Source-linked AI summary
Improved Protein-ligand Binding Affinity Prediction with Structure-Based Deep Fusion Inference
Derek Jones, Hyojin Kim, Xiaohua Zhang, Adam Zemla, Garrett Stevenson, William D. Bennett, Dan Kirshner, Sergio Wong, Felice Lightstone, Jonathan E. Allen
TL;DR
Accurate protein–ligand binding-affinity prediction remains difficult, while existing deep models depend on structural representations with distinct limitations. The paper fuses 3D-CNN and SG-CNN representations and reports improved overall performance over individual models, including on docking poses, with greater computational efficiency than related biophysics-based scoring.
Problem
Accurate protein–ligand binding-affinity prediction remains challenging despite biophysics-based scoring methods and deep learning approaches with representation-dependent limitations.
Method
The paper combines complementary feature representations and predictions from independently trained 3D-CNN and SG-CNN models using mid-level and late fusion.
Results
Fusion models improve overall prediction compared with individual neural-network models and maintain improved accuracy over docking and MM/GBSA scores for docked poses.
Takeaways & Limitations
The Fusion model is presented as a more computationally efficient alternative to MM/GBSA rescoring, with robust prediction on new ligands in recognized pockets.
Takeaways & Limitations
Prediction accuracy drops on docked poses, indicating sensitivity to incorrectly scored poses; the authors propose jointly classifying correct and incorrect poses as future improvement.
Abstract
from arXiv · showhide
Predicting accurate protein-ligand binding affinity is important in drug discovery but remains a challenge even with computationally expensive biophysics-based energy scoring methods and state-of-the-art deep learning approaches. Despite the recent advances in the deep convolutional and graph neural network based approaches, the model performance depends on the input data representation and suffers from distinct limitations. It is natural to combine complementary features and their inference from the individual models for better predictions. We present fusion models to benefit from different feature representations of two neural network models to improve the binding affinity prediction. We demonstrate effectiveness of the proposed approach by performing experiments with the PDBBind 2016 dataset and its docking pose complexes. The results show that the proposed approach improves the overall prediction compared to the individual neural network models with greater computational efficiency than related biophysics based energy scoring functions. We also discuss the benefit of the proposed fusion inference with several example complexes. The software is made available as open source at https://github.com/llnl/fast.
1 Introduction
Accurate protein–ligand binding-affinity prediction is a fundamental drug-development challenge. Deep learning offers an alternative to physics-based scoring by learning interaction rules from atomic representations, but 3D-CNNs and SG-CNNs encode structure differently and have distinct limitations.
- Binding-affinity prediction between small molecules and target proteins remains a fundamental challenge in drug development.
- Deep learning models can learn binding interaction rules directly from atomic representations without relying on hand-curated features.
- 3D-CNNs represent atoms in a 3D voxel grid and implicitly model pairwise relationships through relative atomic positions.
- The voxel representation requires learning many parameters because it does not predetermine which atomic interactions to represent beyond minimum atomic resolution.
2 Materials and Methods
The study uses curated PDBBind protein–ligand complexes, standardized atomic preprocessing, and non-overlapping training and testing subsets. Docking complexes are generated and rescored with Vina and MM/GBSA, while fusion combines model representations for evaluation.
- 2.1 Data: The general, refined, and core subsets are organized using complex, binding-data, and crystal-structure quality criteria.
- 2.1 Data: PDBBind 2016 contains 13,308 general, 4,057 refined, and 290 core protein–ligand complexes.
- Docking complexes: Docking poses are generated with Vina and the top 10 poses are rescored using MM/GBSA.
- Preprocessing: The preprocessing pipeline protonates complexes, assigns charges, removes water from docking holdouts, extracts atomic features, and centers coordinates on each ligand.
- Fusion: The proposed fusion architecture combines the 3D-CNN and SG-CNN representations through separate fusion layers.
- 2.1 Data: The core set is held out for testing, while the remaining general and refined complexes provide training data with validation splits designed to address potential bias.
2.2 3D-CNN
The 3D-CNN converts atom-level complex features into a volumetric representation and processes them with convolutional layers and residual blocks to predict binding affinity.
- 3D-CNN methods voxelize atoms and their features into a 3D volume to capture atomic features and implicit interactions.
- The input volume is 48×48×48 with 19 atomic-feature channels, covering approximately 48 Å per dimension at 1 Å voxel resolution.
