Evaluating Time-Series Foundation Models as Zero-Shot Surrogates for Mechanistic Virtual Patients

Ahmad Wisnu Mulyadi, Charlie George Barker, Sanjana Balaji Kuttae, Lilija Wehling, Thomas Rückle, Nicolas BOUCHER, Firas Abdessalem, Sven Jager, Anastasios Siokis, Sommer Anjum, Mohammed H. Mosa, Thomas Klabunde, Tommaso Andreani, Gurdeep Singh

Keywords: quantitative systems pharmacology, virtual patients, time-series foundation model, time-series forecasting

TL;DR: This study explores and benchmarks a couple of time-series foundation models by analyzing the forecasting results comprehensively and determines whether they can act as zero-shot surrogates for QSP-based virtual patient simulations.

Abstract: Mechanistic models such as Quantitative Systems Pharmacology (QSP) models are widely used to simulate the behavior of virtual patients (VPs) under different therapeutic conditions, supporting hypothesis generation and trial design. However, large-scale VP simulations are computationally expensive and require expert calibration. Recent advances in time-series foundation models (TS-FM) have demonstrated strong generalization across diverse temporal domains in a zero shot manner. In this study, we explore whether these models can act as zero-shot surrogates for QSP-based VP simulations. Using simulation outputs from 5 representative QSP models across multiple treatment scenarios and VP parameterizations, we benchmark 3 TS-FMs on their ability to predict future system trajectories. Our results show that TS-FM capture certain pharmacodynamic patterns and VP-level variability without fine-tuning, although performance varies between biological systems. This work highlights both the promise and the current limitations of using TS-FMs to accelerate VP-based in silico experimentation.

Release To Public: Yes, please release this paper to the public

Submission Number: 29