Machine learning–based transcriptomic signatures to predict treatment outcomes across targeted and immunotherapy regimens in renal cell carcinoma.

4526 Background: Despite available tyrosine kinase inhibitors (TKIs) and immune checkpoint inhibitors (ICIs), reliable biomarkers guiding frontline advanced RCC treatment remain limited. Existing signatures lack generalizability across therapeutic regimens. We developed a data-driven machine learning (ML) framework to predict survival outcomes and therapeutic response. Methods: Transcriptomic and clinical data were analyzed from 733 patients across two frontline treatment cohorts, sunitinib (n = 376) and avelumab plus axitinib (n = 357), derived from JAVELIN Renal 101. A multi-algorithm feature selection framework was applied to identify transcriptomic signatures associated with progression-free survival (PFS) and overall survival (OS). Prognostic performance was evaluated using the concordance index (C-index). Predictive models for therapeutic response, including disease control, were developed using PFS-derived gene signatures and assessed by area under the curve (AUC). External validation was performed in an independent cohort from The Ohio State University Total Cancer Care (OSU TCC) (n = 114). Results: ML-derived transcriptomic models consistently stratified patients into distinct risk groups with improved prognostic discrimination compared with standard clinical classifiers. In the sunitinib cohort, the best-performing models achieved C-indices of 0.72 for PFS and 0.81 for OS, outperforming IMDC (0.59 and 0.66). In the validation set of the sunitinib cohort, high-risk patients exhibited worse outcomes, with hazard ratios of 3.00 for PFS (P < 0.001, 95% CI, 2.06–4.39) and 13.42 for OS (P < 0.001, 95% CI, 7.78–23.13). In the avelumab plus axitinib cohort, C-indices reached 0.70 for PFS and 0.79 for OS. Consistent risk stratification was observed in the validation set, with hazard ratios of 3.16 for PFS (P < 0.001, 95% CI, 2.07–4.83) and 4.69 for OS (P < 0.001, 95% CI, 2.65–8.30). For response prediction, the models demonstrated predictive performance, with the Naive Bayes model achieving a validation AUC of 0.83 for disease control in both sunitinib and avelumab plus axitinib cohorts. The model showed significant risk stratification in an external validation cohort (OSU TCC). Conclusions: This study presents a multi-cohort transcriptomic framework with prognostic and predictive utility in advanced RCC. By outperforming established clinical risk classifiers and enabling prediction of regimen-specific therapeutic responses, this ML-based approach supports biomarker-informed frontline treatment selection.