Source-linked AI summary
Improving detection of protein-ligand binding sites with 3D segmentation
Marta M. Stepniewska-Dziubinska, Piotr Zielenkiewicz, Pawel Siedlecki
TL;DR
Finding protein binding pockets remains challenging, with existing methods reported to detect only 30%-40% of pockets. This work frames pocket detection as 3D segmentation and presents Kalasanty, which correctly located 44.6% of test pockets versus DeepSite’s 23.8%.
Problem
Existing methods detect only 30%-40% of binding pockets, and accurately locating sites remains a challenging task relevant to drug design.
Source: §1 · p. 1Method
Kalasanty adapts a 3D fully convolutional U-Net to segment protein grids into probability maps of binding-pocket locations.
Source: §1 · p. 2, §2.1 · p. 3Results
44.6% of test pockets were correctly located by Kalasanty, compared with 23.8% for DeepSite.
Source: §5 · p. 10Takeaways & Limitations
Kalasanty was stable, with cross-validation results comparable to results on new data not used for training or validation.
Source: §5 · p. 10Because sc-PDB contains only deep cavities, the model cannot detect binding sites on flat surfaces; closely located sites may also merge during post-processing.
Source: §5 · p. 10
Abstract
from arXiv · showhide
In recent years machine learning (ML) took bio- and cheminformatics fields by storm, providing new solutions for a vast repertoire of problems related to protein sequence, structure, and interactions analysis. ML techniques, deep neural networks especially, were proven more effective than classical models for tasks like predicting binding affinity for molecular complex. In this work we investigated the earlier stage of drug discovery process - finding druggable pockets on protein surface, that can be later used to design active molecules. For this purpose we developed a 3D fully convolutional neural network capable of binding site segmentation. Our solution has high prediction accuracy and provides intuitive representations of the results, which makes it easy to incorporate into drug discovery projects. The model's source code, together with scripts for most common use-cases is freely available at http://gitlab.com/cheminfIBB/kalasanty
1 Introduction
Finding protein binding pockets is an important and challenging part of rational drug design. The paper introduces Kalasanty, a deep-learning approach that treats pocket detection as 3D image segmentation.
1 Introduction
- Existing methods detect only 30%-40% of pockets, while DeepSite places about 50% within 4 Å despite outperforming Fpocket and Concavity.The authors describe that localization rate as insufficient.Source: §1 · pp. 1–2
- Traditional pocket-detection methods use geometry, chemical-probe binding energy, sequence conservation, or combinations.Some tools identify candidate pockets before scoring them; Fpocket is given as an example.Source: §1 · p. 1
- Pocket detection has relatively available protein-structure inputs but harder-to-acquire site labels, while deep learning can learn features from unprocessed data with less manual engineering.The introduction describes supervised learning as relating data to desired labels and deep learning as using internal layers to extract features.Source: §1 · p. 2
- The paper recasts pocket finding as semantic image segmentation, which assigns scores representing the probability that each image pixel belongs to the target object.The proposed task applies this segmentation framing to binding pockets.Source: §1 · p. 2
- Kalasanty adapts U-Net to protein structures, converting a 3D protein grid into pocket-probability densities and exportable pocket representations.Predictions can be saved as .cmap or .cube files, and pocket-forming protein parts can be saved as .mol2 files.Source: §1 · p. 2
2 Methods
Kalasanty frames pocket detection as 3D segmentation, using a U-Net-style network to map protein grids to pocket probabilities. The methods describe its training and data safeguards, then define how predictions are evaluated.
2.1 Our approach
- Kalasanty treats the protein and pocket as same-sized 3D grids, predicting a pocket probability at each input grid point.The network adapts U-Net with 3D layers and skip connections that pass encoder features to corresponding decoder blocks for localization.Source: §2.1 · pp. 3–4
- Training uses random translations and rotations, L2 weight regularization, and a negative continuous Dice objective optimized with Adam.The authors report 1.5·10^6 optimization steps, batches of 10, and a learning rate of 10^-6.Source: §2.1 · p. 4
2.2 Data
- To limit data leakage, structures from the benchmark were removed from sc-PDB, and validation folds were grouped by UniProt ID.After additional structure-loading exclusions, 15,860 structures were used for training; the final model was tested on 149 structures containing 269 binding sites.Source: §2.2 · pp. 4–5
- The authors caution that their results cannot be directly compared with Chen et al. because they used only a subset of the original dataset.DeepSite’s test set also included 12 proteins used in its training, which the authors note could make its evaluation slightly optimistic.Source: §2.2 · p. 5
2.3 Results evaluation
- Evaluation uses DCC for pocket-location success and DVO for overlap of predicted and actual pocket shapes.Pockets with DCC below 4˚A count as correctly located; Kalasanty probabilities are thresholded at 0.5, and the top n predicted pockets per structure are matched to real pockets.Source: §2.3 · p. 6
3 Results
Across validation folds, Kalasanty’s results were stable, and it showed no systematic performance difference for the most frequent protein types. On the independent test set, it located more pockets within 4˚A than DeepSite, although performance was lower than in cross-validation.
