This research is motivated by acceleration. Molecular simulations make true possibility to simulate the motion from small molecules to big proteins and their combinations in drug-target complexes. It let us to predict their changing confirmation, their stability and a plenty of other properties thanks to the evolution of molecular structure. However, application of molecular simulations is affected by the large computational costs in computing steps that must be in order of femtoseconds, to assure numerical stability to integrate Newton equation of motion. Taking into account this limitation, a typical molecular dynamics simulation is capable to sample only a small fraction of the states available to the simulated system, with the likely catch or unlikely loss of some slow or rarely occurring processes, where likelihood depends on the simulation time. There are numerous techniques to address this limitation and to speed up simulations. Metadynamics is an enhancing method based on biasing Hamiltonian of the system that helps to cross barriers and go head through new unexplored free energy surface areas, thanks to some selected internal coordinates, so called collective variables. Choosing correct collective variables to make metadynamics successful is not a trivial task and it depends first of all on a knowledge and expertise of the user. In last few years there are emerging opportunities for machine learning and artificial neural networks in this field. We decided to develop an adversarial autoencoder as a tool to analyse simulation data and to support user to derive good collective variables to enhance molecular dynamics simulation. The potential of this platform is demonstrated on Trp-cage folding.
We thank D.E Shaw Research to provide us trpcage trajectory for
asmsa training. This work was supported by the Czech Science Foundation
(22-29667S).