Abstract
An accurate energy function is an essential component of biomolecular structural modeling and design. The comparison of differently derived energy functions enables analysis of the strengths and weaknesses of each energy function and provides independent benchmarks for evaluating improvements within a given energy function. We compared the molecular mechanics Amber empirical energy function to two versions of the Rosetta energy function (talaris2014 and REF2015) in decoy discrimination and loop modeling tests. In decoy discrimination tests, both Rosetta and Amber (ff14SBonlySC) energy functions performed well in scoring the native state as the lowest energy conformation in many cases, but several false minima were found in with both talaris2014 and Amber ffl4SBonlySC scoring functions. The current default version of the Rosetta energy function, REF2015, which is parametrized on both small molecule and macromolecular benchmark sets to improve decoy discrimination, performs significantly better than talaris2014, highlighting the improvements made to the Rosetta scoring approach. There are no cases in Rosetta REF2015, and 8/140 cases in Amber, where a false minimum is found that is absent in the alternative landscape. In loop modeling tests, Amber ffl4SBonlySC and REF2015 perform equivalently, although false minima are detected in several cases for both. The balance between dihedral, electrostatic, solvation and hydrogen bonding scores contribute to the existence of false minima. To take advantage of the semi-orthogonal nature of the Rosetta and Amber energy functions, we developed a technique that combines Amber and Rosetta conformational rankings to predict the most near-native model for a given protein. This algorithm improves upon predictions from either energy function in isolation and should aid in model selection for structure evaluation and loop modeling tasks.