A Generative Neuro-Symbolic AI for Protein Sequence Design

Area of Science:

• Computational biology
• Artificial intelligence
• Protein engineering

Background:

• Deep learning excels at protein sequence generation for target backbones.
• Current auto-regressive methods struggle with long-range dependencies crucial for protein function.
• Limitations in current inverse folding tools hinder the design of complex proteins.

Purpose of the Study:

• Introduce EffieDes, a neuro-symbolic AI framework for advanced protein design.
• Overcome limitations of auto-regressive sampling in generating functional protein sequences.
• Enable precise dissection of protein fitness landscapes for optimized performance.

Main Methods:

• Developed EffieDes, integrating deep learning with automated reasoning.
• Used deep learning to encode protein backbone fitness into a probabilistic graphical model (Potts model).
• Employed an automated reasoning prover to explore fitness landscapes and identify optimal sequences.

Main Results:

• Designed orthogonal protein pairs with selective self-assembly capabilities.
• Created a de novo nanobody with high affinity for a SARS-CoV-2 variant.
• Demonstrated EffieDes's ability to satisfy complex design constraints and optimize fitness.

Conclusions:

• EffieDes offers a robust architecture for dissecting learned protein fitness landscapes.
• The neuro-symbolic approach enables precise protein design for sophisticated functional objectives.
• This framework opens new avenues for engineering proteins with highly optimized performance.