AlphaFold distillation for inverse protein design.

Area of Science:

• Computational Biology
• Protein Engineering
• Bioinformatics

Background:

• Inverse protein folding designs sequences for specific 3D structures, vital for bio-engineering and drug discovery.
• Existing methods often depend on limited experimental structures.
• Accurate protein structure prediction models (e.g., AlphaFold) are too slow for inverse folding training.

Purpose of the Study:

• To develop a computationally efficient method for inverse protein folding.
• To integrate fast structure prediction into inverse folding model training.
• To enhance sequence recovery and diversity in protein design.

Main Methods:

• Knowledge distillation applied to protein folding model confidence metrics (pTM, pLDDT).
• Creation of a fast, end-to-end differentiable distilled model.
• Utilizing the distilled model as a structure consistency regularizer during inverse folding training.

Main Results:

• The distilled model significantly speeds up structure consistency checks.
• Achieved up to 3% improvement in sequence recovery compared to baselines.
• Demonstrated up to 45% increase in protein diversity while maintaining structural integrity.

Conclusions:

• Knowledge distillation offers an efficient way to regularize inverse folding models.
• The proposed method enhances protein design capabilities for bio-engineering applications.
• This technique is adaptable for other protein design tasks like sequence-based infilling.