Intelligent cell agents: decoding multicellular morphogenesis through inverse physical modeling & AI – CellAgents

This project aims to uncover how animal cells self-organize into complex structures through innovative inverse modeling and AI-based approaches. Focusing on early embryos and 3D engineered tissues, our research integrates physical modeling, machine learning, and computer vision to reverse-engineer developmental principles and address current challenges in tissue engineering. A central hypothesis of this project is that core functions of embryonic or stem cells (“hardware”) are conserved across species, while their context-dependent information processing (“software”) varies largely but could be learned.
The project is structured around three main objectives. First, we will develop new inverse modeling methods to infer spatiotemporal forces and model parameters from microscopy data, ensuring that our modeling of cellular mechanics and division patterns accurately capture tissue and embryo shape changes. Second, we will build a high-fidelity 3D agent-based simulator of tissue development that incorporates explicitly the stochastic nature of cellular signaling, and will create a surrogate tissue model based on graph neural networks for scalability and differentiability. Third, we will establish a computational framework for inverse morphogenesis, treating cells as trainable agents in interaction, that learn to respond to environmental cues using reinforcement learning (RL), even from sparse data such as microscopy of developing tissues.
The project will deliver advanced tools for mechanical inference, a robust and scalable tissue development simulator, and a novel RL-based framework for inverse morphogenesis. Supported by strong collaborations with international partners for biological validation, the results will contribute powerful computational tools for understanding development and advancing tissue engineering, with potential applications extending beyond biology, including material inverse design and multi-agent RL systems.