Driving Conformational Transitions in the Feature Space of Autoencoder Neural Network

The thermodynamic properties of interests for many physical or chemical processes are closely related to the relative stabilities and distributions of different metastable states. These statistical data are described by the underlying free-energy surfaces (FESs), which are typically embedded in a high-dimensional space that often appears incomprehensible. To resolve this curse of dimensionality in the calculation of the FESs by enhanced sampling techniques such as metadynamics, we introduce an approach based on using an autoencoder to construct a set of low-dimensional collective variables by iterative unsupervised training on a dataset. This is demonstrated for the case of conformational FES of cyclooctane that serves as a benchmark for rare event problems with complex FESs. Using the derived collective variables, the free-energy differences of canonical conformations were computed while preserving the essential symmetries of space.