Early favorable prostate-specific antigen response prediction in metastatic hormone sensitive prostate cancer Authors: Soumyajit Roy, Yilun Sun, Maha Hussain, Kim N. Chi, Karim Fizazi, Ian D. Davis, Susan Halabi, Neeraj Agarwal, Simon Chowdhury, Bertrand Tombal, Scott C. Morgan, Shawn Malone, Pedro C. Barata, Michael Ong, Christopher J. D. Wallis, Alejandro Berlin, Umang Swami, Amar U. Kishan, Angela Y. Jia, Nicholas G. Zaorsky, Jorge A. Garcia, Prateek Mendiratta, Jason R. Brown, Vinod V. Subhash, Martin R. Stockler, Hayley Thomas, Rana R. McKay, Eric J. Small, Neal D. Shore, Fred Saad, Christopher J. Sweeney, Daniel E. Spratt Published: 2025 ABSTRACT There is an unmet need for a tool that could predict early favorable prostate-specific antigen (PSA) response in metastatic hormone sensitive prostate cancer (mHSPC) patients receiving androgen receptor pathway inhibitor (ARPI). Here, we train and validate a multivariable logistic regression model to predict early favorable PSA response (≤0.2 ng/mL by 6 months) in these patients. Patients randomly allocated to the ARPI arms of the LATITUDE (abiraterone), TITAN (apalutamide), and ARASENS (darolutamide) trials, are split 60:40 into training (n = 1030) and internal validation (n = 688) cohorts. The locked model is validated in an independent external validation cohort - the enzalutamide arm of the ENZAMET trial (n = 540). The area under curve and Brier score for the locked model in the external validation cohort are 0.82 (95% confidence interval [CI] = 0.78–0.85) and 0.16, respectively. Stratification by predicted probability tertiles show PSA response rates of 92% (95% CI = 88–96), 74% (95% CI = 68–81), and 39% (95% CI = 32–47), respectively. Pending prospective validation, our model predicts early favorable PSA response supporting its potential role in guiding treatment decisions. Introduction Early PSA response is a well-established prognostic marker in metastatic hormone sensitive prostate cancer (mHSPC) patients treated with androgen deprivation therapy (ADT) alone or a combination of ADT plus androgen receptor pathway inhibitors (ARPI) with or without docetaxel1–7. Achieving a PSA response of ≤0.2 ng/mL within 6 months of treatment initiation predicts improved overall survival (OS) and mediates the treatment effect on subsequent mortality risk8. Conversely, failure to reach this response threshold portends inferior OS. Therefore, this clinical state may warrant an earlier treatment intensification, or a clinical trial enrollment. However, waiting until 6 months to assess PSA nadir may inadvertently delay timely intervention, allowing disease progression and evolution to a more treatment-resistant state, potentially diminishing the benefit of subsequent therapeutic intensification. Predicting PSA responses using pre-treatment variables, especially in ARPI-treated patients, could enable earlier, personalized treatment decisions before the optimal intervention window closes. Conventional risk stratification methods in mHSPC, such as volume of metastatic disease based on conventional imaging9,10, have limited accuracy in predicting early PSA response. For example, one study demonstrated that over half of patients with favorable PSA responses had high-volume metastatic disease, while nearly 1/3 of those with suboptimal PSA response had low-volume disease8. These findings underscore the limitations of relying on a single variable for risk stratification and treatment optimization in mHSPC. Therefore, there is an unmet need of an easily deployable clinical tool, which could predict favorable early PSA response. While such tool will not be a replacement for formal PSA monitoring, it will provide baseline prognostic insights that can enhance early risk stratification, guide treatment planning, and support enrollment in biomarker-driven studies. In this work, we developed and validated the Nadir ARPI-Derived Integrative Response (NADIR) model using individual patient data from four phase III trials (ARASENS, ENZAMET, LATITUDE, and TITAN)11–14. This tool leverages pretreatment clinical and laboratory parameters to estimate the likelihood of achieving a PSA ≤ 0.2 ng/mL by 6 months of initiating ARPI-based therapy. Results Baseline characteristics Overall, 1718 male patients were treated with ARPI agents in the ARASENS, LATITUDE, and TITAN trials, of whom the training cohort consisted of 1030 patients, while the internal validation cohort consisted of 688 patients. The external validation cohort