Source-linked AI summary
AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery
Izhar Wallach, Michael Dzamba, Abraham Heifets
TL;DR
Molecular binding prediction remains inaccurate computationally despite its importance for drug discovery, while existing ligand-based deep models require prior target data and omit target structure. The paper introduces AtomNet, a structure-based deep convolutional network that combines ligand and target information to learn local chemical interactions. AtomNet outperforms previous docking methods and achieves an AUC greater than 0.9 on 57.8% of DUDE targets.
Problem
Existing computational tools remain too inaccurate for routine binding prediction, and ligand-based models depend on prior target data while remaining blind to target structure.
Method
AtomNet combines ligand information with target structure in a deep convolutional architecture using vectorized 1 Å 3D grids of sampled protein–ligand complexes.
Results
AtomNet outperforms previous docking methods across benchmarks, achieving an AUC greater than 0.9 on 57.8% of DUDE targets.
Takeaways & Limitations
AtomNet can predict new active molecules for targets with no previously known modulators while learning meaningful chemical features from structural inputs.
Takeaways & Limitations
Property-matched-decoy benchmarks exclude inactive molecules similar to actives, omitting challenging experimentally relevant cases.
Abstract
from arXiv · showhide
Deep convolutional neural networks comprise a subclass of deep neural networks (DNN) with a constrained architecture that leverages the spatial and temporal structure of the domain they model. Convolutional networks achieve the best predictive performance in areas such as speech and image recognition by hierarchically composing simple local features into complex models. Although DNNs have been used in drug discovery for QSAR and ligand-based bioactivity predictions, none of these models have benefited from this powerful convolutional architecture. This paper introduces AtomNet, the first structure-based, deep convolutional neural network designed to predict the bioactivity of small molecules for drug discovery applications. We demonstrate how to apply the convolutional concepts of feature locality and hierarchical composition to the modeling of bioactivity and chemical interactions. In further contrast to existing DNN techniques, we show that AtomNet's application of local convolutional filters to structural target information successfully predicts new active molecules for targets with no previously known modulators. Finally, we show that AtomNet outperforms previous docking approaches on a diverse set of benchmarks by a large margin, achieving an AUC greater than 0.9 on 57.8% of the targets in the DUDE benchmark.
1 Introduction
AtomNet addresses the need for more accurate molecular-binding prediction by applying deep convolutional architectures to structure-based bioactivity prediction. It combines ligand and target structural information to model local chemical interactions and extend prediction to targets without known modulators.
- Accurate molecular-binding prediction could accelerate treatment discovery, identify toxic molecules earlier, and guide medicinal chemistry.
- AtomNet is introduced as the first deep convolutional neural network for molecular binding affinity prediction and the first deep-learning system incorporating target structure.
- Deep convolutional networks reduce overfitting by sharing local detector weights and hierarchically composing simple features into more complex ones.
- AtomNet applies locality and hierarchical composition because biochemical interactions depend predominantly on nearby atoms, distances, angles, and chemical groups.
- Unlike ligand-based models, AtomNet combines ligand information with target structure, allowing arbitrary interaction features and predictions for targets without known binders.
2 Methods
The methods evaluate AtomNet across complementary structure-based benchmarks, including property-matched decoys and experimentally verified inactives. The model represents sampled protein–ligand complexes as 3D grids and processes them with convolutional and fully connected layers.
- Datasets: AtomNet is evaluated on DUDE, an internal DUDE-like benchmark, and a benchmark containing experimentally verified inactive molecules.
- Datasets: DUDE contains 102 targets, 22,886 actives, and 50 property-matched decoys per active, with 30 targets selected for testing.
- Datasets: The ChEMBL-20 benchmark filters activity, target confidence, binding-site annotation, structural resolution, PAINS, and promiscuity before pairing actives with ZINC decoys.
- Datasets: Property-matched-decoy benchmarks exclude similar inactive molecules, so experimentally verified inactives provide more challenging cases involving structurally similar molecules.
- Structure-based deep-convolutional neural network: The network uses four 3D-convolutional layers, two fully connected layers with 1024 hidden units each, and a logistic output layer for two activity classes.
- Structure-based deep-convolutional neural network: AtomNet converts sampled protein–ligand poses into 1 Å 3D grids within a 20 Å cube, then vectorizes the structural-feature grid.
3 Results
Across three benchmarks, AtomNet substantially outperforms Smina in overall ranking and early enrichment, including held-out targets and experimentally verified inactives.
- AUC measures overall ranked classification performance, whereas adjusted-logAUC emphasizes correct ranking near the top of the list.The adjusted-logAUC subtracts the random-curve area of 0.14462, so positive values indicate better-than-random performance.
- AtomNet’s held-out DUDE-30 mean and median AUC are 0.855 and 0.875, versus 0.7 and 0.694 for Smina, reducing available mean error by 51.6%.AtomNet exceeds 0.9 AUC for 14 held-out targets (46.7%), while Smina does so for one target (3.3%).
- AtomNet reaches AUC ≥0.9 for 10 ChEMBL-20-PMD held-out targets (20%), while Smina reaches it for zero targets.At AUC ≥0.8, AtomNet succeeds on 25 targets (50%), compared with one (2%) for Smina.
- On the experimentally verified inactives benchmark, AtomNet reaches AUC ≥0.9 for 10 targets (6.7%), compared with zero for Smina.At AUC ≥0.8, AtomNet succeeds for 45 targets (30.2%) and Smina for 4 (2.7%); both methods perform worse than on previous benchmarks.
- AtomNet’s mean logAUC on DUDE-30 is 0.321, compared with 0.153 for Smina, indicating stronger early enrichment.For target 1qzy, 35% of actives are concentrated at the top of the ranked list despite an AUC of 0.76, with logAUC 0.44.
4 Discussion
AtomNet is evaluated against Smina across multiple structure-based benchmarks, with indirect filter visualization used to inspect learned chemical features. The reported comparisons emphasize both overall ranking and early enrichment.
- Benchmark comparisons: AtomNet achieves an order-of-magnitude improvement over Smina across the four evaluation datasets at accuracy levels useful for drug discovery.The comparison uses target counts exceeding AUC and adjusted-logAUC thresholds.
- Filter visualization: AtomNet learns locally related chemical features, including a filter specializing as a sulfonyl/sulfonamide detector.The filter inferred a meaningful spatial arrangement of atom types without chemical prior knowledge.
- Benchmark comparisons: AtomNet outperforms prior docking methods on reported DUDE comparisons, including mean AUC of 0.895 versus 0.696 for DOCK3.7.The cited comparisons also report AtomNet median AUC of 0.93 versus 0.76 for Surflex-Dock and mean AUC of 0.852 versus 0.72 for Dock6.7.
5 Conclusion
The paper presents AtomNet as the first structure-based deep convolutional neural network for predicting small-molecule bioactivity. Its locally constrained architecture models molecular binding through hierarchical chemical features and achieves strong DUDE benchmark performance.
- 5 Conclusion: AtomNet is the first structure-based deep convolutional neural network designed to predict small-molecule bioactivity for drug discovery.The model uses local constraints to hierarchically compose proximate chemical features.
- 5 Conclusion: AtomNet can predict new active molecules for targets with no previously known modulators by incorporating structural target information.The conclusion attributes this capability to the model's use of target structure.
- 5 Conclusion: 57.8% of DUDE targets achieve AUC greater than 0.9 with AtomNet, far surpassing previous docking methods.This is reported as the principal benchmark result.