ML framework BioStrucTag unlocks full stereoisomer space via ADH engineering

Excited to share that our research on machine learning-guided enzyme engineering of alcohol dehydrogenase has been published in Nature Synthesis!

In this work, we developed BioStrucTag, a machine learning framework that combines structural and sequence information to guide efficient enzyme engineering. By integrating 3D voxel-grid representations of ligand-enzyme complexes, which capture the structural features of pre-organized catalytic pockets associated with specific chiral products, with enzyme sequence feature extracted using a transformer-based protein language model, BioStrucTag enables systematic exploration of enzyme sequence with minimal screening.

Through just 2-3 rounds of iterative engineering, we successfully accessed all 12 alcohol stereoisomeric alcohol products containing three stereocenters through the asymmetric reductions of 2,2-disubstituted cyclodiketones. 

The work demonstrates the potential of combining structural representations and protein language models to accelerate the discovery and engineering of highly selective biocatalysts.

Read the paper: https://rdcu.be/fBfCJ

BioStrucTag is openly available on GitHub: https://github.com/zhoujiahui01/BioStrucTag/