In silico phase III clinical trial of avelumab plus axitinib versus sunitinib in advanced renal cell carcinoma using a machine learning model transfer approach.

Abstract

e16500 Background: Randomized clinical trials (RCTs) are the gold standard for establishing treatment benefit, but are resource-intensive. An in silico virtual trial that enables exploration of treatment effects from available data could complement this process. We aimed to demonstrate the feasibility of an in-silico clinical trial using a machine learning (ML) model transfer approach, leveraging RCT data to generate counterfactual outcomes and assess whether such a framework reproduces observed RCT behavior in real-world TCGA data. Methods: ML models were trained on the JAVELIN Renal 101 RCT using RNA-seq data and internally validated for overall-survival (OS) and progression-free survival (PFS) within each arm (Avelumab+Axitinib (AA) and Sunitinib). They were then applied to the opposite arm to generate counterfactual survival predictions under alternative therapy. Treatment effects were summarized using Kaplan-Meier analysis and restricted mean survival times (RMST; τ = 24 months) based on available follow-up data. Results: In the available JAVELIN Renal 101 RCT data, AA demonstrated a significant PFS benefit compared with Sunitinib (median PFS 13.6 vs 8.1 months; ΔRMST 2.68 months [95% CI: 1.24-4.03], p = 0.001), but no OS benefit (p = 0.11). In our counterfactual analysis, PFS preserved a robust treatment effect favoring AA (median PFS 12.4 vs 8.3 months; ΔRMST 2.29 [95% CI: 0.72-3.65] months, p = 0.003), while OS did not (p = 0.45) (Table 1). Application to the TCGA cohort revealed similar results. Conclusions: This work demonstrates how the model transfer approach can be used to estimate counterfactual treatment outcomes and support a more informed transition from single-arm phase II to randomized phase III development. By enabling pre-testing of new drugs against established standards within an in-silico setting, this approach provides a practical means to evaluate expected benefit and improve patient selection before launching RCTs. Moreover, this framework would allow for testing drugs in rare diseases where a randomized study is not feasible. Landmark survival probabilities. Outcome Time (months) AA Observed Sunitinib Observed AA Counterfactual Sunitinib Counterfactual PFS 6 68.7% (63.9-73.9) 55.8% (50.8-61.4) 68.4% (63-73.1) 57.6% (51.7-63.1) PFS 12 53.5% (48.3-59.2) 39.5% (34.4-45.3 52.0% (46-57) 40.3% (34.4-45.7) PFS 18 43.7% (38.3-49.9) 29.1% (23.9-35.3) 41.5% (35.4-46.4) 27.8% (21.6-33.3) PFS 24 33.4% (26.1-42.7) 29.1% (23.9-35.3) 29.8% (21.1-37.2) 27.8% (21.6-33.3) OS 6 95.7% (93.6-97.9) 92.4% (89.7-95.1) 95.7% (93.3-97.8) 93.6% (91.2-96) OS 12 87.3% (83.9-90.9) 83.5% (79.7-87.4) 86.9% (82.9-90.2) 84.4% (80.2-88.2) OS 18 78.9% (74.5-83.6) 74.8% (70.2-79.7) 77.3% (71.6-81.1) 75.1% (69.3-79.4) OS 24 70.0% (64-76.6) 65.9% (59.9-72.5) 67.5% (59.1-73) 65.7% (58.1-70.7) AA: Avelumab+Axitinib.