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

INTRODUCTION

• Free-energy calculations can predict binding affinity accurately but require extensive configurational sampling and computation.The introduction identifies simple, accurate, and efficient scoring methods as a drug-design need.Source: INTRODUCTION · p. 1
• CNNs are a deep-learning approach, and prior CNN-based virtual-screening models used sequence, 3D-complex, or graph representations.The introduction also notes that deep learning can learn features from raw data.Source: INTRODUCTION · p. 1
• The authors propose encoding proteins as residues rather than atoms because different residues can give the same elements distinct physical and chemical properties.They describe amino-acid classes as carrying intrinsic features such as polarity and aromaticity.Source: INTRODUCTION · pp. 1–2
• OnionNet-2 uses rotation-free residue–atom contacts across multiple distance shells and improves prediction power by about 3.7% over previous models.The authors present it as a simple 2D CNN-based regression model for protein–ligand interactions.Source: INTRODUCTION · p. 2

Descriptors

• The descriptor counts contacts between specific protein-residue and ligand-atom types across N packed distance shells.The first shell is a sphere of radius d0; subsequent shells have thickness δ.Source: Descriptors · p. 2
• The shell boundaries use a first interval from 0 to d0 and successive intervals of width δ.For shell i ≥ 2, the boundary is d0 + (i − 2)δ ≤ Ki < d0 + (i − 1)δ.Source: Descriptors · p. 2
• Each shell contains counts for residue–ligand atom type combinations, yielding 168 features per shell and 168×N features across N shells.The ligand atoms are classified into eight types; the protein has 20 standard residue types plus OTH for water, ions, and other non-standard residues.Source: Descriptors · p. 2
• Shell membership is determined by the minimum distance between a ligand atom and a protein residue, using d0 = 1 Å and δ = 0.5 Å.Residue–atom distance is measured from the ligand atom to the nearest heavy atom in the residue.Source: Descriptors · p. 2

Evaluation metrics

• The loss is the one adopted in the authors’ previous work.Source: Evaluation metrics · p. 2
• The loss combines Pearson correlation coefficient R and RMSE, with α set to 0.7 to weight the two metrics.R is Pearson correlation; RMSE is root-mean-squared error.Source: Evaluation metrics · pp. 2–3
• Training stops when validation loss changes by less than 0.001 after 20 epochs, and each prediction averages five independent trainings.The model selected for each task is the one that performed best on the validation set.Source: Evaluation metrics · p. 3

The predictive power of OnionNet-2

• R first peaks at N=30, then fluctuates within 0.01 before reaching its global maximum at N=62.The tested shell counts spanned N=10–90; N=30 provides high accuracy at relatively low computational cost.Source: The predictive power of OnionNet-2 · p. 3, Figure 2 · p. 3
• With N=62, predicted and experimental pKa values are highly correlated on the validation set and both core test sets.The re-trained models using older PDBbind versions v.2016 and v.2018 had almost the same R values.Source: The predictive power of OnionNet-2 · p. 3, Figure 3 · p. 3
• OnionNet-2 achieved the highest reported performance: R=0.864 and RMSE=1.164 on core set v.2016, and R=0.821 and RMSE=1.357 on core set v.2013.Source: The predictive power of OnionNet-2 · p. 3, Figure 4 · p. 4
• Despite using residue–atom contacts alone, OnionNet-2 had higher predictive power than the compared 3D convolution scoring functions K deep and Pafnucy.Those models use voxel representations and explicitly include physical properties such as hydrophobicity and hydrogen-bond donor or acceptor status.Source: The predictive power of OnionNet-2 · pp. 3–4

Evaluation of the generalization ability of the model

• On CSAR NRC-HiQ subset 1, OnionNet-2 achieved R=0.89, above K deep (R=0.72) and RF-Score (R=0.78); it also outperformed both models on subset 2.The test subsets do not overlap with the training, validation, or core set v.2016 data.Source: on different test sets · p. 4, Table I · p. 4

structures

• RMSE quickly increased with decoy RMSD, indicating that OnionNet-2 responds to binding-pose changes and distinguishes native structures.The evaluation compares each decoy’s predicted pKa with the corresponding native complex’s pKa.Source: structures · p. 4, Figure 5 · p. 5

