idDock+: Integrating Machine Learning in Probabilistic Search for Protein–Protein Docking

Abstract

Predicting the three-dimensional native structures of protein dimers, a problem known as protein-protein docking, is key to understanding molecular interactions. Docking is a computationally challenging problem due to the diversity of interactions and the high dimensionality of the configuration space. Existing methods draw configurations systematically or at random from the configuration space. The inaccuracy of scoring functions used to evaluate drawn configurations presents additional challenges. Evidence is growing that optimization of a scoring function is an effective technique only once the drawn configuration is sufficiently similar to the native structure. Therefore, in this article we present a method that employs optimization of a sophisticated energy function, FoldX, only to locally improve a promising configuration. The main question of how promising configurations are identified is addressed through a machine learning method trained a priori on an extensive dataset of functionally diverse protein dimers. To deal with the vast configuration space, a probabilistic search algorithm operates on top of the learner, feeding to it configurations drawn at random. We refer to our method as idDock+, for informatics-driven Docking. idDock+is tested on 15 dimers of different sizes and functional classes. Analysis shows that on all systems idDock+finds a near-native structure and is comparable in accuracy to other state-of-the-art methods. idDock+ represents one of the first highly efficient hybrid methods that combines fast machine learning models with demanding optimization of sophisticated energy scoring functions. Our results indicate that this is a promising direction to improve both efficiency and accuracy in docking.

Principles and Overview of Sampling Methods for Modeling Macromolecular Structure and Dynamics

“…Machine learning methods, though not the focus of this review, are showing promise in elucidating features of native interaction interfaces so as to bypass the employment of interaction energy functions at a global layer [ 265 – 268 ]. For instance, work in [ 269 ] proposes a learned model to be used as a top filter to label sampled protein-protein dimers before attempting to refine them with more accurate and computationally costly interaction energy functions. Rather than employing information from machine learning models, methods such as HADDOCK [ 243 ], the Integrative Modeling Platform (IMP) [ 270 ] and others [ 271 , 272 ], employ wet-laboratory data to restrict sampling of bound conformations to those that reproduce the wet-laboratory data.…”

A stochastic roadmap method to model protein structural transitions

“…EAs are investigated in detail in our lab in diverse protein modeling scenarios, including de novo structure prediction [74][75][76] and protein-protein docking. [77][78][79] The EA we employ here has been recently proposed 9,10 to further populate the structure space of a protein for which many experimental structures already exist in the Protein Data Bank (PDB). 80 A detailed analysis of the energy landscapes obtained from the algorithm on various H-Ras sequences has also been published recently.…”