Predicting the native poses of ligands correctly is one of the most important steps towards successful structure-based drug design. Binding affinities (BAs) estimated by traditional scoring functions (SFs) are typically used to score and rank-order poses to select the most promising conformation. This BA-based approach is widely applied and some success has been reported, but it is inconsistent and still far from perfect. The main reason for this is that SFs are trained on experimental BA values of only native poses found in co-crystallized structures of protein-ligand complexes (PLCs). However, during docking, they are needed to discriminate between native and decoy poses, a task for which they have not been specifically designed. To overcome this limitation, we propose to build task-specific SFs that model binding affinities (scoring task) as well as conformations (docking task) using the root mean square deviation (RMSD) of a ligand pose from the native pose. Our models are based on boosted ensembles of neural networks and other state-of-the-art machine learning (ML) algorithms in conjunction with multi-perspective interaction modeling techniques for accurate characterization of PLCs. We assess the docking and scoring/ranking accuracies of the proposed ML SFs as well as three conventional SFs in the context of the 2014 CSAR benchmark exercise that encompasses three high-quality protein systems and a diverse set of drug-like molecules. Our proposed docking-specific SFs provide a substantial improvement in the docking task. We find that RMSD-based SFs for BsN, an ensemble neural networks (NN) model based on boosting, and six other ML models provide more than 120% improvement, on average, over their BA-based counterparts. In terms of scoring/ranking accuracy, we find that the approach of using RMSD-based BsN to select the top ligand pose followed by applying BA-based BsN to rank ligands using predicted BA scores leads to consistent and correctly ranked ligands for the two protein targets Spleen Tyrosine Kinase (SYK) and tRNA (m1G37) methyltransferase (TrmD). In addition, the ensemble NN SF BsN is at least 250% more accurate than a single neural network (SNN) model. We further find that ensemble models based on NNs surpass SFs based on other state-of-the-art ML algorithms such as BRT, RF, SVM, and [Formula: see text]NN. Finally, our RF model fitted to PLCs characterized by multiple sets of descriptors from four different sources (X-Score, AffiScore, RF-Score, and GOLD) substantially outperforms the SF RF-Score that uses only one set of features, underlining the value of multi-perspective modeling.
Keywords: CSAR; Protein-ligand binding affinity; descriptors; interaction modeling; machine learning; neural networks.