Protein sequence design on given backbones with deep learning.

Area of Science:

• Computational biology
• Protein engineering
• Bioinformatics

Background:

• Deep learning models protein sequence design by sampling amino acid distributions based on backbone structure.
• Physically foldable protein sequences require proper consideration of inter-residue couplings, addressed explicitly in iterative or autoregressive methods.
• Non-autoregressive models offer computational efficiency but require experimental validation.

Purpose of the Study:

• To evaluate the current state of computational metrics used for protein sequence design.
• To highlight the limitations of existing metrics and the need for experimental validation.
• To encourage the use of wet experiments for validating deep learning-based protein design methods.

Main Methods:

• Review of deep learning approaches for protein sequence design, including iterative, autoregressive, and non-autoregressive models.
• Analysis of current evaluation metrics such as native sequence recovery rate and native sequence perplexity.
• Discussion of complementary metrics like sequence-structure compatibility derived from energy calculations or structure prediction.

Main Results:

• Existing computational metrics for protein design have limitations that may not accurately predict real-world performance.
• Non-autoregressive models, while efficient, have not yet been sufficiently validated by experimental testing.
• Native sequence recovery and perplexity are common but potentially insufficient metrics for assessing design success.

Conclusions:

• Computational metrics for protein design require further development and validation.
• Experimental validation through wet experiments is crucial for assessing the true performance of protein design methods.
• A combination of computational and experimental approaches is necessary for robust protein sequence design.