Generative diffusion model surrogates for mechanistic agent-based biological models

Area of Science:

• Computational Biology
• Artificial Intelligence
• Biophysics

Background:

• Mechanistic, multicellular, agent-based models (MCMs) like the Cellular-Potts Model (CPM) are crucial for single-cell resolution biological investigations.
• Computational expense of MCMs at large scales hinders their application.
• Stochasticity in MCMs complicates the development of surrogate models.

Purpose of the Study:

• To develop a generative AI surrogate model for CPMs using denoising diffusion probabilistic models (DDPMs).
• To accelerate the evaluation of complex biological systems simulated by CPMs.
• To enable the creation of digital twins for stochastic biological systems.

Main Methods:

• Leveraged denoising diffusion probabilistic models (DDPMs) to train a generative AI surrogate of a CPM.
• Employed an image classifier to identify unique regions within a 2D parameter space.
• Utilized the classifier for surrogate model selection and verification.

Main Results:

• The DDPM-based surrogate model successfully generated CPM configurations 20,000 timesteps ahead of a reference.
• Achieved an approximate 22x reduction in computational time compared to native CPM code execution.
• Demonstrated the feasibility of using AI for accelerating complex biological simulations.

Conclusions:

• DDPMs can be effectively implemented to create efficient surrogate models for stochastic agent-based models.
• This approach significantly reduces computational burden, facilitating the study of complex biological processes.
• The developed surrogate models represent a significant step towards creating accurate digital twins of biological systems.