Source-linked AI summary
OnionNet: a multiple-layer inter-molecular contact based convolutional neural network for protein-ligand binding affinity prediction
Liangzhen Zheng, Jingrong Fan, Yuguang Mu
TL;DR
Protein–ligand binding-affinity scoring functions lack sufficient accuracy for large-scale lead screening. OnionNet uses rotation-free, element-pair-specific contacts across distance ranges in a CNN, achieving stronger benchmark performance and comparable predictions from docked complexes. Its reported CASF-2013 performance is R 0.78 and RMSE 1.503, while the authors identify general lead-discovery suitability as a future scope.
Problem
Existing binding-affinity scoring functions are not accurate enough, motivating improved computational prediction for large-scale lead screening.
Method
OnionNet is a CNN based on rotation-free element-pair-specific protein–ligand contacts grouped across distance ranges to represent local and non-local interactions.
Results
R 0.78 and RMSE 1.503 are reported for OnionNet on the CASF-2013 benchmark, with better performance than current deep-learning and classic scoring functions.
Takeaways & Limitations
OnionNet’s stability and robustness were verified through retraining and predictions using docked protein–ligand complexes.
Takeaways & Limitations
The authors state that the model’s general suitability for lead-discovery tasks remains a future consideration.
Abstract
from arXiv · showhide
Computational drug discovery provides an efficient tool helping large scale lead molecules screening. One of the major tasks of lead discovery is identifying molecules with promising binding affinities towards a target, a protein in general. The accuracies of current scoring functions which are used to predict the binding affinity are not satisfactory enough. Thus, machine learning (ML) or deep learning (DL) based methods have been developed recently to improve the scoring functions. In this study, a deep convolutional neural network (CNN) model (called OnionNet) is introduced and the features are based on rotation-free element-pair specific contacts between ligands and protein atoms, and the contacts were further grouped in different distance ranges to cover both the local and non-local interaction information between the ligand and the protein. The prediction power of the model is evaluated and compared with other scoring functions using the comparative assessment of scoring functions (CASF-2013) benchmark and the v2016 core set of PDBbind database. When compared to a previous CNN-based scoring function, our model shows improvements of 0.08 and 0.16 in the correlations (R) and standard deviations (SD) of regression, respectively, between the predicted binding affinities and the experimental measured binding affinities. The robustness of the model is further explored by predicting the binding affinities of the complexes generated from docking simulations instead of experimentally determined PDB structures.
1. Introduction
Existing scoring functions are insufficiently accurate or efficient for large-scale protein–ligand binding-affinity prediction. OnionNet addresses these limitations with rotation-invariant, element-specific contacts across distance ranges and performs strongly on benchmark datasets and docked poses.
- Motivation and existing methods: Current binding-affinity scoring functions can be inaccurate, while accurate physics-based alternatives may be too computationally expensive for large-scale screening.Docking scores have high false-positive rates; MMPBSA and alchemical methods involve accuracy, absolute-value, time, or resource limitations.
- Motivation and existing methods: Pafnucy’s localized 20 Å grid may not fully capture long-range electrostatic interactions, and empirically calculated partial charges may introduce noise.These limitations motivate features that extend beyond a ligand-centered grid while reducing dependence on potentially inaccurate charge estimates.
- OnionNet approach: OnionNet represents protein–ligand complexes using rotation-invariant, element-specific intermolecular contacts grouped into distance ranges covering local and non-local interactions.The resulting multiple-layer intermolecular features are used to train a CNN on PDBbind v2016 data.
- Evaluation and results: OnionNet achieves RMSE values of 1.278 and 1.503 on the PDBbind v2016 core set and CASF-2013, respectively, versus Pafnucy’s 1.42 and 1.69.Its correlation coefficients are R=0.812 and R=0.786 on the two benchmark datasets, and docked-pose predictions are comparable to predictions from experimentally determined structures.
2. Methods and materials
OnionNet featurizes protein–ligand complexes with element-pair contact counts across concentric distance shells, then predicts binding affinity with a deep CNN. The model uses disjoint PDBbind-derived datasets and is evaluated on benchmark and docking-generated complexes.
- 2.1 Featurization of protein-ligand complexes: Element-pair-specific contacts are calculated between ligand and protein atoms within multiple concentric shells.The shell design captures interaction information across different distance ranges.
- 2.1 Featurization of protein-ligand complexes: Eight element types produce 64 protein–ligand element-pair combinations for each shell.The categories include C, N, O, H, P, S, halogens, and a dummy category for remaining elements.
