Source-linked AI summary
DeepSurf: A surface-based deep learning approach for the prediction of ligand binding sites on proteins
Stelios K. Mylonas, Apostolos Axenopoulos, Petros Daras
TL;DR
Identifying potential protein binding sites is an important preliminary step in drug discovery. DeepSurf combines a surface-based representation with local 3D voxel grids and a 3D convolutional network. It outperformed competing deep-learning methods on localization and overlap accuracies across three diverse datasets, while remaining competitive with non-data-driven methods.
Problem
Identifying potentially druggable binding sites on proteins is an important preliminary step toward drug discovery.
Source: §1 · p. 1Method
DeepSurf places local 3D voxel grids on protein-surface points and processes their chemical features with a 3D convolutional network.
Source: §1 · p. 2, §6 · p. 15Results
DeepSurf outperformed competing deep-learning methods in localization and overlapping accuracies across three diverse datasets and was competitive with non-data-driven methods.
Source: §6 · p. 15Takeaways & Limitations
DeepSurf reaches maximum performance by preserving only the points the network assigns most reliably.
Source: §6 · p. 16Incomplete training data and the BSP task’s inherent peculiarities may limit performance; sites without known ligands may be labeled negative despite being positive in practice.
Source: §5.5 · p. 14
Abstract
from arXiv · showhide
The knowledge of potentially druggable binding sites on proteins is an important preliminary step towards the discovery of novel drugs. The computational prediction of such areas can be boosted by following the recent major advances in the deep learning field and by exploiting the increasing availability of proper data. In this paper, a novel computational method for the prediction of potential binding sites is proposed, called DeepSurf. DeepSurf combines a surface-based representation, where a number of 3D voxelized grids are placed on the protein's surface, with state-of-the-art deep learning architectures. After being trained on the large database of scPDB, DeepSurf demonstrates superior results on three diverse testing datasets, by surpassing all its main deep learning-based competitors, while attaining competitive performance to a set of traditional non-data-driven approaches.
1 Introduction
Predicting protein binding sites can support drug discovery, but existing approaches use varied geometric, energetic, template-based, and deep-learning representations. DeepSurf combines local surface-centered voxel grids with a 3D residual network and reports superior performance among state-of-the-art approaches.
1 Introduction
- Structure-based binding-site prediction supports drug discovery by estimating in silico properties that would otherwise require substantial experimental effort.The paper describes knowledge of potential small-compound binding sites as a preliminary step in structure-based drug discovery.Source: §1 · p. 1
- Existing structure-based methods include geometry-based cavity detection, energy-based probe interactions, template alignment, and consensus approaches.Examples include ConCavity, Fpocket, FTSite, Findsite, and metaPocket2.0.Source: §1 · p. 1
- P2Rank uses a random-forest classifier to score solvent-accessible surface points from local chemical and geometric features, then clusters high-scoring points into sites.Source: §1 · p. 2
- DeepSite scores feature-based subgrids with a shallow CNN, while FRSite detects binding sites and Kalasanty segments them in whole-protein voxel grids.Kalasanty was reported to achieve higher accuracies than DeepSite.Source: §1 · p. 2
- Whole-protein voxelization in DeepSite and Kalasanty can neglect surface morphology and lose information through fixed discretization, motivating a surface-based representation.The authors cite as reporting worse results when focusing on grid points or atoms rather than surface points.Source: §1 · p. 2
- DeepSurf combines local voxel grids centered on protein-surface points with a three-dimensional LDS-ResNet and reports superior performance among state-of-the-art approaches.The paper presents the network as an extension of LDS-ResNet for volumetric data.Source: §1 · p. 2
2 Proposed method
DeepSurf predicts ligandability at sampled protein-surface points using locally aligned voxel grids and a 3D CNN, then clusters high-scoring points into binding sites. Its method combines surface-point preprocessing with a 3D residual architecture that includes an LDS module.
