OnionNet-2: A Convolutional Neural Network Model for Predicting Protein-Ligand Binding Affinity Based on Residue-Atom Contacting Shells

Abstract

One key task in virtual screening is to accurately predict the binding affinity (△G) of protein-ligand complexes. Recently, deep learning (DL) has significantly increased the predicting accuracy of scoring functions due to the extraordinary ability of DL to extract useful features from raw data. Nevertheless, more efforts still need to be paid in many aspects, for the aim of increasing prediction accuracy and decreasing computational cost. In this study, we proposed a simple scoring function (called OnionNet-2) based on convolutional neural network to predict △G. The protein-ligand interactions are characterized by the number of contacts between protein residues and ligand atoms in multiple distance shells. Compared to published models, the efficacy of OnionNet-2 is demonstrated to be the best for two widely used datasets CASF-2016 and CASF-2013 benchmarks. The OnionNet-2 model was further verified by non-experimental decoy structures from docking program and the CSAR NRC-HiQ data set (a high-quality data set provided by CSAR), which showed great success. Thus, our study provides a simple but efficient scoring function for predicting protein-ligand binding free energy.

Protein–Ligand Docking in the Machine-Learning Era

“…This method used a CNN model with inputs based on rotation-free element-pair specific contacts between protein and ligand in different shells. OnionNet, as well as the subsequent OnionNet-2.0 (2021) [ 119 ], achieved excellent scoring performance on the CASF-2016 benchmark.…”

Improving protein–ligand docking and screening accuracies by incorporating a scoring function correction term

“…Their learning labels were artificial binding affinities calculated as per the docking pose root mean square deviation (RMSD) and Vina scores with the hypothesis that these poses having larger RMSD should be labeled to have a lower binding affinity [ 40 ]. Inspired by these prior studies and our own DL exercises [ 19 , 20 ], we suggest that the robustness of Vina score could be optimized by augmenting a ML-based score term. Accordingly, the screening power and power to differentiate the ‘good’ molecules from ‘bad’ ones could be enhanced.…”

“…Both ML and DL models have the ability to capture high-level complexities in protein–ligand complexes and have the potential to formulate a powerful scoring function for more practical docking and screening tasks [ 20 , 21 , 29 ]. Previous models such as OnionNet [ 19 ], OnionNet2 [ 20 ], Pafnucy [ 34 ], Kdeep [ 35 ] and RosENet [ 23 ] have been reported to have high scoring and ranking scores; however, their performance on virtual screening is not as good as expected [ 28 , 29 ], possibly because of the over-fitting to the experimental complex structures datasets. A recent docking application Gnina [ 22 , 36 ] adopts a 3D convolutional neural network (CNN) model [ 37 ] to provide a score to the docking poses.…”

“…Unlike what the authors did for developing Δ vina XGB and Δ vina RF20, we did not apply artificial labels in model training; instead, only the RMSD values were used as labels [ 38 ], and the trained model works as a scoring function correction term (SFCT). The input features of this model are the multiple layers of protein–ligand intermolecular contacts (between protein residues and ligand atoms) [ 19 , 20 , 43 ]. The performance of adding OnionNet-SFCT term to the Vina score is verified by docking (in particular redocking and cross-docking) tasks, virtual screening tasks and reverse docking tasks.…”

Delta Machine Learning to Improve Scoring-Ranking-Screening Performances of Protein–Ligand Scoring Functions

“…As can be seen, our Δ Lin_F9 XGB model ranks 6 th among these start-of-the-art models. The top 5 performers are graphDelta 17 (graph-convolutional neural network model, Pearson's R = 0.87), ECIF::LD-GBT 18 (gradient boosting tree model incorporating extended connectivity interaction features and RDKit ligand features, Pearson's R = 0.866), OnionNet-2 15 (convolutional neural network model with inputs based on rotation-free specific contacts between the protein and the ligand in different shells, Pearson's R = 0.864), TopBP 12 (a consensus model incorporating different ML methods and with inputs based on the algebraic topology for characterizing biomolecular complexes, Pearson's R = 0.861), and ECIF::GBT 18 (gradient boosting tree model incorporating only extended connectivity interaction features, Pearson's R = 0.857). Other methods, such as persistent spectral-based ML models (Mol-PSI 13 and PerSpect ML 14 ), the algebraic graph theory-based model (AGL-Score 19 ), and the usage of diverse ligand-based features in the previous ML model (RFScore v3+RDKit 22 ), also show very good scoring power in the CASF- Moreover, we also compared with several ML scoring functions that were evaluated with at least three different metrics for the CASF-2016 benchmark.…”