DeepBindPoc: a deep learning method to rank ligand binding pockets using molecular vector representation.

Area of Science:

• Computational biology
• Structural bioinformatics
• Drug discovery

Background:

• Accurate identification of ligand-binding pockets is crucial for structure-based drug design.
• Existing deep learning models predict pockets but struggle to rank native ones effectively.
• There is a need for enhanced models trained on realistic data with ligand information.

Purpose of the Study:

• To propose DeepBindPoc, a novel deep learning method for identifying and ranking ligand-binding pockets in proteins.
• To leverage pocket and ligand information for improved pocket prediction accuracy.
• To provide a valuable tool for ranking near-native pockets for modeled proteins.

Main Methods:

• Developed DeepBindPoc, a deep learning model utilizing pocket and ligand data.
• Employed mol2vec to represent ligands and pockets as vectors for a fully connected neural network.
• Trained the model to automatically extract key features for pocket-ligand binding.

Main Results:

• DeepBindPoc effectively identifies and ranks ligand-binding pockets.
• The method shows complementary advantages when combined with traditional tools like fpocket and P2Rank.
• Validated on multiple datasets, including G-protein Coupled receptors, demonstrating robust performance.

Conclusions:

• DeepBindPoc is a valuable tool for ranking near-native ligand-binding pockets.
• The model is particularly useful for theoretically modeled proteins with known ligands but unknown active sites.
• The DeepBindPoc model and webserver are publicly available for research use.