(702e) Deep Learning of Slow Collective Variables to Understand and Accelerate Biomolecular Folding and Assembly

The variational approach to conformational dynamics provides a framework to estimate from time series data the leading eigenfunctions of the transfer operator that propagates the system dynamics. In the context of molecular simulation, this allows the estimation of the slowest collective modes governing the microscopic time evolution of the system from molecular dynamics trajectories. These collective variables are extremely valuable in understanding the important facets of protein folding and self-assembly, and in furnishing good coordinates in which to perform accelerated sampling. We introduce a novel neural network architecture and loss function that we term state-free reversible VAMPnets (SRVs) to estimate within a variational approximation the full orthogonal spectral hierarchy of transfer operator eigenfunctions. Applications to the Trp-cage and WW domain mini-proteins and the hybridization of short DNA-oligomers identifies the slow pathways governing molecular folding and provides an optimal basis set for the construction of Markov state models (MSMs) with higher temporal resolution and state decomposition than state-of-the-art techniques. This approach unveils new understanding of the metastable states and folding and assembly pathways of these systems.