Abstract 6724: OncoTwin: A multimodal digital twin framework for predicting treatment response and guiding trial design in ALK-rearranged non-small-cell lung cancer.

Abstract Introduction: Despite advances with next-generation tyrosine kinase inhibitors (TKIs), response and progression patterns in ALK-rearranged NSCLC vary widely, and currently no reliable biomarkers can predict individualized outcomes under alternative therapies. Digital-twin models offer a solution by integrating real-world evidence to reconstruct patient-specific counterfactual disease trajectories for single-arm trials and evaluation of escalation strategies. Methods: We integrated real-world and clinical-trial datasets of ALK-rearranged NSCLC treated with TKIs, including MDACC cohort (n = 103, GEMINI database) for model development, Phase III randomized ALTA-1L (n = 207) for external validation, and Phase II BrightStar (n = 32) for clinical application. Tumor burden was quantified at whole-body and organ levels using CT-derived 3D volumetrics, combined with longitudinal routine blood test and demographic variables. Two digital twin models were developed through machine learning: OncoTwin-2D, a parsimonious model based on serial sum of longest diameters (SLD) and demographics, and OncoTwin-3D, an advanced model integrating longitudinal 3D volumetric, blood, and demographic features. Models were evaluated on MDACC and ALTA-1L cohorts by hazard ratio (HR) and log-rank test. The calibrated OncoTwin-3D was further applied to the single-arm BrightStar trial to simulate a counterfactual brigatinib-only control arm and evaluate the added benefit of local consolidation therapy (LCT). Results: Early tumor response patterns and long-term outcomes differed by TKI generation, with second-generation TKIs showing greater overall and organ-level responses and longer median PFS (29 vs 10 months; HR = 0.45; p 0.001) than first-generation TKI. For our digital twin framework, OncoTwin-2D achieved robust risk stratification (HR = 1.8, p = 0.034 in MDACC cohort; HR = 2.1, p 0.001 for external ALTA-1L cohort). Furthermore, model predicted survival outcomes aligned consistently with observed outcomes in ALTA-1L for first- and second-generation TKI. The advanced OncoTwin-3D further improved prognostic accuracy with significant risk stratification in MDACC cohort (HR = 3.9, p 0.0001) and separately for individual TKI subgroups (p = 0.006 and p 0.001 for first- and second-generation TKI). In the prospective BrightStar phase II trial, OncoTwin-3D was applied to simulate a counterfactual brigatinib-only control arm, which revealed a significant benefit from adding LCT (median PFS 66 vs. 22 months; HR = 2.8, p = 0.002). Conclusion: We introduced OncoTwin, an AI-driven multimodal digital twin for individualized response prediction and novel escalation evaluation. This scalable framework extends beyond thoracic disease, offering a generalizable paradigm that bridges real-world data and clinical trials to accelerate precision oncology. Citation Format: Hui Xu, Yasir Y. Elamin, Lingzhi Hong, Kyle Concannon, Maliazurina Binti Saad, Xinyan Xu, Muneer Amgad, Hui Li, Kang Qin, Xiaoyu Han, Sherif Ismail, Yuliya Kitsel, Saumil Gandhi, Mara B. Antonoff, Carol C. Wu, Brett W. Carter, Girish S Shroff, Simon Heeke, Xiuning Le, Tina Cascone, Natalie Vokes, Mehmet Altan, Don L. Gibbons, David Jaffray, Joe Y Chang, Zhongxing Liao, David Rice, Ara Vaporciyan, Stephen G G. Swisher, J Jack Lee, Jianjun Zhang, John V. Heymach, Jia Wu, . OncoTwin: A multimodal digital twin framework for predicting treatment response and guiding trial design in ALK-rearranged non-small-cell lung cancer abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 6724.