Improved drug-target interaction prediction with intermolecular graph transformer.

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Drug-target interaction prediction is crucial for virtual screening and drug discovery.
• Current deep learning models often overlook critical intermolecular information, limiting prediction accuracy.
• Molecular docking and existing deep learning methods have limitations in capturing complex interactions.

Purpose of the Study:

• To introduce a novel approach, the Intermolecular Graph Transformer (IGT), for accurate drug-target interaction prediction.
• To address the limitations of existing models in capturing topological and spatial intermolecular information.
• To improve the performance and generalization ability of drug-target binding prediction models.

Main Methods:

• Developed a three-way Transformer-based architecture incorporating a dedicated attention mechanism.
• Utilized graph representations to model intermolecular information effectively.
• Evaluated IGT against state-of-the-art methods on binding activity and binding pose prediction tasks.

Main Results:

• IGT achieved superior performance, outperforming the second-best approach by 9.1% in binding activity and 20.5% in binding pose prediction.
• Demonstrated enhanced generalization ability on unseen receptor proteins compared to state-of-the-art methods.
• Successfully identified 83.1% of experimentally validated active drugs against SARS-CoV-2 with near-native binding poses.

Conclusions:

• The Intermolecular Graph Transformer (IGT) represents a significant advancement in drug-target interaction prediction.
• IGT's ability to model intermolecular information leads to improved accuracy and generalization.
• IGT shows practical utility in drug screening, particularly for identifying potential therapeutics like those for SARS-CoV-2.