Enhanced Sampling of Molecular Dynamics Simulations with Self-learning Collective Variables

BEHROOZ HASHEMIAN; DANIEL MILLÁN; MARINO ARROYO

Barcelona

Encuentro; XXX Annual Meeting of Reference Network of R+D+i on Theoretical and Computational Chemistry; 2014

Institut de Química Avançada de Catalunya (IQAC), CSIC

One of the main outstanding issues in molecular dynamics simulations is accessing the relevant time-scales for conformational transitions in functional molecules, as a result of metastability. This makes it difficult to connect simulations with experimental observables. To overcome this issue, a number of enhanced sampling methods have been proposed, such as metadynamics and adaptive biasing force. However the effectiveness of these methods strongly relies on the choice of a good set of collective variables (CVs), which need to compactly describe molecular conformations, capture metastability, and be differentiable. There is increasing research towards creating CVs from statistical learning techniques. Here, we present a self-learning methodology based on manifold learning techniques that, starting from pre-existing computational or experimental ensembles representative of the conformational variability, builds smooth and nonlinear data-driven collective variables (SandCV). We integrate SandCV within standard MD codes and enhanced sampling methods. This methodology can deal seamlessly with intrinsic manifolds of complex topology described with multiple parameterization charts. SandCV is independent of the enhanced sampling method, minimally invasive with respect to the underlying molecular dynamics code and is being implemented in PLUMED to be available for the scientific community.