Multi-modal AI modeling to predict clinical trial outcomes and benchmarks pan-RAS versus KRAS G12D inhibition in KRAS G12D–mutant PDAC.

Abstract

4020 Background: In pancreatic ductal adenocarcinoma (PDAC), uncertainty persists regarding the optimal positioning of pan-RAS inhibitors, allele-specific KRAS inhibitors, and chemotherapy backbones. While patient-derived organoids (PDOs) provide functional resolution of drug response, their clinical utility has been limited by cohort size and insufficient validation against trial outcomes. Methods: A biobank of 135 colorectal cancer (CRC) and PDAC PDOs was assembled with matched clinical annotation, omics characterization, and functional drug profiling. Multi-modal AI models integrating molecular features, PDO drug response, and clinical covariates were trained to predict patient-level progression-free survival (PFS) and longitudinal response, and to aggregate these predictions into cohort-level clinical trial outcomes. Model performance was evaluated by benchmarking predicted outcomes against published results across 22 treatment arms from multiple clinical trials. A pan-RAS inhibitor (daraxonrasib), a KRAS G12D-directed inhibitor (zoldonrasib), and FOLFIRINOX (FFX) were compared as single agents or in combination in a KRAS G12D-mutant PDAC subset (n = 23), with external benchmarking to published clinical data. Results: The AI models demonstrated robust predictive performance at both the patient and cohort levels. Patient-level predictions were driven by functional PDO drug response and established clinical factors, including performance status and prior treatment exposure. At the cohort level, predicted and published outcomes showed good agreement (R² = 0.63), with correct ranking of treatment arms by efficacy in 78% of cases. In pairwise arm-to-arm comparisons, the model identified the superior regimen in 80% of comparisons. When applied to KRAS G12D-mutant PDAC, virtual clinical trial simulations preserved the relative efficacy hierarchy observed in external datasets. Zoldonrasib showed heterogeneous predicted activity (simulated ORR 20% vs reported 30%), whereas daraxonrasib demonstrated more favorable predicted outcomes (simulated ORR 35% vs reported 36%). Combination with FFX further improved predicted efficacy for zoldonrasib, particularly in first-line simulations (simulated ORR 38% vs reported 63%). Conclusions: Multi-modal AI modeling trained on clinically annotated PDO collections enables calibrated simulation of clinical trial outcomes and comparative assessment of therapeutic strategies in KRAS-mutant PDAC. By capturing patient-level heterogeneity and cohort-level efficacy trends, this framework provides a scalable decision-support tool for arm selection, therapeutic positioning, and trial design in KRAS-directed drug development.