Integrating protein language models and automatic biofoundry for enhanced protein evolution

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Traditional protein engineering methods like directed evolution are slow and labor-intensive.
• Machine learning and automated biofoundries offer new avenues for optimizing protein engineering.
• Automated protein engineering platforms can accelerate the discovery of novel protein functions.

Purpose of the Study:

• To develop a closed-loop, automated protein engineering platform.
• To integrate a protein language model (ESM-2) with an automated biofoundry for rapid protein evolution.
• To enhance the speed and accuracy of protein engineering for industrial applications.

Main Methods:

• A closed-loop Design-Build-Test-Learn system was implemented.
• The protein language model ESM-2 was used for zero-shot prediction of protein variants.
• An automated biofoundry constructed and evaluated variants.
• A multi-layer perceptron trained a fitness predictor for subsequent variant selection.

Main Results:

• Four rounds of automated evolution were completed in 10 days using tRNA synthetase as a model enzyme.
• Enzyme activity was improved by up to 2.4-fold.
• The platform demonstrated significantly enhanced speed and accuracy in protein evolution.

Conclusions:

• The developed platform streamlines protein engineering through automation and machine learning.
• This approach accelerates the discovery of improved enzymes for various applications.
• Automated protein evolution holds significant promise for industrial biotechnology.