Mixtures of Neural Cellular Automata: A Stochastic Framework for Biological Growth Modelling

Neural Cellular Automata (NCAs) offer a powerful framework for modeling self-organizing processes with potential applications in biomedicine. However, their deterministic nature limits their ability to represent the stochastic dynamics of real biological systems.We introduce the Mixture of Neural Cellular Automata (MNCA), a novel extension that incorporates stochasticity and probabilistic rule clustering. By combining intrinsic noise with learned rule assignments, MNCAs can capture heterogeneous local behaviors and emulate the randomness inherent in biological processes.We assess MNCAs on synthetic tissue simulations and spatial transcriptomics data from mouse intestine. Our results show improved reconstruction of biological growth patterns and interpretable segmentation of local rules, establishing MNCAs as a promising tool for modeling complex biological dynamics.