Nowadays molecular dynamics (MD) has become a standard tool to investigate molecular systems. By simulating the dynamic behavior of molecules, it facilitates the exploration of stability, conformational changes and a plenty of other properties essential for understanding molecular functions. However, the application of MD is affected by the large computational costs. It computes steps that must be in order of femtoseconds, to assure numerical stability, for a time scale long enough to likely catch some or rarely occurred processes. There are numerous techniques to enhance simulations to optimize the sampling of the slow motions, one of those is metadynamics. It operates by biasing the Hamiltonian of the system, encouraging it to cross barriers and explore configurations that might otherwise be inaccessible or challenging to sample, in a reasonable amount of time. However, to make metadynamics successful, the scientist has to select the so-called “collective variables'' which are functions of Cartesian coordinates. Designing good collective variables is not a trivial task and it relies on the knowledge of the system and experience of the scientist. Machine learning and artificial neural networks have shown incredible power in supporting the exploration of conformational phase space.
Hereby we show the involvement of two methods based on the control and optimization of the latent space. The developed neural network exhibits the ability to analyze simulation data and to derive optimal combinations of internal coordinates to be used as CVs. The power and the efficiency of the presented approach are demonstrated on Trp-cage foldings.