Abstract
Abstract We present a new approach to predict overall survival (OS) in untested oncology trial designs, which does not require the use of clinical covariates. This approach is based on generating a virtual population from a quantitative systems pharmacology (QSP) model of cancer immunology, and linking this virtual population to real patients from previous clinical trials. QSP has emerged as a dominant paradigm in recent years for investigating disease mechanisms and allowing prediction of drug effects in silico. However, survival cannot be described mechanistically and intrinsically predicted from a QSP model. We develop a weakly supervised learning approach to impute labels of OS and censorship in the virtual population. This approach requires prior matching of the virtual patients to real patients on the basis of longitudinal tumor growth curves. The idea that there exists a predictive relationship between tumor dynamics and OS was motivated by previous work on tumor growth inhibition (TGI-OS) [Claret, L. et al., J. Clin. Onc., 2013]. In contrast to the TGI-OS framework, we rely solely on simulated QSP dynamical variables for predicting survival. This allows us to derive OS predictions for trial designs different than the ones used for model development. Data from 5 clinical trials for atezolizumab in NSCLC (BIRCH, FIR, OAK, POPLAR, and IMpower110, total N = 1641) were used to link survival labels to virtual patients. 90% of the data was used for training, and 10% was held out for model validation. The imputed OS labels were used to train a log-normal accelerated failure time model on the training data, and predictions of survival in the test data were in good agreement with a Kaplan-Meier estimate of the clinical survival. The model predicted a median OS of 471 days (95% CI 453-490) compared to the observed 475 days, (95% CI 440-517). The rate of change of tumor size under treatment in the QSP model was the most predictive feature of survival, with a hazard ratio of 1.79 between 95th percentile and median of the dynamical range (95% CI 1.69-1.81). Model variables related to tumor antigens, cytotoxic cell death, and T cell dynamics were also relevant to the prediction of survival. This work provides the first example of generating and validating predictions of OS without using covariates from actual clinical trials. We intend to expand this methodology across different trial designs and different subpopulations. For example, our approach could be used to estimate hazard ratios between different treatments or combination therapies, as well as for patients grouped by PDL1 expression or by line of therapy. While this work only considered overall survival as an endpoint, our approach can potentially be extended to endpoints like progression-free-survival. Citation Format: Matthew West, Kenta Yoshida, Jiajie Yu, Vincent Lemaire. A treatment-agnostic approach to predict patient survival from virtual clinical trials [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 1923.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have