Measuring irreversibility from learned representations of biological patterns

Abstract: Thermodynamic irreversibility is a crucial property of living matter. Irreversible processes maintain spatiotemporally complex structures and functions characteristic of living systems. In high-dimensional biological dynamics, robust and general quantification of irreversibility remains a challenging task due to experimental noise and nonlinear interactions coupling many degrees of freedom. Here we use deep learning to identify tractable, low-dimensional representations of phase-field patterns in a canonical protein signaling process -- the Rho-GTPase system -- as well as complex Ginzburg-Landau dynamics. We show that factorizing variational autoencoder neural networks learn informative pattern features robustly to noise. Resulting neural-network representations reveal signatures of mesoscopic broken detailed balance and time-reversal asymmetry in Rho-GTPase and complex Ginzburg-Landau wave dynamics. Applying the compression-based Ziv-Merhav estimator of irreversibility to representations, we recover irreversibility trends across complex Ginzburg-Landau patterns varying widely in spatiotemporal frequency and noise level. Irreversibility estimates from representations similarly recapitulate cell-activity trends in a Rho-GTPase patterning system undergoing metabolic inhibition. Additionally, we find that our irreversibility estimates serve as a dynamical order parameter, distinguishing stable and chaotic dynamics in these nonlinear systems. Our framework leverages advances in deep learning to offer robust, model-free measurements of nonequilibrium and nonlinear behavior in complex living processes.