Efficient Generation of Protein Pockets with PocketGen

Area of Science:

• Computational biology
• Protein engineering
• Drug discovery

Background:

• Designing novel protein-binding proteins is crucial for advancing drug discovery.
• Current AI-based protein design faces challenges due to complex interactions, molecular flexibility, and sequence-structure dependencies.

Purpose of the Study:

• To introduce PocketGen, a deep generative model for simultaneous residue sequence and atomic structure generation of protein binding sites.
• To ensure sequence-structure consistency in protein pocket design using advanced AI techniques.

Main Methods:

• PocketGen utilizes a bilevel graph transformer for multi-scale structural encoding (atom, residue, ligand levels).
• It incorporates a sequence refinement module with a structural adapter integrated into a protein language model for enhanced prediction alignment.

Main Results:

• PocketGen generates high-fidelity protein pockets with improved binding affinity and structural validity.
• The model demonstrates a 95% success rate in generating pockets with superior binding affinity and an amino acid recovery rate over 64%.

Conclusions:

• PocketGen offers a significant advancement in AI-driven protein design for drug discovery.
• The model provides a faster and more effective approach compared to physics-based methods, achieving high accuracy and efficiency.