Predicting locations of cryptic pockets from single protein structures using the PocketMiner graph neural network

Area of Science:

• Computational biology
• Drug discovery
• Structural bioinformatics

Background:

• Cryptic pockets are transient protein binding sites crucial for drug discovery.
• Identifying these pockets is currently slow and labor-intensive.
• Targeting cryptic pockets can unlock proteins previously considered undruggable.

Purpose of the Study:

• To develop a rapid and accurate method for predicting cryptic pocket formation.
• To accelerate the identification of druggable pockets for drug discovery.
• To expand the range of targetable proteins in the human proteome.

Main Methods:

• Developed PocketMiner, a graph neural network model.
• Trained PocketMiner on a dataset of 39 experimentally confirmed cryptic pockets.
• Applied PocketMiner to predict cryptic pocket formation in molecular dynamics simulations.
• Validated PocketMiner's performance using ROC-AUC metric.

Main Results:

• PocketMiner accurately identifies cryptic pockets with a ROC-AUC of 0.87.
• The method is over 1,000-fold faster than existing techniques.
• Application across the human proteome suggests over half of proteins may contain cryptic pockets.
• Predicted pockets were confirmed to open during simulations.

Conclusions:

• PocketMiner significantly accelerates the discovery of cryptic pockets.
• This AI-driven approach vastly expands the pool of druggable proteins.
• The findings have major implications for future drug development strategies.