Rotational Variance-Based Data Augmentation in 3D Graph Convolutional Network.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Protein-ligand binding is crucial for drug action but is sensitive to molecular orientation.
• Current computational models often struggle with the rotational variance inherent in binding events.
• Accurate prediction of binding requires robust handling of 3D molecular structures and their orientations.

Purpose of the Study:

• To develop a data augmentation strategy using molecular rotation to improve the orientation recognition of computational models.
• To evaluate the impact of this augmentation on the prediction accuracy of a 3D graph convolutional network (3DGCN) for protein-ligand binding.
• To demonstrate the utility of the augmented model in identifying active ligands for drug screening.

Main Methods:

• Proposed data augmentation by rotating known active ligands (1-labeled) to generate inactive (0-labeled) data.
• Utilized a 3D graph convolutional network (3DGCN) to analyze protein-ligand binding events.
• Investigated the rotation-dependent prediction accuracy of the 3DGCN before and after data augmentation.

Main Results:

• Data augmentation by molecular rotation significantly improved the 3DGCN's ability to recognize ligand orientation.
• The enhanced orientation recognition led to a significant improvement in the classification accuracy for human beta-secretase 1 (BACE-1)/ligand binding.
• The data-augmented 3DGCN demonstrated improved performance in predicting active ligands from a candidate dataset.

Conclusions:

• Molecular rotation as a data augmentation technique effectively enhances the orientation-recognizing capabilities of 3DGCNs for protein-ligand binding prediction.
• This approach offers a promising method to improve the accuracy of computational models used in virtual drug screening and discovery.
• The improved orientation recognition is key to advancing the identification of novel therapeutic agents.