Location
MC 6460
Candidate
Saranya Varakunan | Applied Mathematics, University of Waterloo
Title
Scientific Machine Learning for Model Discovery and Inference in Mathematical Oncology
Abstract
Understanding and predicting cancer–immune dynamics remains a central challenge, given nonlinear interactions, limited clinical data, and patient-specific variability. In this seminar, I present several complementary approaches for model discovery and inference in cancer immunology, centered on integrating mechanistic modeling with machine learning methods. I first consider mechanistic models of tumor-immune interactions, focusing on a CAR-T cell therapy framework that captures lineage dynamics. I demonstrate how this model can be used to identify critical parameters governing treatment response and, using a feed-forward neural network, enable prediction of treatment outcomes while preserving interpretability. Next, I introduce a generative modeling approach to inverse problems. In particular, I develop a framework using GFlowNets and physics-informed neural network (PINN) losses to discover candidate dynamical models from sparse, noisy synthetic data; this is demonstrated on a simple gene regulatory network. Finally, I outline a broader research direction that integrates mechanistic and data-driven modeling approaches into a unified framework for model discovery and inference in mathematical oncology, with applications to emerging challenges such as immune system aging. In summary, this seminar shows how scientific machine learning methods enable principled discovery of dynamical models from sparse data, identification of key immunological mechanisms, and robust prediction of treatment outcomes.