DeepLigType: Predicting Ligand Types of Protein-Ligand Binding Sites Using a Deep Learning Model.

Area of Science:

• Computational biology
• Drug discovery
• Bioinformatics

Background:

• Protein-ligand binding site analysis is vital for early drug discovery.
• Predicting ligand types improves drug design decisions.
• Experimental methods are common for characterizing binding sites.

Purpose of the Study:

• To develop a computational method, DeepLigType, for predicting protein-ligand binding site types.
• To classify binding sites into five categories: antagonist, agonist, activator, inhibitor, and others.
• To leverage deep learning for enhanced prediction accuracy.

Main Methods:

• Utilized Fpocket to identify protein-ligand binding sites.
• Employed a Convolutional Block Attention Module (CBAM) with ResNet (CBAM-ResNet) deep learning model.
• Created a novel dataset, LigType5, from PDBbind and scPDB for training and testing.

Main Results:

• The CBAM-ResNet model achieved 74.30% accuracy in predicting ligand types.
• The model obtained an Area Under the Curve (AUC) of 0.83 on a novel test dataset.
• Successfully classified protein-ligand binding sites into five distinct functional categories.

Conclusions:

• DeepLigType offers a robust computational alternative to experimental methods for binding site analysis.
• The developed deep learning architecture demonstrates significant potential in predicting ligand types.
• Accurate ligand type prediction can accelerate and refine the drug discovery process.