Pocket2Drug: An Encoder-Decoder Deep Neural Network for the Target-Based Drug Design.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning in drug discovery

Background:

• Modern drug discovery relies heavily on computational modeling to identify potential drug candidates for target proteins.
• Advances in structural biology reveal numerous disease-related protein binding sites, creating opportunities for computational prediction models.

Purpose of the Study:

• To introduce Pocket2Drug, a deep graph neural network model designed to predict small molecules that bind to specific ligand binding sites.
• To leverage data mining and machine learning for efficient computational models in drug candidate selection.

Main Methods:

• Pocket2Drug utilizes a deep graph neural network architecture.
• The model is trained on a large dataset of protein pocket structures to learn the probability distribution of small molecules.
• Drug candidates are sampled from the trained model for prediction.

Main Results:

• Pocket2Drug significantly enhances the success rate of identifying molecules that bind to target pockets compared to traditional methods.
• The model successfully generated known binders for 80.5% of targets in a diverse, independent testing set.
• Performance was validated against a testing set distinct from the training data.

Conclusions:

• Pocket2Drug represents a promising computational approach for accelerating the discovery of novel biopharmaceuticals.
• The model's ability to predict binding molecules offers a valuable tool for informing drug discovery pipelines.