consisted of 540 male patients from the ENZAMET trial (Fig. 1). Baseline characteristics have been summarized in Supplementary Table S1 (supplementary information). Overall, 54% (n = 551) in the training cohort, 56% (n = 383) in the internal validation cohort, and 69% (n = 371) in the external validation cohort experienced an early PSA response to ≤0.2 ng/mL by 6 months.Fig. 1The CONSORT diagram.The bold font texts highlight the composition of the training and two validation cohorts, respectively. The blue boxes specifically demonstrate the independent external validation cohort. The CONSORT diagram. The bold font texts highlight the composition of the training and two validation cohorts, respectively. The blue boxes specifically demonstrate the independent external validation cohort. Model training and validation After the NADIR model (https://bit.ly/42Ec1lP) was built utilizing the training cohort, it was locked and applied to the two validation cohorts. The model coefficients with 95% confidence intervals (CI) are summarized in Table 1. In the internal validation cohort, the AUC for the NADIR model was 0.76 (95% CI = 0.72–0.79) with a Brier score of 0.20. In the external validation cohort, the AUC was 0.82 (95% CI = 0.78–0.85) with a Brier score of 0.16. The receiver operating curves are shown in Fig. 2. The calibration plots showed alignment between the predicted probabilities and the actual outcomes in the training, internal and external validation cohorts (Supplementary Figs. S1, S2, S3, respectively, in the supplementary information). In the external validation cohort, the sensitivity, specificity, PPV, and NPV of the model were 0.83 (95% CI = 0.79–0.86), 0.62 (95% CI = 0.54–0.69), 0.83 (95% CI = 0.79–0.86), and 0.62 (95% CI = 0.54–0.69), respectively (Supplementary Table S2, supplementary information). On LOOCV, the model demonstrated consistent performance (AUC values ranging between 0.78 and 0.82; Brier scores ranging from 0.16 to 0.19) across trials (Supplementary Table S3, supplementary information). There was substantial heterogeneity in the distribution of predicted probability of achieving early PSA response based on metastatic volume by CHAARTED definition across the internal validation cohort (Supplementary Fig. S4, supplementary information) and external validation cohort (Supplementary Fig. S5, supplementary information), respectively. Removing baseline PSA from the multivariable model led to a substantial drop in predictive accuracy in external validation, confirming PSA as an important predictor but not a sufficient one (Table 2). Importantly, the full NADIR model outperformed PSA-only model across exploratory subgroups within the external validation cohort stratified by baseline PSA levels with greater benefits among patients with baseline PSA value higher than 5 ng/mL (Supplementary Table S4, supplementary information).Fig. 2Receiver operating curve for the logistic regression (NADIR) model to predict favorable early PSA response to ≤0.2 ng/mL by 6 months in the internal (n = 688) and external validation cohorts (n = 540).In the internal validation cohort, the AUC for the NADIR model was 0.76 (95% CI = 0.72–0.79). In the external validation cohort, the AUC was 0.82 (95% CI = 0.78–0.85).Table 1Logistic regression model coefficients (NADIR model) for predicting early PSA response by 6 monthsVariableLog odds ratio2.5% CI97.5% CIP-value (two-sided)(Intercept)−16.11−24.35−7.880.000125Volume (High versus Low)−0.4−0.77−0.040.031506Log of baseline BMI−0.29−1.230.650.547319Log of age0.53−0.621.680.364180Log of hemoglobin at baseline3.522.324.730.0000000097Metastatic stage at presentation (Synchronous vs metachronous)−0.21−0.720.310.428367Visceral metastasis (No vs Yes)−0.08−0.480.320.699544Restricted cubic spline of baseline PSASpline basis 1a−0.14−0.17−0.122e−16Spline basis 2a106.9886.24127.712e−16Spline basis 3a−122.48−146.23−98.742e−16This table displays the estimated log-odds ratios, 95% confidence intervals (CI), and two-sided p-values from a multivariable logistic regression model predicting the likelihood of achieving early PSA drop to ≤0.2 ng/mL by 6 months of ARPI initiation. Continuous variables such as baseline BMI, age, and hemoglobin were log-transformed. Baseline PSA was fitted using restricted cubic splines with 4 degrees of freedom. The respective levels for categorical variables are also shown. The reported P-values are two-sided and not adjusted for multiple comparisons. DOF was 1029 for null and 