Deep Reinforcement Learning for Biological Modeling

A common difficulty in biological modeling is finding both biologically realistic and appropriate sets of parameters for your simulation. The amount of parameters begins to increase significantly and you increase the size and scope of your simulation. One solution is to use a form of optimization or search, to find functional and informative parameters for your model. This project combines two computational modeling platforms: CompuCell3D a multicellular modeling environment and popular deep-learning Python packages. 

Specifically, in this project, we present a deep learning approach for optimizing bacterial chemotaxis within the CompuCell3D framework, a powerful agent-based simulation platform for studying multi-scale cellular phenomena. Our study focuses on the run-and-tumble model of bacterial motion, which is characterized by periods of straight "runs" interspersed with random "tumbles" that change the cell's direction. We aim to optimize the chemotactic behavior of the bacteria in response to a spatially distributed food source by modulating three key output parameters: the probability of tumbling, the persistence of the cell, and its sensitivity to environmental stimuli. Importantly, this work serves as a foundation and a proof of principle that deep learning techniques as derived from Keras and TensorFlow can be implemented within CC3D for a variety of different purposes.

For the relevant code, mathematics, and educational resources, please visit the associated Github repository .