q1 and 59.3% in Q4.
Beyond predicting castration resistance, the researchers evaluated whether the three models could stratify survival outcomes. All three models significantly separated progression-free survival (PFS) curves among the four risk groups, but the genetic models produced sharper separation than the clinical-only model. The C-index for PFS in the clinical, small SNPs, and large SNPs models were 0.617, 0.727, and 0.730, respectively -- the genetic models exceeding the clinically meaningful 0.7 threshold that the clinical model missed.
For cancer-specific survival (CSS), only the large SNPs model produced statistically significant stratification. Its C-index was 0.781, compared to 0.678 for the clinical model and 0.670 for the small SNPs model. This result is clinically important: cancer-specific survival is considered a more meaningful endpoint than progression-free survival because it captures deaths directly attributable to prostate cancer rather than other causes.
For overall survival (OS), only the large SNPs model significantly stratified survival curves (p = 0.0079), achieving a C-index of 0.703 compared to 0.636 and 0.621 for the clinical and small SNPs models. In comparison, the well-established clinical risk tool J-CAPRA (which uses TNM stage, Gleason score, and PSA) stratified PFS but not CSS or OS, with C-indices of 0.588, 0.602, and 0.528 for the three endpoints -- all below 0.7 and consistently outperformed by the genetic models.
The machine learning models identified both expected and unexpected clinical predictors of castration resistance. Known factors such as Gleason score, PSA level, N-category, and extent of disease appeared as important features across models -- consistent with established clinical knowledge. However, the models also highlighted less conventional factors including hypertension, total cholesterol level, lymphocyte percentage, blood urea nitrogen, dyslipidemia, albumin, and liver enzyme levels.
Among these unexpected findings, comorbidity with hypertension was associated with better ADT response, while diabetes mellitus and high blood glucose were associated with worse response. Higher total cholesterol predicted better response, but the presence of dyslipidemia predicted worse response -- a paradox that implies the relationship between lipid metabolism and androgen sensitivity is complex and not captured by a single biomarker. These associations are consistent with previously reported findings from the same research group and suggest metabolic status interacts with the androgen signaling system in important ways.
Among the 19 SNPs in the large SNPs model, several mapped to genes with direct biological relevance to androgen metabolism. The SNP rs1931229 was linked by expression quantitative trait loci (eQTL) analysis to the gene TSPYL1, which regulates CYP17A1 and CYP3A4 -- enzymes involved in androgen biosynthesis and metabolism. Two other SNPs were linked to HSD17B6, an enzyme that converts androgen precursors. Genetic variation in these pathways may influence how effectively ADT suppresses biologically active androgens, offering a mechanistic explanation for the observed predictive associations.
The allele frequencies of 16 key SNPs from the large SNPs model differed substantially across ethnic groups. The estimated combined effect of these SNPs produced a score of 0.94 in East Asians -- indicating a higher predicted probability of responding to ADT -- compared to -1.20 in Europeans. This finding provides a genetic basis for the observed clinical observation that Japanese patients have higher survival rates after primary ADT than Caucasian or African American patients, and suggests that genetic ancestry should be considered when applying such predictive models across populations.
The central clinical implication of this study is that measuring 19 SNPs alongside four clinical parameters can predict with very high accuracy which patients will develop castration resistance within two years of starting ADT. Patients identified by the large SNPs model as likely non-responders might be prioritized for intensive upfront combination therapy -- adding novel androgen receptor inhibitors or docetaxel to standard ADT -- rather than waiting for castration resistance to develop before escalating treatment.
Conversely, patients predicted to be long-term responders might be candidates for de-escalated treatment approaches, potentially reducing the side effects and costs associated with combination therapies that may not be necessary for every patient. As triplet therapy becomes a new standard option for metastatic prostate cancer, having a tool to distinguish patients who truly need that level of treatment intensity from those who would do equally well with simpler therapy has substantial clinical value.
The study has important limitations. The sample size was relatively small at 119 patients, all from a single ethnic group (Japanese). While the SNP model showed remarkable performance, very high AUC values in small validation cohorts are subject to uncertainty -- they could reflect overfitting or the specific characteristics of this particular cohort rather than true generalizability. Prospective validation in larger, more diverse, and multi-ethnic cohorts is essential before clinical adoption.
Additionally, the study used primary ADT alone, which is no longer the standard of care -- most patients now receive ADT combined with other agents. Whether these SNP models retain predictive value in the context of combination therapies is unknown and an important question for future research. Despite these limitations, the work demonstrates that integrating genomic data with clinical information through machine learning can overcome the performance ceiling of clinical prediction alone in advanced prostate cancer management.