- Atoms occupy voxels according to their van der Waals radii, colliding features are added element-wise, and Gaussian blur with σ = 1 spreads features to neighboring voxels.
- The network uses five convolutional layers and two residual blocks, which improved prediction performance relative to the version without residual blocks and earlier models.
2.3 SG-CNN
The SG-CNN represents atoms as graph nodes and encodes covalent and non-covalent spatial relationships using distance-based adjacency thresholds and specialized propagation layers.
- The spatial graph formulation relaxes covalency requirements by allowing local neighborhoods based on Euclidean distance.
- The SG-CNN models atoms as graph nodes and represents covalent and non-covalent bonds through a spatial graph representation.
- Adjacency entries encode interatomic Euclidean distances, with covalent and non-covalent neighborhoods defined by thresholds αc = 1.5 Å and αnc = 4.5 Å.
2.4 Fusion
The paper fuses heterogeneous feature representations from 3D-CNN and SG-CNN models to improve protein-ligand binding affinity prediction. It uses mid-level and late fusion, with the former combining intermediate activations through learned fully connected layers.
- Fusion motivation: Fusion models combine complementary feature representations from multiple neural networks, potentially improving prediction and reducing uncertainty associated with individual representations.The approach is motivated by fusion methods used with multimodal images and different sensors.
- Fusion strategy: The proposed approach adopts mid-level and late fusion to combine representations from the 3D-CNN and SG-CNN models.The paper describes this as the first attempt to combine multiple model representations for protein-ligand binding affinity prediction.
- Mid-level fusion architecture: The mid-level model extracts activations from late layers of both networks, transforms them with fully connected layers, and concatenates them with the original features before further processing.The architecture uses the second- and fourth-last layer outputs from the 3D-CNN and SG-CNN, respectively, with cross-layer connections.
2.5 Structure based clustering
The evaluation clusters PDBBind complexes by whole-protein and binding-site structural similarity to assess model performance across related local environments. Binding-pocket templates are compared using an all-atom structural similarity metric across thousands of complexes.
- Clustering procedure: Binding pockets are clustered at whole-protein and local-substructure levels to compare structurally related protein-ligand complexes.Local regions surrounding each ligand are delineated from an initial 12.0 Å sphere for structure-based evaluation.
- Structural similarity metrics: The clustering uses LGA for structural grouping and GDC to measure similarity between superimposed binding-site regions.GDC evaluates the conformation of all atoms in the compared substructures rather than only Cα positions.
- Evaluation context: Table 1 compares fusion approaches with individual and existing models on the PDBBind 2016 core crystal-structure set, distinguishing refined-set and general-set training with R and G labels.The supplied table caption identifies the evaluation set and the meanings of the training abbreviations.
- Physics-based comparison: Table 2 compares the mid-level fusion model with physics-based scoring functions using 243 core-set complexes scored successfully by all methods.Vina and MM-GBSA correlation coefficients are reported as absolute values.
- Dataset coverage: 4,464 protein-ligand complexes with binding-affinity scores were included in the structure-similarity search.Ligand-pocket templates were searched against local regions from all PDBBind complexes.
3 Results
The results evaluate complementary CNN information, fusion gains, robustness to docked poses, and comparison with MM/GBSA scoring. Fusion improves over individual neural-network models, while maintaining strong performance on docking and classification tasks.
- Evaluation hypotheses: The PDBBind 2016 core dataset was used to test four hypotheses concerning CNN complementarity, fusion gains, docked-pose accuracy, and comparison with MM/GBSA.These hypotheses frame the reported experiments.
- Crystal structures: The SG-CNN trained on general plus refined data performed best among SG-CNN variants, whereas the refined-only 3D-CNN achieved the highest validation and test accuracy for 3D-CNN.General-set training provides more samples but includes noisier affinity measurements and lower-resolution structures.
- Crystal structures: The proposed mid-level fusion model outperformed individual models, while late fusion by averaging also improved accuracy and differed only marginally from mid-level fusion.The fusion inputs were SG-CNN trained on general plus refined data and refined-only 3D-CNN.
- Crystal structures: 0.803 versus 0.647: the fusion model had a higher Pearson correlation coefficient than MM-GBSA scoring on crystal structures.This comparison concerns the correlation coefficient reported for the fusion model and MM-GBSA.
