Mind the gap: An embedding guide to safely travel in sequence space

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Generative algorithms are increasingly vital for designing novel enzymes.
• Current deep learning methods for enzyme design can be difficult to interpret and control.
• Ensuring structural integrity and functional properties in enzyme mutants is crucial for experimental success.

Purpose of the Study:

• To develop a hybrid approach combining protein language models (pLMs) and Monte Carlo (MC) sampling for enzyme mutant generation.
• To create enzyme mutants that preserve the geometry of the catalytic site while allowing significant sequence variation.
• To provide a controllable and interpretable generative process for enzyme design.

Main Methods:

• Utilized a protein language model (pLM) to define an energy landscape.
• Employed Monte Carlo (MC) sampling to explore this landscape and generate enzyme mutants.
• Integrated statistical mechanics concepts, such as temperature, to steer the generative process and control mutant properties.
• Validated the approach by comparing generated mutants to experimentally characterized chorismate mutase variants.

Main Results:

• Successfully generated enzyme mutants with conserved catalytic site geometry despite significant sequence divergence.
• Demonstrated that low embedding energy, as predicted by the pLM, is a necessary condition for catalytic function.
• Provided experimental validation for the energy function underpinning the generative approach.
• Generated over 12,500 sequences for 13 different enzymes involved in key catalytic processes.

Conclusions:

• The hybrid pLM-MC approach offers a controllable and interpretable method for designing enzyme mutants with desired properties.
• The findings provide experimental grounding for using energy landscapes in generative enzyme design.
• This work represents a significant advancement for generative algorithms in enzyme engineering and discovery.
• The released sequences facilitate experimental verification and further research in enzyme optimization.