West Matthew, Yoshida Kenta, Yu Jiajie, Lemaire Vincent
Department of Biostatistics, Harvard T.H. Chan School of Public Health, Boston, MA, USA.
Modeling and Simulation, Clinical Pharmacology, Genentech Inc., South San Francisco, CA, USA.
NPJ Precis Oncol. 2025 Apr 14;9(1):106. doi: 10.1038/s41698-025-00898-6.
Quantitative systems pharmacology (QSP) models of cancer immunity provide mechanistic insights into cellular dynamics and drug effects that are difficult to study clinically. However, their inability to predict patient survival mechanistically limits their utility in anti-cancer drug development. To overcome this, we link virtual patients from a QSP model to real clinical trial patients. Using data from atezolizumab trials in non-small cell lung cancer, we show that tumor-based linkage effectively captures survival outcomes. By treating linked survival and censoring as weak supervision labels, we trained survival models using only QSP model covariates, without clinical covariates. Our approach also predicts survival for treatments not included in training data. Specifically, we accurately estimated survival hazard ratios (HR) for chemotherapy monotherapy and atezolizumab plus chemotherapy combination. The predicted HR of 0.70 (95% prediction interval [PI] 0.55-0.86) closely matches the observed HR of 0.79 (95% PI 0.64-0.98) from the IMpower130 trial.
癌症免疫的定量系统药理学（QSP）模型为细胞动力学和药物效应提供了机制性见解，而这些在临床上很难进行研究。然而，它们无法从机制上预测患者的生存情况，这在很大程度上限制了它们在抗癌药物研发中的应用。为了克服这一问题，我们将QSP模型中的虚拟患者与真实的临床试验患者联系起来。利用非小细胞肺癌阿替利珠单抗试验的数据，我们表明基于肿瘤的关联有效地捕捉到了生存结果。通过将关联的生存和删失视为弱监督标签，我们仅使用QSP模型协变量而不使用临床协变量来训练生存模型。我们的方法还能预测未包含在训练数据中的治疗的生存情况。具体而言，我们准确估计了化疗单药治疗以及阿替利珠单抗联合化疗的生存风险比（HR）。预测的HR为0.70（95%预测区间[PI]0.55 - 0.86），与IMpower130试验中观察到的HR 0.79（95% PI 0.64 - 0.98）非常匹配。
