Characterizing particle dynamics in live imaging through stochastic physical models and machine learning

Abstract

Particle dynamics determine the orchestration of molecular signaling in cellular processes. A wide range of subdiffusive motions has been described at the cell interior and membrane, corresponding to different environmental constraints. However, the standard methods for motion analysis, embedded in a diffusion-based framework, lack robustness for capturing the complexity of stochastic dynamics. This work develops a classification method to detect the five main stochastic laws modeling particle dynamics accurately. The method builds on machine-learning techniques that use features properly designed to capture the intrinsic geometric properties of trajectories governed by the different processes. This guarantees the accurate classification of observed dynamics in an interpretable and explainable framework. The main asset of this approach is its capability to distinguish different subdiffusive behaviors making it a privileged tool for biological investigations. The robustness to localization error and motion composition is proven, ensuring its reliability on experimental data. Moreover, the classification of composed trajectories is investigated, showing that the method can uncover the paths mono-vs bi-dynamics nature. The method is used to study the dynamics of membrane receptors CCR5, involved in HIV infection. Comparing the basal state to an agonist-bound state which displays potent anti-HIV-1 activity, we show that the latter affects the natural dynamic state of receptors, thus clarifying the link between movement and receptor activation.