Enhancing Enzyme Activity With Mutation Combinations Guided by Few-Shot Learning and Causal Inference.

Area of Science:

• Biotechnology
• Enzyme Engineering
• Metabolic Engineering

Background:

• Enhancing enzyme product yield is crucial for metabolic engineering.
• Current methods often focus on thermostability, which may not directly correlate with activity.
• In vivo unit yield (yield/expression) offers a potential surrogate for optimizing enzyme activity.

Purpose of the Study:

• To develop and validate a workflow integrating computational predictions and experimental iteration for enzyme engineering.
• To establish in vivo unit yield as a viable metric for optimizing enzyme activity and product yield.
• To engineer high-yield enzyme variants for increased product formation.

Main Methods:

• Utilized causal inference and dataset analysis to validate in vivo unit yield as a surrogate for activity.
• Employed computational prediction of binding affinities for reactive intermediates to identify single mutants.
• Developed a few-shot learning model, Physics-Inspired Feature Selection of Protein Language Models (PIFS-PLM), for predicting mutation combinations.

Main Results:

• Achieved a 73-fold increase in bicyclogermacrene (BCG) yield and a 15% increase in BCG selectivity.
• Demonstrated the efficiency of optimizing for unit yield over thermostability.
• Provided crystallographic and biochemical evidence for the impact of specific mutations.

Conclusions:

• The developed workflow enables efficient engineering of high-yield enzyme variants.
• In vivo unit yield optimization is a powerful and efficient strategy for metabolic engineering.
• This approach significantly advances the design of enzymes for enhanced product formation.