Speaker
Description
In this talk, I present a computational approach to precision oncology that combines mechanistic modeling with modern machine learning to build predictive digital twins; patient- and subgroup-specific models that forecast tumor progression and response to therapy. Our foundation is quantitative systems pharmacology (QSP) modeling, where tumor–immune dynamics are represented as systems of differential equations. We parameterize these models using patient observations and introduce a clustered inference strategy that stratifies individuals by immune profiles and estimates mechanistic parameters within each cluster. This reduces heterogeneity, improves identifiability, and yields interpretable subgroup phenotypes that link immune state to dynamical behavior. A distinguishing aspect of our evaluation is out-of-treatment generalization: we calibrate tumor growth and immune dynamics without fitting treatment response data, then use the resulting mechanistic parameters to predict outcomes under therapy. This design sharpens the scientific test of the model and targets the real-world challenge of predicting under new regimens. I then describe a digital twin platform that operationalizes this pipeline.
