Collective variable estimation with SEEKR.

Abstract

In biomolecular simulation, enhanced sampling methods improve the variety of states visited by the molecular system. Some enhanced sampling methods require the definition of a collective variable (CV) of interest - which often represents a slow or critical variable within the system, such as a reaction coordinate. Automated computational methods to identify the CV have been developed based on machine learning, principal component analysis, and mutual information - examples include the Automated Mutual Information and Noise Omission (AMINO) and Spectral Gap Optimization of Order Parameters (SGOOP) software. The SEEKR method is an approach to estimate the kinetics and thermodynamics of binding and requires the definition of a CV describing the process of interest. In this work we aim to develop a workflow for calculating single- or multi-dimensional CVs using algorithms that require minimal interaction or input from the user. This workflow consists of the production of MD trajectories through SEEKR using a weighing approach to recover the probability distributions. Trajectories are then analyzed using the AMINO algorithm for the identification of critical order parameters and the SGOOP algorithm to weigh the selected order parameters in a linear combination to form the CV. We test this method on a model two-well potential, a muller potential and real biomolecular systems.