2.1 DeepSurf
- The pipeline simplifies the solvent-accessible surface with K-means and places local grids at representative points to reduce redundant computations.The number of clusters is set to the original point count divided by f; the closest point to each cluster center is retained.Source: §2.1 · p. 3
- Each local grid is oriented with its z-axis parallel to the surface normal, although this alignment does not eliminate rotation sensitivity.The grids are 16 × 16 × 16 at 1 Å resolution, and random rotations are still applied during training.Source: §2.1 · p. 3, Figure 1 · p. 4
- Each voxel receives 18 atom features, and a 3D CNN scores its surface point before thresholded points are clustered and assigned to nearby protein atoms.Mean-shift is used to cluster the retained points without specifying the number of clusters in advance.Source: §2.1 · pp. 3–4
2.2 Network architectures
- The 3D bottleneck LDS-ResNet adds an LDS-module branch parallel to the original convolutional branch and concatenates their outputs.The paper also uses an 18-layer 3D-ResNet as a baseline.Source: §2.2 · p. 4, Figure 2 · p. 5
2.2.1 Bottleneck 3D-LDS-Resnet
- The LDS module iteratively processes 3D patches, applies a hidden-state transition, then maps the resulting volume through learned convolutional filters.The stride factor k controls downsampling; k = 1 preserves spatial dimensions, while k > 1 downsamples them by k.Source: §2.2.1 · pp. 4–5
3 Materials
DeepSurf is trained and validated using scPDB structures divided into five protein-separated folds. Its evaluation uses three test datasets with different structural compositions and coverage.
3 Materials
- The 2017 scPDB release provided 15,182 retained structures for training and validation after exclusions from its original 16,034 entries.The release contained 4,782 proteins and 17,594 binding samples before exclusions.Source: §3 · p. 5
- The five scPDB folds were split by UniProt ID so structures from the same protein would not occur in both training and testing within a split.The authors state this separation allows a more robust and fair assessment.Source: §3 · p. 5
- Testing used COACH420, HOLO4K, and the apo/holo subsets of CHEN, covering single-chain structures, larger multi-chain structures, and non-homologous proteins.COACH420 contains 420 structures; HOLO4K contains 4,009 structures.Source: §3 · pp. 5–6
4 Implementation
DeepSurf’s implementation preprocesses protein structures and surface points, balances binding and non-binding training samples, and uses a CUDA-implemented custom layer. Training uses regularization, batch normalization, and Adam optimization.
4 Implementation
- Surface-point simplification increased the average minimum neighbor distance from 0.7 Å to 2.3 Å using f = 10.DMS generated solvent-accessible surface points and normal vectors; simplification was chosen to balance retained surface information against computational cost.Source: §4 · p. 6
- Proteins were stripped of water, ions, and ligands, protonated when needed for feature computation, then stripped of hydrogens before binding-site extraction.Source: §4 · p. 6
- Non-binding samples were randomly undersampled per protein to match the binding-sample count, producing a 50/50 class balance.Binding points were surface points within 4 Å of any ligand atom; the remaining points supplied non-binding samples.Source: §4 · p. 6
- The LDS custom layer calculates a transition matrix and performs patch-level multiplication; it was implemented in CUDA to parallelize this computationally heavy operation.The implementation was integrated into TensorFlow.Source: §4 · p. 7
- All models trained for 20 epochs with batches of 64, Adam at a 10^-3 learning rate, L2 regularization of 10^-4, and default batch normalization.Source: §4 · p. 7
5 Results and Discussion
Across cross-validation and test-set comparisons, DeepSurf performs strongly against deep-learning competitors and competitively against traditional methods. Results assess pocket localization and, in some comparisons, pocket shape and the number of extracted pockets.
