AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery

Abstract: 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.

• The paper introduces AtomNet, a deep convolutional neural network that revolutionizes structure-based bioactivity prediction by applying 3D convolutional filters to molecular data.
• The methodology employs 3D grid representations of target-ligand complexes and achieves an AUC greater than 0.9 on 57.8% of DUDE benchmark targets.
• The findings demonstrate that incorporating spatial chemical features in DCNNs enhances drug discovery processes by reducing dependence on extensive target data.

AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery

The paper "AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery" introduces an innovative approach to molecular binding affinity prediction using deep convolutional neural networks (DCNNs). This research marks a significant advancement in leveraging DCNNs for structure-based drug discovery beyond their traditional applications in fields such as image and speech recognition.

Significance of AtomNet

AtomNet is noteworthy for several reasons. Prior models in drug discovery predominantly utilized DNN architectures for QSAR and ligand-based predictions, lacking the spatial and structural information crucial for accurately modeling bioactivity and chemical interactions. In contrast, AtomNet innovatively applies local convolutional filters to structural data, effectively predicting active molecules for targets without known modulators—a domain previously challenging for computational models.

Methodology and Approach

AtomNet integrates DCNNs with a structure-based modeling approach, capitalizing on the local interaction of biochemical features. The convolutional filters are adept at discerning intricate chemical interactions such as hydrogen bonding and π-stacking through localized feature recognition. This ability parallels the recognition of simple patterns in image processing but applied within the chemical space.

The architecture comprises input representation through 3D grids over target-ligand complexes, followed by convolutions and fully connected layers. The model capitalizes on features detected through this grid, optimizing predictions of molecular interactions.

Experimental Evaluation

AtomNet is extensively validated on three challenging benchmarks: the DUDE benchmark, a DUDE-like internal dataset, and a dataset with experimentally verified inactives. Notably, AtomNet achieved an AUC greater than 0.9 on 57.8% of the DUDE targets, a substantial improvement over traditional docking methods, which managed similar performance for only 1% of the targets.

Implications and Future Directions

The implementation of AtomNet demonstrates a practical enhancement in structure-based drug discovery workflows, offering a predictive tool that reduces reliance on extensive prior data about molecular targets. Practically, this can expedite drug discovery cycles, identify promising leads earlier in the development process, and provide insights into molecular interaction networks that were previously inaccessible.

From a theoretical perspective, AtomNet exemplifies how DCNNs can be tailored beyond conventional domains to accommodate the complexities of chemical data, setting the stage for more nuanced applications in computational chemistry and pharmacology.

Future developments might focus on enhancing the interpretability of detected features and expanding this approach to additional types of molecular interactions. Continued improvement in data representation techniques and model robustness will further elevate the potential of DCNNs in structural biology and drug design.

Conclusion

AtomNet represents a significant progression in the application of machine learning to biochemical and pharmacological research, reflecting the growing intersection of computational methodologies and biological sciences. The model's efficacy and innovation suggest promising advancements in AI-driven drug discovery, potentially transforming how novel drugs are identified and developed.