Cross-validation
- Validation results were stable across folds, with F1 0.69, recall 0.66, and precision 0.73; the model returned no pockets for 4.7% of structures.Figure 2 shows success rates across DCC thresholds and DVO distributions for correctly located pockets.Source: §3 · p. 6, Figure 2 · p. 7
- The most frequent protein types showed no systematic performance difference from the average, which the authors interpret as evidence of generalization to unseen structures.The dataset was unbalanced, and performance did differ between some protein groups.Source: §3 · p. 7
Test-set evaluation
- 44.6% of test binding sites were correctly located by Kalasanty, versus 23.8% for DeepSite, using the DCC < 4˚A criterion.Kalasanty’s test F1 was 0.45 (precision 0.64, recall 0.35); DeepSite’s was 0.26 (precision 0.36, recall 0.20).Source: §3 · p. 7
- On the test set, Kalasanty was slightly less successful than in cross-validation, and 5% of its correctly located pockets had DVO below 0.25.Only 3 test structures (1.3%) had no detected pockets.Source: §3 · p. 7
4 Discussion
Kalasanty detected nearly twice as many test-set pockets as DeepSite, and more often placed predictions closer to the binding-site centers. Differences in predicted pocket size and protein-specific failures qualify that advantage; observed protein-property trends were not shown to be causal.
4 Discussion
- 84.6% of pockets detected by at least one model had lower DCC with Kalasanty, which correctly located 44.6% versus DeepSite’s 23.8%.The comparison covers the 136 pockets detected by at least one model.Source: §4 · p. 7, Figure 4 · p. 9
- DeepSite’s predicted pockets averaged twice the size of Kalasanty’s, helping explain why its correctly located pockets often had inaccurate shapes.The authors suggest that DeepSite’s more voluminous predictions account for the difference in performance.Source: §4 · pp. 7–8, Figure 5 · p. 9
- DeepSite correctly located all holo and apo pockets for L-histidinol dehydrogenase and 7,8-dihydroneopterin aldolase, while Kalasanty missed them.The two proteins were highlighted as cases where the models differed.Source: §4 · p. 8, Figure 4 · p. 9
- The test set was too small for statistical analysis of whether the two proteins’ failures reflected a broader trend.The authors examined protein properties using cross-validation results instead.Source: §4 · p. 8
- Cross-validation showed higher DCC values for both proteins’ superfamilies and significant differences for both source organisms, but these relationships were not established as causal.The authors noted that another factor correlated with these properties could underlie the differences.Source: §4 · pp. 9–10
5 Conclusions
Kalasanty adapts U-Net for 3D protein binding-cavity detection and locates test-set pockets more precisely than DeepSite. Its scope is limited by training data containing only deep cavities, and closely spaced pockets may be merged during post-processing.
5 Conclusions
- 44.6% of test-set pockets were correctly located by Kalasanty, versus 23.8% for DeepSite.Cross-validation results were comparable to those on new data.Source: §5 · p. 10
- Because sc-PDB contains only deep cavities, the model cannot detect binding sites on flat surfaces; obtaining a model for those sites requires other datasets.The authors note that such datasets were outside this study’s scope.Source: §5 · p. 10
- Post-processing may merge closely located binding sites into one pocket; local probability maxima are proposed as a possible extension.Source: §5 · p. 10
- The source code, trained model, and helper scripts are freely available, and CPU inference takes one second on an Intel Core i7.Loading the model takes five seconds; GPU is needed for training, not inference.Source: §5 · p. 10
- Kalasanty adapts U-Net to 3D protein structures and binding-cavity detection, extending deep-learning methods developed in other fields to structural bioinformatics.Source: §5 · p. 10
Supplementary Figures
Figure S1 relates prediction accuracy to protein superfamily, fold, and source organism.
Supplementary Figures
- Panels A–D group prediction accuracy by superfamily, fold, and source organism; compare groups, with n indicating binding-site counts.P-values were calculated using the Mann–Whitney U-test.Source: Figure S1 · p. 13