Ensemble Neural Networks Scoring Functions for Accurate Binding Affinity Prediction of Protein-Ligand Complexes

摘要

Accurately predicting the binding affinities (BAs) of large sets of protein-ligand complexes (PLCs) is a key challenge in computational biomolecular science, with applications in drug discovery, chemical biology, and structural biology. Since a scoring function (SF) is used to score, rank, and identify drug leads, the fidelity with which it predicts the affinity of a ligand candidate for a protein's binding site has a significant bearing on the accuracy of virtual screening. Despite intense efforts in developing conventional SFs, which are either force-field based, knowledge-based, or empirical, their limited predictive power has been a major roadblock toward cost-effective drug discovery. Therefore, in this work, we present two novel SFs employing a large ensemble of neural networks (NNs) in conjunction with a diverse set of physicochemical and geometrical features characterizing PLCs to predict BA. We build the first ensemble SF, BgN-Score, using the bootstrap aggregation technique, in which hundreds of neural networks are fitted independently to sets of PLCs sampled randomly without replacement from the training data. BgN-Score calculates the final BA score by computing the average of the predictions of the constituent NNs. The boosting approach is used to construct the second ensemble SF, BsN-Score, which is based on an additive stagewise fitting of NNs to random sets of PLCs sampled from the training data. The calculated BA of BsN-Score is a weighted sum of the predictions of the individual NNs in the ensemble. These ensemble SFs are trained on 1105 high-quality PLCs retrieved from the 2007 version of PDBbind. Each PLC in this set is characterized using physicochemical features employed by the empirical SFs X-Score (6 features) and AffiScore (30 features) and calculated by GOLD (14 features), and geometrical features used in the machine-learning SF RF-Score (36 features).