Neural-Network Model for Characterizing Stochastic Dynamic Variability of Cellular Particles

Particle dynamics characterization is fundamental for understanding the biophysical laws orchestrating cellular processes. To classify the dynamic behaviors governing biological particles, we develop a neural network model built on geometric descriptors of trajectories. The model infers the stochastic laws governing the trajectory, enabling the detection of a large family of dynamic behaviors, especially within the subdiffusive regime that characterizes cell signaling processes. Finally, we propose a framework to robustly detect dynamic changes in composed trajectories based on the variability of prediction scores on successive sub-trajectories. The method is validated on simulated composed trajectories simulating the activation pathway of receptors CCR5.