1020 residual.BMI body mass index, CI confidence intervals, PSA prostate-specific antigen.aCoefficients for spline terms may appear large due to the scaling and orthogonalization of spline basis functions. This is expected and does not indicate model misspecification. The model has been thoroughly checked, and predicted effects were stable and interpretable via plotting or marginal estimates.Table 2Variable importance in NADIR model summarized in terms of loss of index of prediction accuracy (IPA) in the validation cohortsInternal validation cohortExternal validation cohortIPAIPA dropIPAIPA dropFull model18.625.8Volume (CHAARTED definition)17.31.324.01.8Log of baseline BMI18.30.326.0-0.2Log of age18.50.125.50.3Log of baseline hemoglobin18.40.2241.8Metastatic stage18.50.125.30.5Visceral metastasis18.30.325.70.1Baseline PSAa8.610.07.118.7Higher IPA indicates better prediction accuracy. IPA drop for a variable is estimated as IPA of full model – IPA. The higher the drop, the more important the variable for the full model is.aFitted with restricted cubic splines with 4 degrees of freedom.IPA values are multiplied by 100 and given in percentages. Receiver operating curve for the logistic regression (NADIR) model to predict favorable early PSA response to ≤0.2 ng/mL by 6 months in the internal (n = 688) and external validation cohorts (n = 540). In the internal validation cohort, the AUC for the NADIR model was 0.76 (95% CI = 0.72–0.79). In the external validation cohort, the AUC was 0.82 (95% CI = 0.78–0.85). Table 1Logistic regression model coefficients (NADIR model) for predicting early PSA response by 6 monthsVariableLog odds ratio2.5% CI97.5% CIP-value (two-sided)(Intercept)−16.11−24.35−7.880.000125Volume (High versus Low)−0.4−0.77−0.040.031506Log of baseline BMI−0.29−1.230.650.547319Log of age0.53−0.621.680.364180Log of hemoglobin at baseline3.522.324.730.0000000097Metastatic stage at presentation (Synchronous vs metachronous)−0.21−0.720.310.428367Visceral metastasis (No vs Yes)−0.08−0.480.320.699544Restricted cubic spline of baseline PSASpline basis 1a−0.14−0.17−0.122e−16Spline basis 2a106.9886.24127.712e−16Spline basis 3a−122.48−146.23−98.742e−16This table displays the estimated log-odds ratios, 95% confidence intervals (CI), and two-sided p-values from a multivariable logistic regression model predicting the likelihood of achieving early PSA drop to ≤0.2 ng/mL by 6 months of ARPI initiation. Continuous variables such as baseline BMI, age, and hemoglobin were log-transformed. Baseline PSA was fitted using restricted cubic splines with 4 degrees of freedom. The respective levels for categorical variables are also shown. The reported P-values are two-sided and not adjusted for multiple comparisons. DOF was 1029 for null and 1020 residual.BMI body mass index, CI confidence intervals, PSA prostate-specific antigen.aCoefficients for spline terms may appear large due to the scaling and orthogonalization of spline basis functions. This is expected and does not indicate model misspecification. The model has been thoroughly checked, and predicted effects were stable and interpretable via plotting or marginal estimates. Logistic regression model coefficients (NADIR model) for predicting early PSA response by 6 months This table displays the estimated log-odds ratios, 95% confidence intervals (CI), and two-sided p-values from a multivariable logistic regression model predicting the likelihood of achieving early PSA drop to ≤0.2 ng/mL by 6 months of ARPI initiation. Continuous variables such as baseline BMI, age, and hemoglobin were log-transformed. Baseline PSA was fitted using restricted cubic splines with 4 degrees of freedom. The respective levels for categorical variables are also shown. The reported P-values are two-sided and not adjusted for multiple comparisons. DOF was 1029 for null and 1020 residual. BMI body mass index, CI confidence intervals, PSA prostate-specific antigen. aCoefficients for spline terms may appear large due to the scaling and orthogonalization of spline basis functions. This is expected and does not indicate model misspecification. The model has been thoroughly checked, and predicted effects were stable and interpretable via plotting or marginal estimates. Table 2Variable importance in NADIR model summarized in terms of loss of index of prediction accuracy (IPA) in the validation cohortsInternal validation cohortExternal validation cohortIPAIPA dropIPAIPA dropFull model18.625.8Volume (CHAARTED definition)17.31.324.01.8Log