Speaker
Description
Active biological systems often operate far from thermodynamic equilibrium, leading to violations of detailed balance and irreversible stochastic dynamics. Red blood cells provide a model system to study nonequilibrium phenomena through membrane flickering, spontaneous fluctuations driven by thermal noise and active intracellular processes. The statistics of these fluctuations reflect the mechanical and dynamical state of the membrane.
The aim of this work is to determine whether the observed dynamics depart from thermodynamic equilibrium. Classical approaches detect nonequilibrium behaviour by estimating entropy production or temporal irreversibility from stochastic trajectories. However, these methods require an explicit discretization of the state space and reliable statistical estimation, which is challenging in noisy experimental data.
We propose a data-driven machine-learning framework that learns tokenized stochastic time-series embeddings, transforming continuous trajectories into sequences of discrete latent states (tokens) through vector quantization. This representation provides a data-driven coarse-graining of the dynamics. Token distributions are analysed using a Self-Organizing Map, yielding a topological embedding of the learned dynamical states. The method is validated on Brownian dynamics simulations and applied to erythrocyte membrane fluctuations, enabling identification of dynamical states associated with nonequilibrium behaviour.
