From high-throughput evaluation to wet-lab studies: advancing mutation effect prediction with a retrieval-enhanced...

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Enzyme engineering is crucial for industrial and research applications, traditionally relying on directed evolution and rational design.
• Deep learning models offer advanced, cost-effective alternatives by capturing coevolutionary patterns in proteins.
• Understanding protein sequence, structure, and function relationships remains a key challenge in enzyme engineering.

Purpose of the Study:

• To introduce VenusREM, a retrieval-enhanced protein language model for enzyme engineering.
• To evaluate VenusREM's performance in predicting mutation effects on enzyme properties.
• To validate VenusREM's utility through computational and experimental assessments.

Main Methods:

• Developed VenusREM, a protein language model incorporating retrieval mechanisms.
• Assessed VenusREM on 217 ProteinGym benchmark assays.
• Conducted experimental validation using VHH antibody and DNA polymerase mutants.

Main Results:

• VenusREM achieved state-of-the-art performance on the ProteinGym benchmark.
• Computational and experimental analyses confirmed VenusREM's ability to improve VHH antibody stability and binding affinity.
• Designed DNA polymerase mutants with enhanced activity at elevated temperatures, validated experimentally.

Conclusions:

• VenusREM is a reliable computational tool for enzyme engineering.
• The study provides a comprehensive framework for evaluating mutation effect prediction models.
• Deep learning approaches significantly advance protein engineering capabilities.