5.1 Cross-validation on scPDB
- 62.1% average cross-validation success with DeepSite’s architecture, versus its derived estimate of about 50%; ResNet-18 reaches 68.1% top-n.Removing surface-grid alignment reduced both top-n and top-(n+2) accuracy by 1.3%.Source: Table 1 · p. 8, §5.1 · p. 8
- 1.7 million parameters and an 11.3% accuracy drop for bottleneck ResNet; LDS raises accuracy by 1.5% and beats parameter-matched bottleneck ResNet-34 by 0.7% top-n and 0.9% top-(n+2).Bottleneck LDS-ResNet-18 has more than 10 times fewer parameters than ResNet-18 while achieving similar or better cross-validation performance.Source: §5.1 · p. 8
5.2 Comparison to DL-based methods
- 4% top-n and 3% top-(n+2) in COACH420 versus Kalasanty, and 4.5% top-n and 2.5% top-(n+2) in HOLO4K versus DeepSite; DeepSurf beats all DL competitors overall.The comparisons use DCA with a 4 Å cutoff.Source: Table 2 · p. 9, §5.2 · p. 9
- Among correctly located binding pockets (DCA < 4 Å), DeepSurf achieves higher OVR than Kalasanty in both COACH420 and HOLO4K, especially COACH420; Kalasanty returns no pocket for 475 of 4009 HOLO4K proteins.DeepSurf and Kalasanty tend to return fewer pockets than DeepSite and Jiang’s method in both datasets. In COACH420, their counts are similar to the average 1.2 true pockets; in HOLO4K, they return fewer than the actual average of 2.8.Source: §5.2 · p. 10
5.3 Comparison between apo and holo structures
- On apo structures, DeepSurf’s accuracy drops only 1–1.5%, the smallest decline; it also leads competitors on the most dissimilar proteins, with accuracies between 35% and 38%.The apo and holo evaluation uses 104 matched protein pairs; the most similar bin contains only four proteins, limiting statistical conclusions there.Source: §5.3 · pp. 10–11, Table 5 · p. 11
5.4 Sensitivity on ligandability threshold
- Lower ligandability thresholds generally reduce performance, especially on HOLO4K, while increasing pocket counts; for ResNet-18, T can be lowered to 0.6 when larger sites are needed.For LDS-ResNet, T = 0.9 significantly outperforms lower values; for ResNet-18, values above 0.6 are statistically equivalent.Source: §5.4 · pp. 11–12
5.5 Comparison to non-data-driven methods
- AutoSite achieves the highest accuracies on HOLO4K, while COFACTOR is slightly worse than DeepSurf top-n but clearly better top-(n+2), at hours rather than seconds of execution time.Fpocket is inferior to the other methods on both datasets; COFACTOR was evaluated only on COACH420 because of its execution time.Source: §5.5 · p. 13, Table 7 · p. 14
- Large candidate lists can rescue top-(n+2) misses, yet distance-based success can count a wide, physically meaningless pocket; the authors identify incomplete training data as a possible performance constraint.They note that sites without known ligands may be labeled negative and suggest that adding more diverse pocket samples could improve performance.Source: §5.5 · pp. 13–14
6 Conclusion
DeepSurf combines protein-surface grids with a 3D convolutional network to predict candidate binding sites, outperforming competing deep-learning methods across diverse datasets while remaining competitive with non-data-driven methods.
6 Conclusion
- DeepSurf predicts potential druggable sites by processing surface-placed 3D voxel grids through a 3D convolutional network and clustering the resulting ligandability scores.The method computes chemical features for the grids before network processing.Source: §6 · p. 15
- Across three diverse datasets of holo and apo structures, DeepSurf outperformed competing deep-learning methods in localization and overlapping accuracies and achieved competitive results against non-data-driven methods.Overlapping accuracy remained small, leaving extraction of ideally shaped binding sites an open issue.Source: §6 · p. 15
- DeepSurf reaches maximum performance by retaining only the network's most reliably assigned points, according to the ligandability-threshold sensitivity analysis.Source: §6 · p. 16
Funding
The work was supported by the ATXN1-MED15 PPI project, funded by the GSRT - Hellenic Foundation for Research and Innovation.
Funding
- The ATXN1-MED15 PPI project funded the work through the GSRT - Hellenic Foundation for Research and Innovation.Source: Funding · p. 16