Improving Structure-Based Virtual Screening with Ensemble\nDocking and Machine Learning

Abstract

One of the main challenges of structure-based\nvirtual screening\n(SBVS) is the incorporation of the receptor’s flexibility,\nas its explicit representation in every docking run implies a high\ncomputational cost. Therefore, a common alternative to include the\nreceptor’s flexibility is the approach known as ensemble docking.\nEnsemble docking consists of using a set of receptor conformations\nand performing the docking assays over each of them. However, there\nis still no agreement on how to combine the ensemble docking results\nto obtain the final ligand ranking. A common choice is to use consensus\nstrategies to aggregate the ensemble docking scores, but these strategies\nexhibit slight improvement regarding the single-structure approach.\nHere, we claim that using machine learning (ML) methodologies over\nthe ensemble docking results could improve the predictive power of\nSBVS. To test this hypothesis, four proteins were selected as study\ncases: CDK2, FXa, EGFR, and HSP90. Protein conformational ensembles\nwere built from crystallographic structures, whereas the evaluated\ncompound library comprised up to three benchmarking data sets (DUD,\nDEKOIS 2.0, and CSAR-2012) and cocrystallized molecules. Ensemble\ndocking results were processed through 30 repetitions of 4-fold cross-validation\nto train and validate two ML classifiers: logistic regression and\ngradient boosting trees. Our results indicate that the ML classifiers\nsignificantly outperform traditional consensus strategies and even\nthe best performance case achieved with single-structure docking.\nWe provide statistical evidence that supports the effectiveness of\nML to improve the ensemble docking performance.