the binding pockets on the prediction accuracy

• OnionNet-2 achieved higher prediction accuracy than other scoring functions across the H-, S-, and V-series subsets.The comparison covers subsets classified by hydrophobic scale, buried solvent-accessible area, and excluded pocket volume.Source: the binding pockets on the prediction… · p. 5, Table II · p. 5
• Contact counts in multiple shells may capture pocket hydrophobic scale and reflect ligand burial and excluded volume.The authors suggest these contact features may reflect buried solvent-accessible area and excluded volume, particularly within the binding pocket.Source: the binding pockets on the prediction… · p. 5
• OnionNet-2’s R value was lower on V2 than on other subsets, though still higher than other scoring functions, suggesting less sensitivity to medium-sized pockets.The authors note that recognizing binding-pocket size and shape may remain challenging for current scoring functions.Source: the binding pockets on the prediction… · p. 5
• Predicted pKa showed almost no dependence on H, S, or V values, prompting the authors to suggest that more realistic ligand-in-pocket descriptors may be needed.The authors base this suggestion on scatter plots of predicted and experimental pKa across the descriptor ranges.Source: the binding pockets on the prediction… · p. 5

CONCLUSION

• OnionNet-2 achieved the highest accuracy among previous scoring functions on CASF-2013 and CASF-2016.The conclusion also reports generalization checks on CSAR NRC-HiQ and decoy structures.Source: CONCLUSION · p. 5
• The model uses a 2D convolutional neural network trained on protein-residue and ligand-atom contact-pair counts.The conclusion describes these contact counts as the model’s features for predicting binding free energy.Source: CONCLUSION · p. 5
• The conclusion says OnionNet-2 can recognize pocket-interaction properties including hydrophobic scale, ligand burial, and excluded volume.The listed properties characterize ligand–binding-pocket interactions.Source: CONCLUSION · pp. 5–6

Supporting Information

• The supplement is titled “OnionNet-2: A Convolutional Neural Network Model for Predicting Protein-Ligand Binding Affinity based on Residue-Atom Contacting Shells.”The title appears across two passages.Source: Supporting Information · p. 8
• The supporting-information author list includes Zechen Wang, Liangzhen Zheng, Yang Liu, Yuanyuan Qu, Yong-Qiang Li, Mingwen Zhao, Yuguang Mu, and Weifeng Li.The names are supplied across two passages.Source: Supporting Information · p. 8

PART I. PREPARATION OF DATASET AND ARCHITECTURE

• The supplement begins with a section on preparing training and validation sets and architecture.The supplied heading also identifies a dataset-and-architecture part.Source: PART I. PREPARATION OF DATASET AND… · p. 8

testing sets

• Testing used CASF-2016 and CASF-2013 benchmarks, with CSAR NRC-HiQ as a third test set.The passage describes CASF-2016 as containing 285 complexes and CASF-2013 as containing 195 complexes in 65 clusters.Source: testing sets · p. 8

Architecture

• The model uses 2D convolution to learn the relationship between contact features and binding free energy.The described architecture includes three convolutional layers and two fully connected layers.Source: Architecture · p. 8
• The architecture uses ReLU activations, batch normalization, SGD, and L2 regularization to reduce overfitting.The reported SGD learning rate is 0.001, and the batch size is 64.Source: Architecture · p. 8
• R values were almost the same across PDBbind v.2016, v.2018, and v.2019 despite large differences in database size.The comparison was conducted on CASF-2013 and CASF-2016; the figure caption specifies shell counts of 58, 60, and 62.Source: Architecture · pp. 8–9, Figure S1 · p. 9

• Table S2 compares the predictive power of scoring functions on the CASF-2016 and CASF-2013 core sets.Source: Table S2 · p. 9
• Figure S2 reports Pearson correlations for different scoring functions on the CASF-2013 benchmark.Source: Figure S2 · p. 9
• Figure S3 presents OnionNet-2 results on CSAR-HiQ subsets.Source: Figure S3 · p. 9

PART IV.THE DETAILED INFORMATION OF THE GENERATION OF DECOYS

• AutoDock Vina generated non-native poses in a 27 Å × 27 Å × 27 Å sampling cube centered on the native ligand pose, with exhaustiveness set to 12.Because Vina ignores water and ion effects, receptor-decoy complexes omitted water molecules and ions.Source: PART IV.THE DETAILED INFORMATION OF… · p. 9
• Up to 20 decoys were generated per receptor, with fewer possible when the binding pocket's size and shape limited sampling.Each decoy's RMSD relative to the native structure was calculated.Source: PART IV.THE DETAILED INFORMATION OF… · p. 9

2. We used 10 RMSD intervals, [0 ˚A, 2 ˚A], [2 ˚A, 3 ˚A],

• Ten RMSD intervals defined the subsets, and the decoy with the smallest RMSD was selected for each ligand in each interval.Source: §2 · p. 10
• The resulting ten test subsets contained non-experimental complexes and were used for OnionNet-2 training.Source: §2 · p. 10