Analysis of Feature-Combination Dependency in Collective-Variable-Driven Enhanced Sampling Methods

Sammanfattning

Determining free-energy landscapes is vital to understanding thermodynamic behaviour of molecular systems like proteins. Molecular dynamics (MD) simulations are a powerful tool in obtaining structural insight through simulating time evolution of the system. To facilitate sampling of rough free-energy surfaces, enhanced sampling techniques are often implemented to apply a bias the simulation. In order to define the bias in a meaningful way, a low-dimensional function of atomic coordinates, known as a collective variable (CV), can be introduced. However, the inclusion of collective variables is not trivial, as they are required to capture large-scale dynamics to accelerate the sampling process. In this thesis, the prominence of how collective variables are constructed was investigated through applying the accelerated weight histogram (AWH) method to a small model system. Three sets of combinations composed of three atom--atom distances each were used to construct two-dimensional collective variables, for which the resulting free-energy profiles were compared to an established reference profile. Clear distinctions in the ability to reproduce an accurate profile is seen across the different sets while some less evident discrepancies within each set are present. While the results suggest even distance-combination impacts the functionality of the collective variable, improvements to the atom--atom-distance selection process, a substantiated error estimation method, and a larger sample size are necessary to further support this claim.