XENet: Using a new graph convolution to accelerate the timeline for protein design on quantum computers

Area of Science:

• Computational biology
• Machine learning
• Protein engineering

Background:

• Graph representations are standard for protein structures in sequence design when backbone conformation is known.
• Existing graph convolution algorithms struggle with protein environments due to limited focus on edge attributes and shallow architectures.

Purpose of the Study:

• To introduce an improved message-passing operation, XENet, for enhanced protein design.
• To address limitations in current graph convolution methods for representing protein environments.

Main Methods:

• Developed XENet, a novel message-passing operation emphasizing both incoming and outgoing edge attributes.
• Compared XENet against existing graph convolutions for protein side-chain optimization and sequence design within Rosetta's protocol.
• Evaluated XENet's performance in decreasing rotamer sample counts and its tolerance for deeper neural network architectures.

Main Results:

• XENet decreased rotamer counts by 40% without compromising quality.
• Reduced classical pre-computation memory consumption by over 3x.
• Decreased quantum algorithm qubit consumption by 40% and solution space size by 165x.
• Demonstrated superior performance and greater tolerance for deeper architectures compared to existing models.

Conclusions:

• XENet offers a significant advancement in modeling local kinematics for protein design.
• The method enhances computational efficiency for classical and quantum approaches to protein sequence design.
• XENet's ability to handle deeper architectures opens possibilities for more complex protein design challenges.