We develop mechanistic models at the cell and tissue level based on PhysiCell and PhysiBoSS, combining probabilistic Boolean networks for intracellular signalling with agent-based rules for cell growth, migration and interactions. These simulations have been applied to predict drug synergies, modes of metastasis, and responses to metabolic conditions. However, building patient-specific digital twins remains complex and cumbersome. Mechanistic models offer interpretability and causal reasoning but struggle with parameter identifiability and sparse patient-specific data. 

To address these problems we are working in embedding mechanistic agent-based simulations within an AI-assisted architecture. The system combines our multiscale simulator with a data integrator that learns patient-specific network structures and model parameters through a bidirectional co-design loop, enabling AI to inform mechanistic parameterisation while simulations generate training constraints for AI models. The initial developments are explored in the context ofthe simulation of perturbations in organoids with drugs and drug combinations and cancer progression and treatment response. 

In my view, this approach illustrates a paradigm for AI for Science where machine learning and mechanistic modelling work as integrated components of a unified framework for understanding and predicting complex biological systems.