Speaker
Description
Virtual clinical trials (VCTs) hold significant promise for improving the drug development process, yet their predictive reliability depends on design decisions that remain poorly understood. This presentation investigates how model complexity, prior parameter distributions, and virtual patient (VP) inclusion criteria interact to shape VCT outcomes.
Using oncolytic virotherapy in murine tumors as a case study, we compared three models of varying complexity using different parameter priors and methods for including/excluding a parameterization in a virtual population. Our results demonstrate that while the simplest model inadequately spans the feasible trajectory space, potentially missing critical interpatient heterogeneity, there are diminishing returns beyond intermediate complexity. Both intermediate and complex models captured similar ranges of patient responses across various dosing protocols.
Further, we show that methods that simply reject “out-of-bound” parameterizations can generate posterior distributions that overly resemble the chosen prior, artificially reducing variability in treatment responses. In contrast, patient generation methods that instead perturb “out-of-bound” VPs to make them feasible produced results less sensitive to prior assumptions. These findings suggest that VCT design should prioritize intermediate-complexity models to capture key biological mechanisms, paired with perturbation-based inclusion criteria that prevent unrealistic prior assumptions from overconstraining the virtual population. [Joint with Dr. Joanna Wares]
