Congratulations to Tri-Institutional PhD Program in Chemical Biology (TPCB) graduate student and NSF GRFP recipient Chaya Stern on receiving a 2018 MolSSI Phase I Fellowship to support her work in developing new algorithms and open source software for Bayesian inference of force field parameters from experimental and quantum chemical data! You can learn more about Chaya's work in this area by listening to her PyData NYC 2017 talk or reading her MolSSI Fellowship Proposal, and hear more about what Chaya is up to by following her twitter feed.
Congratulations to postdoc Gregory Ross for being awarded an inaugural MolSSI Postdoctoral Software Science Fellowship! Greg's project focuses on building a toolbox of self-tuning Monte Carlo methods for use alongside molecular dynamics based Markov chain Monte Carlo (MCMC) sampling to greatly enhance the efficiency of molecular simulations and facilitate the construction of new sampling algorithms. Read on for a more detailed description of his project.
An automatic sampling toolbox for molecular simulation
Gregory Ross, DPhil
In the biomolecular simulation community, molecular dynamics (MD) is the most popular method for sampling molecular configurations. MD is well suited to capturing global, concerted movements, but it struggles to sample configurations that are separated by large energetic barriers. As a result, configurations taken from MD simulations tend to be from the same local region. An equivalent problem is encountered in Bayesian inference when drawing samples from a high-dimensional, multimodal posterior distribution. MD-related techniques are a small subset of a large class of Markov chain Monte Carlo (MCMC) techniques, and practitioners in both the biomolecular simulation and Bayesian inference communities have, in principle, a great deal of choice on which sampling method to use. However, it is difficult to know which MCMC method will the most efficient at generating uncorrelated samples for a particular system. In my project, I will develop a toolbox for MCMC that is agnostic with regard to what is being sampled, whether a protein conformation or posterior distribution. I will focus on using adaptive MCMC methods to help select the sampling technique that works best for a given problem. The general idea is to write a library of MCMC tools and a software interface that allows for the construction of MCMC “blocks” which can be mixed together and easily applied. The goal is to simplify sampling from difficult distributions and, ultimately, to widen the range of phenomena currently accessible by molecular simulation.