GENzyme

An enzyme is not defined by what it binds but by what it does to what it binds: a substrate enters the pocket, bonds break and re-form, and a chemically different product leaves. Design tools built around binding optimise for a static lock-and-key fit, which describes an inhibitor complex well and a catalytic cycle poorly, discarding induced fit and the substrate-to-product transition. GENzyme inverts the conditioning: given the SMILES strings of a substrate and its product, it returns a catalytic pocket, a full enzyme structure and sequence built around that pocket, and a predicted enzyme-substrate complex.

GENzyme was released in November 2024 by a group led from McGill and Shanghai Jiao Tong University, with co-authors at Mila, HKUST, Yale, the Institute for Protein Design and DeepMind. The pipeline runs in three stages: an SE(3) flow-matching module that generates the catalytic pocket and co-designs its sequence, an inpainting stage that grows the rest of the enzyme around that pocket and inverse-folds the full sequence, and a docking module that screens substrate conformations against the designed pocket. It ships as a single pretrained checkpoint, so targeting a new transformation is a configuration edit rather than a training run.

It sits in a growing family of enzyme-specific generative models alongside EnzyGen2, which conditions co-design on a bound ligand, and EC-class-conditioned sequence generators such as ZymCTRL: here the design objective is a transformation, not a class label or a static partner.

#Key Features

• The reaction is the specification: Conditioning on substrate and product SMILES points the model at transformations with no characterised natural catalyst, not only at reactions already assigned an EC number.
• Pocket first, scaffold second: Generating the pocket alone, then inpainting it into a complete enzyme, concentrates the generative budget on the active site.
• Enzyme-specialised language model: EnzymeESM, a fine-tune of ESM-3 on enzyme data, drives inpainting and full-enzyme inverse folding, with masking targeted at catalytic residues.
• Binding and screening in the loop: A docking module scores substrate conformations against the designed pocket, so each design arrives with a proposed binding pose.
• Design, refinement and repurposing: An existing enzyme structure switches the same checkpoint from de novo generation to refinement or repurposing.

#Technical Details

Training uses EnzymeFill, a curated corpus of 328,192 enzyme-reaction pairs spanning 145,782 unique enzymes and 17,868 unique reactions, assembled from Rhea, MetaCyc and BRENDA. Catalytic pockets were extracted from predicted structures by transplanting ligands from homologous complexes within a 10 Å radius; the 128,940 pairs whose pockets reach 64 residues form the training set. Pocket generation combines conditional flow matching on SE(3) residue frames with a continuous-time Markov chain over amino acid types, regularised by FAPE, lDDT, TM and violation losses, and represents the substrate as a 2D molecular graph so that no precomputed conformer is required. Sequence design uses a PiFold inverse-folding head fine-tuned on the pocket data; inpainting is masked discrete diffusion over ESM-3 structure tokens. The docking module, Uni-Mol Docking v2, is the only component not trained end to end.

Evaluation covers 30 enzyme-reaction pairs, five per EC class, clustered at a 10% homology threshold, with eight designs sampled per reaction and filtered on pTM and pLDDT. At the pocket level GENzyme reproduces reference geometry more closely than the reported baselines — on EC1, top-1 RMSD of 1.88 Å and TM-score of 0.67, against 3.71 Å and 0.28 for RFdiffusionAA. Whole-enzyme designs diverge further from their natural counterparts than the pockets do. Function is assessed by prediction rather than assay — optimal pH, turnover number and mutation stability.

#Applications

The reaction-first framing suits metabolic engineering and biocatalysis, where the starting point is a desired chemical step — a substrate to convert, a product to reach — and no suitable enzyme is known. Protein engineers can also supply an existing structure and ask for a refined variant or one redirected at a new substrate — the common practice of starting from a scaffold that already folds. Because every design arrives with a pocket, a structure and a docked complex, filtering can weigh binding geometry alongside sequence plausibility.

#Impact

GENzyme is a preprint, and its authors are explicit that it is a step rather than a solution: no designed enzyme has been expressed or assayed, and every reported gain is computational. The pretrained checkpoint is distributed through a shared drive link rather than a versioned release, the EnzymeFill corpus has not been published alongside the code, and the licence prohibits commercial use and use in drug discovery. Running the pipeline also requires gated access to ESM-3 and a separate install for the docking module. Its contribution is the framing — a catalytic reaction, rather than a binding partner or a functional label, as the conditioning signal for protein generation — with evidence that pocket geometry survives it.