Three birds with one stone: reaction coordinate, thermodynamics and kinetics from all-atom molecular simulations

Many molecular systems involve processes with intertwined spatio-temporal resolutions ranging from femtoseconds to days, making it hard to probe them completely using traditional experimental tools. It has been a holy grail to simulate these in all-atom resolution using molecular dynamics methods, but these can go up to only a few hundred microseconds even with the most powerful and custom-built supercomputers. Thankfully, over the decades several sampling algorithms have been proposed that can simulate these complex systems in an accelerated but controllable manner. However, a large class of these methods (arguably all!) need an a priori sense of a low-dimensional reaction coordinate (RC) even before performing the sampling. This has severely limited the usefulness of such sampling methods. In order to deal with this cyclic problem where one needs extensive sampling of the rare events to know the RC, but also needs to know the RC in the first place to perform sampling, it is thus extremely desirable to construct methods that learn the RC as they perform the sampling. Here we will describe two such methods, namely SGOOP [1,2] and RAVE [2] developed by us that use flavors of statistical mechanics and deep learning to solve this problem. We will demonstrate the generality and power of these methods by showing how they give direct predictive insight into biologically important problems such as mechanisms of ligand-protein (including T4L99A lysozyme and tyrosine kinases), and transcription factor-DNA (including Epstein Barr virus binding domain) interactions. Our findings includes an all-atom characterization of metastable and transition states, their stabilities, various rate constants, as well as prediction of deleterious point mutations in the system which could upend the functioning of the protein, DNA or the ligand.