REXzyme translates a chemical reaction into an enzyme sequence, and its formate dehydrogenase design C20 reaches 2.5 times wild-type activity
Synopsis
Enzyme engineering has relied on directed evolution, which screens thousands of variants even though fewer than 5% of single mutations are beneficial. REXzyme recasts enzyme design as translation: an encoder reads a reaction written in SMILES and a decoder generates an amino acid sequence token by token, after training on about 21 million reaction-enzyme pairs. Steered toward the reverse reaction of formate dehydrogenase, the reduction of CO2 to formate, all 20 designs generated after family-specific fine-tuning and reinforcement-learning alignment expressed as soluble proteins and showed activity, and design C20 reached a specific activity of 80.8 mU/mg, 2.5 times that of wild-type CtFDH.
Interpretation
Enzyme sequence generation can now be driven by a complete user-defined chemical reaction, with substrates and products written at the atomic level in SMILES rather than specified indirectly through sequence prompts or broad functional annotations. Protein language models built for protein generation have mostly used decoder-only architectures, which offer no direct mechanism for conditioning on a complete chemical reaction, while enzyme engineering needs the target transformation stated explicitly. The model was trained on a self-curated set of about 21 million non-redundant reaction-enzyme pairs drawn from Rhea, BKMS, UniProt and BRENDA, covering roughly 16,000 unique reactions and about 15 million unique enzyme sequences, with training and test sets split so their chemical reactions do not overlap.
The model translates a reaction into a sequence by autoregressive generation on an encoder-decoder architecture: the encoder reads the reaction and the decoder predicts each next amino acid from the tokens already generated. On the reaction side a custom ABPE tokenisation algorithm first treats atomic units such as Cl as indivisible and then merges frequently recurring SMILES fragments into one token, while enzymes are split character by character into one-letter amino acids. Training minimises the negative log-likelihood of the target enzyme sequence given its reaction, letting the model extract shared patterns across similar pairs, as when different nitrilases vary in substrate preference for nitrile hydrolysis but conserve the overall mechanism and chemical groups.
On formate dehydrogenase, the designed variants raise CO2-reduction specific activity to 2.5 times that of wild-type CtFDH, and every substitution sits outside the catalytic site. Most natural formate dehydrogenases show very low activity in the CO2-reduction direction, and earlier engineering of CtFDH focused on single or double substitutions at a small number of residues within or adjacent to the active site. All 20 designs expressed as soluble proteins in Escherichia coli BL21(DE3) and were purified, and all showed measurable CO2-reduction activity when assayed at pH 7.5 and 25 plus or minus 4 degrees Celsius; C20 reached 80.8 plus or minus 2.1 mU/mg against 32.0 plus or minus 3.5 mU/mg for wild-type CtFDH, while the two best literature variants reassayed under the same conditions gave D188R below wild type and only a modest gain for G93H/I94Y.
Perspective
Experimental validation is confined to the NAD+-dependent formate dehydrogenase reaction that reduces CO2 to formate, with activity measured at pH 7.5 and 25 plus or minus 4 degrees Celsius against wild-type CtFDH and two literature-reported CtFDH variants; other reactions were assessed only through computational structure and function predictions.
Structure prediction and CLEAN functional classification show that designs are computationally plausible but do not establish catalytic activity; the training data come from public databases whose annotations are often homology-based rather than direct activity measurements, which introduces uncertainty into the relationships the model learns; and why the 11 substitutions in C20 raise activity remains a structural hypothesis that the authors say requires systematic mutational studies for concrete evidence.