Diffusion models learn underlying trends in actomyosin networks and predict behavior at unseen filament turnover

Generative diffusion models have demonstrated an ability to produce novel images sampled from the learned underlying data distribution. These models are able to interpolate system characteristics for underlying parameter combinations that were not seen during training. We consider cortical flow in a simulated actomyosin system across a range of filament turnover rates. We then train a diffusion model on curvature and density heatmap images, which is able to interpolate nonlinear trends in average curvature even from two extreme turnover rates. The strong linear relationship between average density and filament turnover in the system can allow the model to use the joint curvature and density distributions to predict curvature distributions for new turnover rates. Capturing these trends in curvature and density are sufficient to predict the cortical flow.