181 - Validation of a Digital Twin-Based Clinical Trial Design Approach Combining Radiotherapy and Tyrosine Kinase Inhibitors: Prospective Virtual Trial Predictions Match Real-World Data

Purpose/Objective(s): The immense resources needed to run randomized controlled trials (RCTs) pose a practical challenge. Multimodal therapy, combining chemo-, targeted, immuno-, and radio-therapy, poses an additional burden with a myriad of treatment schemes that need to be tested. 50% of late-stage clinical trials fail to meet their endpoint, further highlighting the need for improved RCT design. Here, we demonstrate that prospective predictions of a digital twin-based virtual clinical trial accurately match analogous RCT results published after the model, highlighting the potential of this method to aid RCT design.

Materials/Methods: In 2020, we published a biomathematical model where stochastically drawn digital twins undergo treatment combinations of chemoradiation (CRT) and tyrosine kinase inhibitors (TKIs). This framework simulates both intratumoral and patient-level heterogeneity by calibrating to both patient-level data and population-level RCT results of EGFR-mutated, locally advanced non-small cell lung cancer (EGFRmut-LA-NSCLC). Our model determined an optimal regimen of upfront CRT without TKI induction, followed by TKIs until progression. Since model publication, 3 studies investigating TKI+CRT regimens in EGFRmut-LA-NSCLC with available KM data were identified for validation of prospective predictions: 1) LOGIK0902 Phase II trial - progression-free survival (PFS) after 8 weeks of 1^st gen. TKI induction followed by CRT, 2) LAURA Phase III RCT - PFS after upfront CRT followed by 3^rd gen. TKI therapy until progression, and 3) LAURA Phase III RCT - freedom from distant failure (FFDF) for the same regimen.

Results: The LOGIK0902 1 & 2 year PFS rates were 58.1% (95CI [33.4 - 76.4]) & 36.9% [16.6% - 57.6%], similar to model predicted rates of 61.7% & 31.5%. The published LAURA 1 & 2 year PFS were 74% [64% - 80%] & 65% [56% - 73%], while the model predicted rates were 75.2% & 55.5%. The published LAURA 1 & 2 year FFDF rates were 85% and 81% (CI not reported), while model predicted rates were 84.3% and 67%. Our model correctly predicted the benefit of omitting TKI induction and the size of the effect. The model was calibrated to data from 1^st gen. TKIs, explaining the excellent agreement with the 1^st gen. LOGIK0902 trial and minor overestimation of progression in the 3^rd gen. TKI LAURA trial. Despite this, the model still accurately predicted upfront CRT with TKIs until progression to be an effective treatment with both LAURA trial PFS and FFDF having a log-rank p<0.001 (vs. CRT) and 1000 virtual simulations of this trial with the same sample size also having a median p<0.001 for both endpoints.

Conclusion: We have shown that digital twin approaches can accurately and prospectively predict real-world outcomes of multimodal treatments combining systemic and radiation therapy. Digital twin systems can analyze the entire multimodal treatment space in silico and and their demonstrated utility makes them a potential tool to improve the design of efficacious RCTs investigating multimodal treatment regimens.