of baseline BMI18.30.326.0-0.2Log of age18.50.125.50.3Log of baseline hemoglobin18.40.2241.8Metastatic stage18.50.125.30.5Visceral metastasis18.30.325.70.1Baseline PSAa8.610.07.118.7Higher IPA indicates better prediction accuracy. IPA drop for a variable is estimated as IPA of full model – IPA. The higher the drop, the more important the variable for the full model is.aFitted with restricted cubic splines with 4 degrees of freedom.IPA values are multiplied by 100 and given in percentages. Variable importance in NADIR model summarized in terms of loss of index of prediction accuracy (IPA) in the validation cohorts Higher IPA indicates better prediction accuracy. IPA drop for a variable is estimated as IPA of full model – IPA. The higher the drop, the more important the variable for the full model is. aFitted with restricted cubic splines with 4 degrees of freedom. IPA values are multiplied by 100 and given in percentages. Risk stratification by NADIR model When stratified by decreasing order of tertiles of predicted probability of early PSA response, the proportions achieving a favorable early PSA response to ≤0.2 ng/mL in the upper (n = 229), middle (n = 229), and lower tertile (n = 230) were 83% (95% CI = 78–88), 52% (95% CI = 46–59), and 32% (95% CI = 26–38) in the internal validation cohort (Supplementary Fig. S6, supplementary information). Similarly, the proportions achieving a favorable early PSA response to ≤0.2 ng/mL in the upper (n = 180), middle (n = 180), and lower tertile (n = 180) were 92% (95% CI = 88–96), 74% (95% CI = 68–81), and 39% (95% CI = 32–47) in the external validation cohort (Supplementary Fig. S7, supplementary information). Quartile-based stratification demonstrated similar differences in the proportion of patients with early favorable PSA responses in the internal and external validation cohort, respectively (Supplementary Figs. S8 and S9, supplementary information). On decision curve analysis in the external validation cohort, the NADIR model was most useful for identifying patients at a range of risk that is clinically meaningful (threshold probabilities between 25% and 60%) for early PSA drop in patients treated with ARPI with a higher net benefit in this range compared to both baseline assumptions (no patient has early PSA response, and all patients have early PSA response) (Supplementary Fig. S10, supplementary information). We subsequently determined the OS estimates across the tertiles of predicted probability from the NADIR model using Kaplan–Meier methods. The 4-years OS probabilities in patients stratified by tertiles of predicted probability were 71% (95% CI = 65–79); 61% (95% CI = 55–69); and 47% (95% CI = 39–56), respectively in the internal validation cohort (n = 672) (Fig. 3) and 80% (95% CI = 75–86), 74% (95% CI = 68–81), and 56% (95% CI = 49–64), respectively in the external validation cohort (n = 540) (Fig. 4). On Cox proportional hazard regression modeling, compared to upper tertile, patients in the middle and lower tertile had inferior OS in the internal validation cohort (HR for middle tertile = 1.57 [95% CI = 1.11–2.21]; HR for lower tertile = 2.28 [95% CI = 1.65–3.16]) and in the external validation cohort (HR for middle tertile = 1.35 [95% CI = 0.92–1.69]; HR for lower tertile = 2.72 [95% CI = 1.91–3.88], respectively.Fig. 3Kaplan–Meier based overall survival curves with 4-year overall survival (OS) estimates and 95% confidence intervals (within parenthesis) across patient groups stratified by tertiles (upper, middle, and lower) of predicted probability of early PSA response to ≤0.2 ng/mL from the NADIR model in the internal validation cohort (n = 672).The tables below the risk show the numbers at risk and number of events respectively. The 4-years OS probabilities in the upper (n = 228), middle (n = 223), and lower (n = 221) tertile were 71% (95% CI = 65–79); 61% (95% CI = 55–69); and 47% (95% CI = 39–56), respectively.Fig. 4Kaplan–Meier based overall survival curves with 4-year overall survival (OS) estimates and 95% confidence intervals (within parenthesis) across patient groups stratified by tertiles (upper, middle, and lower) of predicted probability of early PSA response to ≤0.2 ng/mL from the NADIR model in the external validation cohort (n = 540).The tables below the risk show the numbers at risk and number of events respectively. The 4-years OS probabilities in the upper (n = 180), middle (n = 180), and lower (n = 180) tertile were 80% (95% CI = 75–86), 74% (95% CI = 68–81), and 56% (95% CI = 49–64), respectively. Kaplan–Meier based overall survival curves with 4-year overall survival (OS) estimates and 95% confidence intervals (within parenthesis) across patient groups stratified by tertiles (upper, middle, and lower) of predicted probability of early PSA response to ≤0.2 ng/mL from the NADIR model in the internal validation cohort (n = 672). The tables below the risk show the numbers at risk and number of events respectively. The 4-years OS probabilities in the upper (n = 228), middle (n = 223), and lower (n = 221) tertile were 71% (95% CI = 65–79); 61% (95% CI = 55–69); and 47% (95% CI = 39–56), respectively. Kaplan–Meier based overall survival curves with 4-year overall survival (OS) estimates and 95% confidence intervals (within parenthesis) across patient groups stratified by tertiles (upper, middle, and lower) of predicted probability of early PSA response to ≤0.2 ng/mL from the NADIR model in the external validation cohort (n = 540). The tables below the risk show the numbers at risk and number of events respectively. The 4-years OS probabilities in the upper (n = 180), middle (n = 180), and lower (n = 180) tertile were 80% (95% CI = 75–86), 74% (95% CI = 68–81), and 56% (95% CI = 49–64), respectively. Exploratory analysis of interaction with docetaxel effect In the combined validation cohort, we did not find any considerable evidence of difference in proportion of patients with early PSA response with and without use of docetaxel across the three risk score groups (Supplementary Table S5, supplementary information). We found a differential effect from docetaxel on OS across the predicted scores from the NADIR model (HR for the interaction term 0.41; 95% CI = 0.19–0.88; two-sided p-interaction=0.0215) (Supplementary Fig. S11, supplementary information). Kaplan–Meier method-based OS estimates with and without docetaxel use in these groups are shown in Supplementary Figs. S12 and S13, respectively (supplementary information). For patients in the lower tertile of predicted probability of early favorable PSA response (n = 401), 4-year OS estimates with and without docetaxel were 55% (95% CI = 48–63) and 53% (95% CI = 46–61), respectively while for patients in the upper to middle tertiles (n = 811), the 4-year OS estimates with and without docetaxel were 66% (95% CI = 60–72) and 72% (95% CI = 68–77), respectively (Supplementary Fig. S12, supplementary information). For patients in the upper-tertile of predicted probability of early favorable PSA response (n = 408), 4-year OS estimates with and without docetaxel were 72% (95% CI = 65–80) and 78% (95% CI = 72–83), respectively while for patients in the lower to middle tertiles (n = 804), 4-year OS estimates with and without docetaxel were 58% (95% CI = 52–63) and 60% (95% CI = 56–65), respectively (Supplementary Fig. S13, supplementary information). Similar evidence of differential treatment effect from on OS across the predicted scores from the NADIR model (HR for the interaction term 0.28; 95% CI = 0.09–0.92; two-sided p-interaction=0.03377) was found in the external validation cohort (n = 540) (Supplementary Fig. S14, supplementary information). Additional sensitivity analyses As an exploratory analysis, we applied the NADIR model on patients treated with ADT with/without docetaxel. In patients treated with ADT with/without docetaxel, the AUC of the NADIR model was 0.77 (95% CI = 0.75–0.79) with a Brier score of 0.24. On decision curve analysis, the NADIR model was useful for identifying patients at a range of risk that is clinically meaningful (threshold probabilities between 15% and 50%) for early PSA response to ≤0.2 ng/mL by 6 months (Supplementary Fig. S15, supplementary information). On sensitivity analysis, the random forest classifier model had an AUC and Brier score similar to the logistic regression model in the internal validation cohort (AUC: 0.75 [95% CI = 0.72–0.80]; Brier score: 0.20) and in the external validation cohort (AUC: 0.83 [95% CI = 0.79–0.86]; Brier score: 0.17), respectively. Additional metrics for the random forest classifier model are summarized in Supplementary Table S6 (supplementary information). The decision curve analysis showed similar net benefit as the logistic regression model when compared to both baseline assumptions (Supplementary Fig. S16, supplementary information). Discussion Using IPD from four randomized controlled trials data we have trained and validated a prognostic model that can predict