Regimen-Aware Forecasting for Mechanistic Virtual Patients with Time-Series Foundation Models

Quantitative Systems Pharmacology (QSP) models provide mechanistic representations of disease progression and drug response in virtual patient populations, but their construction and maintenance are resource-intensive. Large pretrained probabilistic time-series foundation models (TSFMs) exhibit strong cross-domain generalization, yet their application in mechanistically informed drug development remains limited. Here, we present a framework that augments QSP-based virtual patient generation with a fine-tuned TSFM surrogate. We fine-tuned Chronos-2 on a synthetic dataset generated from a curated portfolio of mechanistic QSP models and evaluated its performance on an unseen inflammatory bowel disease (IBD) QSP model. Forecasts were informed with explicit context in the form of historical and future drug dosing trajectories as exogenous inputs. Regimen-aware fine-tuning reduced predictive error and tightened uncertainty estimates relative to zero-shot inference. Incorporating dosing context alongside fine-tuning lowered both mean prediction error and variability across virtual patients. These results demonstrate that context-informed TSFMs can serve as scalable, reproducible surrogates for QSP workflows, enabling rapid scenario exploration in pharmacological modeling pipelines.