Engineering Dehalogenase Enzymes Using Variational Autoencoder-Generated Latent Spaces and Microfluidics.

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Enzymes are vital for sustainable industrial processes, but their optimization is challenging due to complex residue interactions.
• Current computational methods for enzyme design often rely on evolutionary data, facing difficulties in analyzing residue variability and dependencies.

Purpose of the Study:

• To develop a novel machine learning method to overcome limitations in computational enzyme design.
• To engineer improved haloalkane dehalogenases with enhanced stability and activity.

Main Methods:

• A machine learning approach combining variational autoencoders (VAEs) with an evolutionary sampling strategy was developed.
• The method was applied to design novel sequences for haloalkane dehalogenases.
• Three cycles of design-build-test were performed, incorporating experimental validation using a microfluidic device (MicroPEX).

Main Results:

• The VAE-based method successfully improved enzyme solubility from 11% to 75%.
• Twenty multiple-point variants were generated, with nine showing significant improvements despite low sequence similarity (67%) to the template.
• Variants exhibited increased melting temperatures (up to 9 °C) and enhanced activity (3.5-fold increase in the best variant).

Conclusions:

• The developed machine learning method effectively addresses challenges in computational enzyme design.
• The engineered haloalkane dehalogenases demonstrate superior stability and activity, suitable for industrial applications.
• The generated dataset provides valuable data for validating new protein design strategies.