early favorable PSA response to ≤0.2 ng/mL by 6 months of ARPI initiation. Given the similar performance, but easier interpretability and ease of deployment in routine clinical practice, we preferred the logistic regression model over the random forest model. The NADIR model was validated for a clinically relevant endpoint for mHSPC patients treated with an ARPI agent, now standard of care for patients with mHSPC15,16. As evident from the decision curve analysis, the model showed validation across a range of probabilities of early PSA response that is clinically relevant for mHSPC patients treated with an ARPI agent. Our model leverages readily available clinical and laboratory variables which are part of standard of care and have been shown to have prognostic relevance for mHSPC patients10,17–21. In patients with lower baseline PSA levels (≤5 ng/mL), the multivariable model performed similarly to, or slightly better than, PSA alone. This likely reflects the inherently higher probability of achieving an early nadir in this subgroup, reducing the clinical need for additional predictive modeling. Conversely, in patients with higher baseline PSA (>5 ng/mL) - where nadir prediction remains an unmet clinical need - the multivariable model demonstrated superior predictive performance compared to baseline PSA alone. An important aspect when designing prognostic models is the use of high-quality, ideally randomized trial data, and independent validation in more than one dataset. These features, among others, are required to achieve Simon Level 1 evidence for a prognostic tool22. Another strength of our model is the inclusion of all currently FDA-approved ARPI agents, as well as patients who received triplet therapy, to ensure broad applicability to clinical practice. While the model also validated with good discriminatory performance for patients treated with ADT with or without chemotherapy, these patients were intentionally not included in the training and primary validation given the standard of care has evolved to include a doublet of at least ADT with an ARPI for most patients with mHSPC. Nonetheless, the model still needs prospective validation. Our NADIR model bears relevance to the evolving landscape of risk stratification in mHSPC. While informative and better than the commonly used risk stratification tools like disease volume or metastatic stage at presentation, the NADIR model may not be sufficient to fully capture the complexity of early treatment response. Integrating molecular, genomic, or imaging biomarkers could enhance predictive accuracy. Although several genomic and multimodal artificial intelligence-based biomarkers have been found to be prognostic in mHSPC23–25, their clinical use in advanced disease remains limited. Our model may provide a foundation to measure the added value of such advanced tools in risk stratification and treatment personalization in mHSPC. While PSA monitoring is essential in mHSPC patients during ARPI therapy, it is plausible that delaying treatment intensification decisions until 6 months could potentially portend a risk of emerging therapeutic resistance. This may in turn reduce the potential benefit of treatment intensification at or beyond that time point26. Early identification of high-risk patients at diagnosis could support upfront intensification and improve outcomes - an approach supported by oncology trends favoring concurrent over sequential treatment regimens27–29. While this hypothesis requires prospective validation in mHSPC setting, its relevance is reflected in ongoing trials that use PSA nadir status as a criterion for treatment intensification or de-intensification (NCT06592924, NCT05241860). With prospective validation, the NADIR model may help enrich such clinical trials by identifying patients less likely to achieve early PSA response (<33% predicted probability), who also experience nearly 20% lower 4-year OS compared to those with ≥67% predicted probability - independent of disease volume or metastatic stage. In our study, we did not observe an OS benefit with addition of docetaxel in those with a relatively higher predicted probability of early PSA response to ≤0.2 ng/mL by 6 months. Given the post-hoc exploratory nature, potential for residual confounding, and the lack of a trial randomizing to ADT+ARPI +/- docetaxel, this finding is merely hypothesis-generating and is liable to both type I and type II errors. Furthermore, patients referred for docetaxel on trial might