(520e) Understanding and Applying Multistate Reweighting in a Broad Molecular Simulation Context

A recurring challenge in simulations of molecular systems of interest to chemical engineers is estimating physical quantities from sampled ensembles. Collecting sufficient statistics for reliable estimates often requires collecting multiple data sets under varying conditions, each from a different ensemble consistent with the different macrostate.

A large number of methods have been developed to analyze these data sets. The multistate Bennett Acceptance ratio method, based on recent advances in statistical sampling theory, unifies ideas from a number of different analysis methods for multiple ensembles. MBAR is provably the lowest variance unbiased estimator of both free energies and ensemble averages, and has a number of other features, such as no requirement for histograms, scaling with data set size rather than dimension of the problem, and an explicit error estimate.

In this talk, we analyze a number of insights about the nature of multistate reweighting that have emerged since the approach was introduced. We discuss how these methods can be applied to a wide variety of problems such as computing potentials of mean force, force field parameterization, and calculating free energy differences, with an emphasis on multidimensional parameter spaces. We resolve common confusions that may occur when setting up a problem to perform multistate analysis and discuss MBAR implementations that can drastically reduce the amount of time required and significantly increase robustness.