Translating chemical reactions into enzyme sequences enables formate dehydrogenase engineering

Enzyme engineering faces challenges such as high costs and low success rates. However, generative artificial intelligence, like machine translation, can expand the sequence space for enzyme engineering. The study introduces REXzyme, a model that translates between chemical reactions and amino acid sequences. In silico evaluations assess REXzyme's designs, revealing strengths and limitations of its generalization capacity.

The model is further steered to generate enhanced formate dehydrogenases for the reverse natural reaction of reducing carbon dioxide. All 20 generated enzymes were expressed and purified as soluble proteins, demonstrating CO2-reduction activity. The most active design reached 2.5-fold the specific activity of the reference enzyme.

Importantly, all substitutions in the designed variants occurred on distal positions outside the catalytic site. The findings suggest that reaction-to-sequence translation is a promising approach for efficiently engineering enhanced biocatalysts.