have other poor risk features not captured in the model which needs further exploration with additional biomarkers30. Our model has limitations. The model interpretation (including sensitivity, specificity, and predictive values) depends on the threshold being used. For example, at 0.5 threshold, the model identified 31% (169/540) of patients as predicted non-responders in the external validation cohort. Among these, 64 (38%) patients were falsely classified. If all predicted non-responders were to receive treatment intensification (e.g., with docetaxel), the model-based approach would result in overtreatment in ~12% (64/540) of patients out of the total cohort. While the multivariable NADIR model improved prediction over PSA alone, the absolute gain in predictive accuracy was modest across PSA strata, especially in those with pre-ARPI PSA ≤ 5 ng/mL; in such cases, additional modeling may offer limited incremental value. While the current treatment landscape has been shaped by all the clinical trials included in this study, ongoing advancements in the staging and treatment are not fully reflected in these trials. These include increasing use of positron emission tomography (PET) in practice to detect metastatic disease at an early state31,32. Additionally, the treatment of the primary was not mandatory in de novo low volume disease, which is now an integral part of the current standard of care33–36. Other common variables were not readily available for all trials or all patients, including baseline Gleason score, serum testosterone level, PSA level, and alkaline phosphatase level prior to and on ADT. Although we included laboratory variables before ARPI initiation, some of these variables might have been influenced by prior systemic treatment. Finally, the p-values for interaction between model prediction and docetaxel effect were not adjusted for multiple testing and should be interpreted as exploratory only. Therefore, these findings warrant further validation. To conclude, using readily available clinical and laboratory variables from four phase III randomized trials, we developed and validated a logistic regression model that predicts favorable early PSA response to ≤0.2 ng/mL by 6 months in mHSPC patients initiating treatment with an ARPI. Prospective validation is needed before its routine clinical application and integration with additional biomarkers to achieve meaningful predictive gains. Methods Study cohorts Our research complies with all relevant ethical regulations by the institutional review board of University Hospitals Seidman Cancer Center (STUDY20251115) and data-sharing agreements with trial sponsors. All patients provided informed consent before they were enrolled into the respective trials. A literature search was performed using Medline and Clinicaltrials.gov to identify phase III randomized controlled trials of patients solely with mHSPC randomized to ADT with or without an ARPI, irrespective of docetaxel use. Each trial group or industry partner was contacted to obtain IPD. Four trials met criteria for which IPD access was granted for analysis: ARASENS (NCT02799602), ENZAMET (NCT02446405), LATITUDE (NCT01715285), and TITAN (NCT02489318). For ARASENS, LATITUDE, and TITAN, data were accessed via a peer-reviewed data access request on the Vivli platform (Project ID: 009912), while data from the ENZAMET (NCT02446405) trial were accessed through the Australian and New Zealand Urogenital and Prostate Cancer Trials Group. This post-hoc analysis was conducted in accordance with a pre-specified analysis plan approved by all relevant parties. Patients eligible for this analysis were adult (≥18 years) patients with histologically or cytologically confirmed mHSPC, an Eastern Cooperative Oncology Group (ECOG) performance-status score of 0 or 1, and evidence of metastases on bone scans or cross-sectional imaging. All study participants were male, and sex was determined by clinical diagnosis. Patients were included if they were randomly assigned to and treated with ADT with or without docetaxel plus ARPI and had complete baseline and follow-up data for clinical outcomes, treatment assignments, and baseline (pre-ARPI) demographic and clinical characteristics, and laboratory variables such as age (in years), body mass index (BMI), metastatic disease volume (high versus low per CHAARTED criteria)9,10, visceral metastases (yes/no), baseline PSA (ng/mL), hemoglobin (g/L), docetaxel use (yes versus no), and metastatic stage at