This work has demonstrated that by matching clinical patients to virtual patients simulated from a QSP model, it is possible to impute weak supervision labels for survival and censoring in virtual patients. Furthermore, we have demonstrated that using QSP covariates alone, these labels can be used to train survival models that can generalize to different treatment groups, producing estimates of treatment effect consistent with those observed in clinical trials. The fact that we can make a prediction of survival based on QSP covariates alone is a remarkable property of our approach, allowing us to train the model only once and apply it to different clinical settings (that can be represented in the QSP model).
The model predictions for the OS HR demonstrate good alignment with the observed HR in IMPower130. However, it is important to note that the observed HR was derived from an intention to treat (ITT) population. In contrast, our analysis consisted of virtual patients who required at least one post-baseline tumor size measurement to be eligible for matching, which could result in shorter survival times being truncated. Fortunately, patients with only one baseline tumor measurement usually represent only a very small subset of the clinical trial population. While acknowledging the limited presence of patients with a single tumor measurement, we may consider addressing their inclusion in future work to enhance the precision of our findings. Additionally, incorporating stratification of hazard ratio predictions based on other model covariates could help mitigate violations of the proportional hazard assumption and the influence of any other potential sources of bias.
We also investigated using this method to impute clinical covariates in the virtual population, in addition to OS labels, in order to use these clinical covariates as features in the survival model. While we found that they did not significantly alter the performance of the model, future work could explore hybridized approaches to predicting outcomes in virtual populations, that would involve imputing clinical features from real patients in a virtual population, in addition to the survival endpoints imputed in this work.
A unique and valuable aspect of our approach in the context of survival analysis is that our survival model not only benefits from the inclusion of tumor-derived covariates but also from non-tumor-derived covariates, particularly those derived from immune cell populations. This demonstrates that features in the QSP model that are not directly related to tumor growth can exhibit a predictive effect on survival. To the extent that QSP models accurately describe the underlying biology, future work may be able to leverage this kind of methodology to infer molecular or immunological events that are ultimately predictive of survival.
Further improvement in model predictions could be made possible by alternative approaches to matching real patients to virtual patients, for example using immunological markers like PDL1 expression in matching, in addition to tumor dynamics. Further investigation of QSP feature extraction methods would be of interest, both as a means of improving the predictive power of this approach, as well as aiding in providing more detailed biological interpretations.
While survival analysis is a methodology more familiar within traditional statistics, deep learning approaches may add additional predictive power in providing in silico trial endpoints, though perhaps at the cost of some interpretability 21,37. Deep learning approaches would be suited to unstructured data like time-series, learning informative representations of QSP signals that would not require the heuristic feature extraction performed in this study.
This work provides the first example of an approach that could be applied to QSP models and survival data for other indications and pathways, allowing progression beyond this proof-of-concept by validation across other diseases and drug effects. While this work only considered overall survival as an endpoint, similar approaches could be extended to endpoints like progression-free-survival, which may be of clinical interest.