Beyond RECIST: mathematical modeling and Bayesian inference reveal the importance of immune suppressive parameters in metastatic breast cancer

Abstract

Successful immunotherapies must overcome patient- and organ-specific tumor heterogeneities to mount an effective response. Yet tumor dynamics remain poorly characterized in organ-specific contexts and associated immune environments. To quantify heterogeneous tumor responses, we developed methods to fit mathematical models of the tumor-immune dynamics to patients undergoing combination therapy for metastatic breast cancer: checkpoint inhibition via nivolumab + ipilimumab combined with entinostat, as measured by RECIST criteria. In a subset of patients additional immune dynamics were quantified by spatial proteomics. Bayesian parameter inference revealed that only immune-modulatory parameters controlled response; parameters controlling cytotoxicity were uninformative. Through posterior parameter sampling and simulation, we created virtual tumor cohorts, enabling extrapolation beyond the data to predict probabilities of response in metastatic lesions where no data exist. We validated pre-dictions from our virtual tumor population using held-out data characterizing off-target lesions from the patient cohort. Profile-wise likelihood analysis revealed that scans in the week immediately following treatment hold particularly high value in identifying the tumor dynamics. Overall, we demonstrate how through modeling & inference cohort size limitations can be over-come through the creation of virtual tumor populations, giving insight into the site-specific mechanisms of disease progression and response.