Patient-level Prediction of Oncology Clinical Trial Outcomes Using a Calibrated Pan-Cancer Foundation Model for Synthetic Comparator Arms and Indirect Treatment Comparisons

We present a framework that calibrates patient-level SOC predictions from RWD to published clinical trial evidence. The framework has two complementary components: a pre-trained pan-cancer foundation model that enables transfer learning across biomarkers, indications, and treatments, enhancing the utility of small and sparse datasets; and a principled calibration procedure that anchors patient-level simulations to gold-standard aggregate trial evidence.