presentation (synchronous versus metachronous), respectively. All analyzes complied with the terms set by the data providers, and no identifiable patient information was used. Statistics and reproducibility The primary endpoint was early favorable PSA response, defined as a decline in serum PSA level to ≤0.2 ng/mL by 6 months of random allocation among patients treated with an ARPI. Patients from ARASENS, LATITUDE, and TITAN trials were pooled and randomly split into training (60%) and internal validation (40%) cohorts using unstratified random allocation. We used a 60:40 split to prioritize a larger training cohort for model development, ensuring stable coefficient estimation while retaining a sufficiently large internal validation cohort for performance testing. The training cohort was used to build the model and once locked, the model’s performance was evaluated in the internal and external validation cohorts. Model development utilized logistic regression, including baseline PSA fitted with restricted cubic splines with 4 degrees of freedom, log of baseline hemoglobin, baseline body mass index, and age at enrollment respectively, in addition to metastatic disease volume (low versus high by CHAARTED definition)9, visceral metastasis (yes versus no), and metastatic stage at presentation (synchronous versus metachronous). Variable selection was based on clinical relevance, with final model specification determined by the Akaike Information Criterion. Model discrimination was assessed using the area under the receiver operating characteristic curve (AUC) with 95% confidence intervals from 5000 bootstrap iterations. Calibration was evaluated by comparing bootstrapped estimates of early PSA nadir probabilities with 95% confidence intervals across tertiles of predicted probabilities in both internal and external validation cohorts. We selected tertiles to ensure adequate sample sizes and event counts within each subgroup. This avoided over-stratification, which could lead to unstable estimates and wide confidence intervals. Additional performance metrics included Brier score (which measures the accuracy of predicted probabilities with lower Brier score indicating better calibration), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and positive and negative likelihood ratio. No pre-specified probability threshold was selected to define clinical actionability; thresholds were explored post hoc to evaluate model performance across a range of values using decision curve analysis. Variable importance was estimated using loss of index of prediction accuracy (IPA). IPA is calculated by rescaling the Brier score. External validation employed individual patient data from the ARPI arm of the ENZAMET trial to assess the model’s generalizability. A post-hoc power analysis confirmed adequate sample size to detect an AUC of 0.75 with a significance level of 0.05. To determine the validation of our model, we performed a leave-one-trial-out-validation (LOOCV). Each trial was sequentially excluded and used as an external validation cohort while the model was trained on the remaining three trials. Discrimination was assessed using the area under the AUC with 95% confidence intervals. Calibration was evaluated using Brier scores. This approach allows for evaluation of model generalizability and reduces potential overfitting to any single dataset. As an additional sensitivity analysis, a random forest classifier was developed and evaluated using the same framework. Given the prognostic implications of early PSA response, overall survival (OS) was explored in a 6-month landmark population stratified by tertiles (upper, middle, and lower) of predicted probability from the NADIR model. Overall survival (OS) was reported at 4 years because follow-up beyond 4 years was limited in the training and validation cohorts. Cox proportional hazards models evaluated hazard of deaths across the tertiles, and the potential differential effect of docetaxel based on predicted probabilities from the NADIR model using an interaction term. Proportionality assumption was tested using Grambsch–Therneau test. No adjustment for multiple comparisons was made, and all reported p-values were 2-sided. All statistical analyzes were performed using R version 4.4.1 with its associated packages (tidymodels, tidyverse, pROC, riskRegression, survival, rms)37–40. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Supplementary Information Reporting Summary Transparent Peer Review file