Emulating Docking Results Using a Deep Neural Network: A New Perspective for Virtual Screening.

Area of Science:

• Computational chemistry
• cheminformatics
• artificial intelligence in drug discovery

Background:

• Molecular docking is crucial for virtual screening but computationally intensive.
• It's often a bottleneck in evaluating large compound libraries.
• Predicting interactions directly from 2D structures could accelerate this process.

Purpose of the Study:

• To investigate the feasibility of using deep neural networks to predict docking outputs.
• To develop a computationally efficient alternative to traditional docking methods.
• To enable rapid screening of vast compound libraries based on predicted interactions.

Main Methods:

• Developed a deep neural network protocol to predict docking results from 2D compound structures.
• Utilized interaction fingerprints to encode ligand-receptor complexes.
• Employed graph convolutional networks (GCNs) for predicting docking outcomes.
• Validated the approach on G protein-coupled receptor and CYP enzyme targets.

Main Results:

• The deep learning protocol is orders of magnitude faster than conventional docking software.
• The method accurately predicts ligand-receptor interactions and generates interaction fingerprints.
• A variant of graph convolutional network demonstrated high effectiveness in emulating docking results.
• Retrospective virtual screening experiments confirmed the method's utility.

Conclusions:

• Deep learning offers a computationally efficient and viable alternative for predicting molecular docking.
• The developed protocol accelerates virtual screening, enabling large-scale library evaluation.
• The method's speed and accuracy facilitate drug discovery pipelines.
• The readily available code promotes community adoption and further research.