- Docking poses: 0.685 versus 0.629 and 0.616: on docking poses, fusion had higher Pearson correlation than MM/GBSA and Vina scores.The analysis scored the top 10 Vina poses and retained the highest predicted affinity for each of 257 test complexes with completed MM/GBSA calculations.
- Classification: ROC AUC values of 0.82 and higher were obtained for classifying non-binders with pKi/pKd < 5 and binders with pKi/pKd > 8.These thresholds evaluate recognition at both extremes of the affinity distribution.
4 Discussion
Prediction error varies substantially across structure-based clusters, while the two constituent models show distinct strengths and the Fusion model does not achieve the lowest MAE in every cluster.
- Cluster-dependent performance: MAE varies across structure-based clusters, exceeding 2 log units in some cases.Figure 3 groups complexes by pocket and ligand positioning and displays MAE with standard deviation.
- Model-specific strengths: Six clusters show consistently lower error for SG-CNN, whereas four clusters show consistently lower error for 3D-CNN.The model advantage reverses across identified clusters, indicating distinct performance profiles.
- Model-specific strengths: Figure 4 highlights eight compounds with the largest prediction differences between the two models.The examples include the top four cases where error is lower for either 3D-CNN or SG-CNN.
- Fusion behavior: The Fusion model performs better overall in more clusters than its constituent models but does not attain the lowest MAE in every cluster.This result preserves cluster-level variation despite the Fusion model’s broader advantage.
5 Conclusion
The study concludes that combining complementary CNN predictions improves overall binding-affinity prediction, including for docked poses, while cluster structure helps characterize prediction difficulty and applicability.
- Fusion models: The Fusion models improve overall performance compared with the individual CNN models by combining complementary predictions.The two CNNs differ in their representations: 3D-CNN may help when molecular shape matters, while SG-CNN explicitly represents pairwise interactions with fewer parameters.
- Docked-pose evaluation: Docked poses reduce prediction accuracy, but Fusion models remain more accurate than docking and MM/GBSA scores.The authors attribute the reduction to greater sensitivity to incorrectly scored poses.
- Structural clustering: Structure-based pocket clustering exposes variation in prediction error and can classify new docked poses or identify complexes outside existing cluster types.Different cluster categories may represent binding interactions with different prediction difficulty.
- Conclusion: Prediction error appears robust for new ligands in recognized pockets, and the authors expect accuracy to improve as experimental data increases.The paper concludes that Fusion could become a more computationally efficient alternative to MM/GBSA rescoring.
Funding
The work was supported by the American Heart Association and performed at Lawrence Livermore National Laboratory under a U.S. Department of Energy contract.
- Funding: The study received support from the American Heart Association and the U.S. Department of Energy through Lawrence Livermore National Laboratory.The work was conducted under Contract DE-AC52-07NA27344.
Spatial Graph Convolutional Network Architecture
The SG-CNN propagates information through separate covalent and non-covalent interaction layers, aggregates ligand-node features, and predicts binding affinity with a fully connected network. Its performance analysis also examines model bias and comparisons across scoring methods and screening tasks.
- Spatial Graph Convolutional Network Architecture: SG-CNN uses propagation, aggregation, and a fully connected output network to pass messages between atoms and compute binding-affinity predictions.The propagation layer performs repeated message passing before node features are gathered and sent to the output network.
- Spatial Graph Convolutional Network Architecture: The architecture separates covalent and non-covalent propagation, using inter-node Euclidean distance as the scalar edge feature in both layers.Both propagation types use two message-passing rounds; non-covalent propagation starts from covalent-attention features.
- Spatial Graph Convolutional Network Architecture: Non-covalent features are summed across ligand nodes, flattened, and passed through the output network to produce the binding-affinity prediction ˆy.The ligand subgraph defines the nodes included in this aggregation.
- Performance Analysis: Fusion methods show improved Pearson correlation with experimental log(ki/kd) compared with physics-based scoring methods on 242 PDBBind 2016 core complexes.The comparison includes Mid-level Fusion, Late Fusion, MM/GBSA, and Vina scores.
- Performance Analysis: For bind detection, MM/GBSA slightly outperforms Fusion, while Fusion has the highest no-bind ROC AUC but only a negligible margin over Vina.The reported bind-detection improvement over Vina for both methods is 0.04; the no-bind margin between Fusion and Vina is 0.011.