Chasing collective variables using temporal data-driven strategies

Abstract The convergence of free-energy calculations based on importance sampling depends heavily the choice collective variables (CVs), which in principle, should include slow degrees freedom biological processes to be investigated. Autoencoders (AEs), as emerging data-driven dimension reduction tools, have been utilised for discovering CVs. AEs, however, are often treated black boxes, and what AEs actually encode during training, whether latent from encoders suitable CVs further remains unknown. In this contribution, we review their time-series-based variants, including time-lagged (TAEs) modified TAEs, well closely related model variational approach Markov networks (VAMPnets). We then show through numerical examples that learn high-variance modes instead modes. stark contrast, time series-based models able capture Moreover, both TAEs with extensions feature analysis state-free reversible VAMPnets (SRVs) can yield orthogonal multidimensional As an illustration, employ SRVs discover isomerizations N -acetyl- ′-methylalanylamide trialanine by iterative learning trajectories biased simulations. Last, experiments anisotropic diffusion, investigate potential relationship committor probabilities.