Improving quality of ligand-binding site prediction with Bayesian optimization

Ligand binding site prediction from protein structure plays an important role in various complex rational drug design efforts. Its applications include drug side effects prediction, docking prioritization in inverse virtual screening and elucidation of protein function in genome wide structural studies. Currently available tools have limitations that disqualify them from many possible use cases. In general they are either fast and relatively inaccurate (e.g. purely geometric methods) or accurate but too slow for large scale applications (e.g. methods that rely on a large template libraries of known protein-ligand complexes). P2Rank is a recently introduced machine learning based method that have already exhibited speeds comparable to fastest geometric methods while providing much higher identification success rates. Here we present an improved version that brings speed-up as well as higher quality predictions. A leap in predictive performance was achieved thanks to the technique of Bayesian optimization, which allowed simultaneous optimization of numerous arbitrary parameters of the algorithm. We have evaluated our method with respect to various performance and prediction quality criteria and compared it to other state of the art methods, as well as to it's previous version, with encouraging results.

I. Introduction

Ligand binding site prediction from protein structure has many applications related to rational drug design. It can find employment in various tasks such as drug side-effects prediction, docking prioritization, structure based virtual screening and structure-based target prediction. Increasingly it can be seen applied in genome-wide structural studies that try to analyze and compare all known and putative binding sites. Many of those use cases imply the need for fast standalone tool that can be used as a stable part of larger pipeline. This disqualifies many currently available tools that are available only as web servers and/or are simply too slow.