A memetic algorithm enables efficient local and global all-atom protein-protein docking with backbone and sidechain flexibility

Abstract

Protein-protein docking plays a central role in the characterization and discovery of protein interactions in the cell. Complex formation is encoded by specific interactions at the atomic scale, but the computational cost of modeling proteins at this level often requires the use of simplified energy models, coarse-grained protein descriptions and rigid-body approximations. In this study we present EvoDOCK, which is an evolutionary-based docking algorithm that enables the identification of optimal docking orientations using an atomistic energy function and sidechain flexibility, employing a global search without prior information of the binding site. EvoDOCK is a memetic algorithm that combines the strength of a differential evolution algorithm for efficient exploration of the global search space with the benefits of a local optimization method, built on the Monte Carlo-based RosettaDOCK program, to optimize detailed atomic interactions. This approach resulted in substantial improvements in both sampling efficiency and computation speed compared to calculations using the local optimization method RosettaDOCK alone, with up to 35 times of reduction in computational cost. For all the ten systems investigated in this study, a highly accurate docking prediction could be identified as the lowest energy model with high efficiency. While protein-protein docking with EvoDOCK is still computationally expensive compared to many methods based on Fast Fourier Transforms (FFT), the results demonstrate the tractability of global docking proteins using an atomistic energy function while exploring sidechain flexibility. Comparison with FFT global docking demonstrated the benefits of using an all-atom energy function to identify native-like predictions. The sampling strategy in EvoDOCK can readily be tailored to include backbone flexibility in the search, which is often necessary to tackle more challenging docking challenges.

Induced fit with replica exchange improves protein complex structure prediction

“…Thus, ReplicaDock 2.0 can perform exhaustive global sampling on the protein energy landscape with better near-native discrimination. One limitation is that global docking with ReplicaDock 2.0 requires 600–800 CPU hours, compared with 35 CPU hours (as reported by Varela et al [ 40 ]) for ClusPro [ 35 ]. Rigid-backbone global docking results from either ClusPro or ReplicaDock 2.0 can serve as the input to a local, flexible-backbone docking search.…”

Development and Application of Computational Models for Peptide-Protein Complexes

“…Although FFT-docking is most commonly used when performing rigid-body docking with distance-based energy functions, genetic algorithms and other optimization methods are also employed [93].…”