Predicting survival in prospective clinical trials using weakly-supervised QSP

Quantitative systems pharmacology (QSP) models of cancer immunity offer a mechanistic understanding of cellular dynamics and drug effects that are often challenging to investigate clinically. Despite their success, these models are limited by their inability to mechanistically represent patient survival as an output, which restricts their utility in the clinical development of anti-cancer drugs. To address this limitation, we propose a novel approach that links virtual patients, generated from a QSP model, to real patients from clinical trials. Utilizing data from atezolizumab clinical trials in non-small cell lung cancer (N = 1641), our findings demonstrate that tumor-based linkage methods can effectively capture survival outcomes. The linked properties of survival and censoring serve as weak supervision labels, enabling the training of survival models on QSP model covariates alone, without the need for clinical covariates. Furthermore, our approach can predict drug effects on survival for treatments not included in the training datasets. Specifically, we accurately estimated the survival hazard ratio (HR) for two external treatments: a chemotherapy monotherapy arm and an atezolizumab + chemotherapy combination arm. The predicted HR was 0.70 (95% PI 0.55–0.86), closely matching the observed HR of 0.79 (95% PI 0.64–0.98) from the IMpower130 clinical trial.