Combining molecular simulations and semi-reliable data to determine protein structures

Bhumika Singh, Alberto Perez

University of Florida

Our group is interested in integrating molecular simulations with data that is insufficient on its own, to determine protein structures. Modeling Employing Limited Data (MELD) is a physics-based Bayesian inference that combines prior belief from a given force field with semi-reliable data to resolve data ambiguity and produce a statistically consistent ensemble of protein structures. One general challenge in using any data is estimating the noise and ambiguity in the original data set, which informs simulations of how much data to enforce. This is a critical step in Bayesian inference because trusting too much data will incorporate noise into the predictions, resulting in incorrect models. While at the opposite, if a low amount of data is trusted, simulations will face problems of convergence and efficiency in sampling phase space and ultimately reduce the performance of the method. In the presented study, a new version of MELD is used that optimizes the fraction of data to be trusted “on the fly.” The methodology is tested on a dataset of 10 proteins and compared to the previous results.