Speaker
Description
The integration of mechanistic models with machine learning is becoming increasingly important for predicting treatment response and optimizing dose schedules in cancer therapy. Mechanistic models based on differential equations capture biological processes such as tumor growth and drug dynamics, while machine learning provides flexible tools for learning unknown components from data. However, two major challenges arise in this integration: identifiability of model components and the computational cost of repeated inverse problems. In this talk, I present recent work addressing both challenges. First, we develop a framework that establishes conditions for identifiability when simultaneously inferring unknown parameters and functional terms in differential equation models, improving the interpretability and reliability of hybrid mechanistic–ML approaches. Second, we introduce efficient inverse modeling methods based on physics-informed neural networks that enable rapid parameter inference through an offline-online decomposition. Together, these approaches provide practical tools for integrating mechanistic knowledge with data-driven learning to support treatment response prediction and exploration of personalized dosing strategies.