- 2.1 Featurization of protein-ligand complexes: 60 shells with d0=1.0 Å and δ=0.5 Å yield 3840 features spanning distances to 30.5 Å.This representation covers both local and non-local protein–ligand interactions.
- 2.3 Deep neural network model: A modified deep CNN is trained with a customized loss combining correlation and RMSE, with α=0.8 used in this study.The reported model achieves the best performance among the explored configurations and shows prediction accuracy with quick convergence on validation data.
- 2.3 Deep neural network model: The model converts each complex’s 3840 features into a (64, 60, 1) tensor and processes it with convolutional and dense layers.The final layer outputs the predicted pKa value.
3. Results
OnionNet was evaluated using RMSE, SD, MAE, and R, with strong correlations between predicted and measured pKa values across validation and testing sets. The model’s accumulated-error curves also indicate that many predictions have deviations near 1.0 pKa units.
- Model selection: The best model was selected at epoch 89 because it achieved minimal validation loss.Training monitored customized loss, RMSE, and R.
- Evaluation: The evaluation used RMSE, SD, MAE, and Pearson correlation coefficient R as prediction-accuracy metrics.These metrics quantify agreement between predicted and experimentally determined pKa values.
- Prediction performance: RMSE values were 1.287, 1.278, and 1.503 for the validating set and two testing sets, while correlations exceeded 0.7.The predicted and measured pKa values were highly correlated across all three datasets.
- Prediction performance: Approximately 60% of validation samples and 50% of testing samples had pKa deviations of about 1.0 from measured values.The accumulated absolute error curves were used to characterize these deviations.
4. Discussion
The discussion attributes OnionNet’s performance to simple, rotation-free distance-based contact features that represent both local and non-local interactions. Comparisons, feature-removal analyses, and docking-pose tests support its accuracy, stability, and robustness, while revealing specific interaction ranges and element pairs as especially informative.
- Feature design: Distance-based contact features reduce feature complexity and bias while covering local and non-local protein–ligand interactions with rotationally invariant coordinates.The features use atom element types and distance-based contacts rather than more assumption-heavy descriptors.
- Performance comparison: OnionNet achieved higher accuracy than popular classic scoring functions, and it outperformed Pafnucy on the CASF-2013 comparisons.The reported advantages included larger R and smaller RMSE, MAE, and SD than Pafnucy on the two testing sets.
- Performance comparison: OnionNet obtained the best correlations and improved SD by 0.16 versus Pafnucy on the second testing set.The comparison concerns predicted pKa versus experimentally measured pKa on the v2013 core testing set.
- Feature importance: The O_P element pair was most important, followed by C_S, while several oxygen-, sulfur-, nitrogen-, and phosphorus-containing contacts also contributed.Protein carbon–ligand hydrogen contacts made minor contributions.
- Stability and robustness: Removing one element-pair combination or one contact shell caused no great performance decrease, indicating model stability.Independent retraining also produced relatively small standard deviations, and a t-test rejected the hypothesis that mean R was no higher than 0.7.
- Stability and robustness: For 219 native-like docked complexes, OnionNet produced R=0.755 and SD=1.523, with performance slightly worse than for experimentally determined structures.A native-like pose was defined by an RMSD below 2 Å from the native conformation.
- Stability and robustness: Predictions from a native-like docking pose closely matched crystal-structure predictions, with values of 3.421 and 3.436, respectively.The corresponding native-pose experimental binding affinity was 2.29.
Conclusion
The paper concludes that OnionNet uses multiple-layer intermolecular contact features to improve protein–ligand binding-affinity prediction. It reports stronger benchmark performance and stability under missing features and docking-pose inputs, while noting that further improvement is needed for general lead-discovery use.
- Conclusion: OnionNet is based on multiple-layer intermolecular contact features for ligand binding-affinity prediction.The model was proposed to improve ligand binding-affinity prediction.
- Conclusion: On CASF-2013, OnionNet achieved R=0.78 and RMSE=1.503, outperforming current deep-learning-based and classic scoring functions.CASF-2013 was used as the benchmark.
- Conclusion: The model’s stability and robustness were verified by retraining with missing features and predicting affinities from docking poses.These tests extended evaluation beyond the original complete experimental-structure inputs.
- Conclusion: Further improvement would be needed before OnionNet is suitable for general lead-discovery tasks.The conclusion states this as a condition for broader applicability.
Corresponding Author
The supplied passage identifies the authors’ institutional address.
- Corresponding Author: The authors are affiliated with the School of Biological Sciences at Nanyang Technological University.The address is given as 60 Nanyang Drive.