EnzymeFlow

A binding pocket and a catalytic pocket look alike and do different work. An inhibitor enters a binding site, sits, and leaves chemically unchanged; a substrate enters a catalytic site and leaves as a different molecule, having passed through a transition state the surrounding residues had to stabilise. Generative pocket-design models trained on protein-ligand complexes learn the first case, so conditioning one on a substrate says nothing about the transformation the pocket is supposed to perform. EnzymeFlow makes the reaction itself the conditioning signal: it takes a substrate and the product that substrate becomes, and generates the catalytic pocket that runs that step.

EnzymeFlow was released in October 2024 by a group led from McGill and Shanghai Jiao Tong University, with co-authors at Mila, Université de Montréal, HKUST, the Institute for Protein Design, Microsoft Research and DeepMind. The pocket is generated as a set of SE(3) residue frames by flow matching, and amino acid identities, co-evolutionary tokens and the Enzyme Commission class are co-designed inside the same sampling trajectory rather than annotated afterwards by a separate classifier.

It designs the catalytic pocket alone and not the enclosing protein. GENzyme, from the same lead author, adopts this module as its first stage and inpaints a complete enzyme around the pocket it produces.

#Key Features

• The reaction is the specification: The substrate is supplied as a 3D atom point cloud and the desired product as a 2D molecular graph, so the design target is a chemical step rather than a static complex.
• Enzyme-reaction co-evolution: A co-evolutionary transformer, coEvoFormer, runs a joint alignment of homologous enzyme sequences against their reaction SMILES, letting the model treat substrate drift and sequence drift as one coupled signal.
• EC class in the generative loop: A discrete flow interpolates the Enzyme Commission class from a masked state to a prediction, so intended function shapes the structure instead of being read off it later.
• Hierarchical pretraining: Training proceeds from bare protein backbones, to ligand-conditioned binding pockets with geometric regularisers, to catalytic pockets, accumulating structural priors that the enzyme corpus alone is too small to supply.
• EnzymeFill corpus: A purpose-built dataset of 328,192 enzyme-reaction pairs covering 145,782 enzymes and 17,868 reactions, with pockets extracted from predicted structures by ligand transplantation rather than inferred from sequence.

#Technical Details

Sampling interpolates Cα translations linearly, rotations geodesically on SO(3), and amino acid types through a masking discrete flow. A 3D graph network encodes the substrate and a 2D graph network the product; cross-attention couples both to an Invariant Point Attention trunk of 20 blocks with 256-dimensional node and 128-dimensional edge embeddings. Backbone pretraining uses the protein corpus from FrameFlow, pocket pretraining uses PDBBind2020 with added ligand-surface intersection, pairwise-distance and binding-affinity losses, and fine-tuning uses the 53,483 EnzymeFill pairs clustered at 60% sequence homology. Pockets shorter than 32 residues are discarded, following the interaction quality threshold used by LigandMPNN. Training ran on four A40 GPUs.

Evaluation covers 100 enzyme-reaction pairs clustered at a 10% homology threshold and spread evenly across EC1 to EC6, with 100 pockets sampled per reaction. Top-1 constrained-site RMSD is 6.94 Å at a TM-score of 0.290, against 9.13 Å and 0.269 for RFdiffusionAA and 7.42 Å and 0.268 for PocketFlow; amino acid recovery is 0.216 and EC accuracy 0.280, the latter a 64.7% relative gain over RFdiffusionAA. Docked substrate affinity, scored with AutoDock Vina, is -5.03 for EnzymeFlow against -12.71 for RFdiffusionAA — a gap the authors read against the -4.65 measured for the reference natural pairs, since catalysis does not reward tight binding.

#Applications

The reaction-first framing fits biocatalysis and metabolic engineering, where the starting point is a transformation with no characterised natural catalyst. Each design arrives as pocket backbone geometry, a residue sequence and a predicted EC class, so candidates can be ranked on docking geometry and predicted function before any scaffolding work begins. The output also serves as a functional-motif specification for a downstream scaffolding or inpainting model, which is the role it plays inside GENzyme.

#Impact

EnzymeFlow remains a preprint with no wet-lab validation: every result is computational, and the absolute EC accuracy of 0.280 shows how far reaction-conditioned pocket design is from reliably specifying function. Only a reduced "mini" checkpoint is published, distributed through a shared drive link rather than a versioned archive, and the licence forbids commercial use and any use in drug discovery. Several training assets — enzyme MSAs, the higher-homology splits, the pocket and molecule structure folders — are referenced in the repository without resolvable download links, so retraining is harder than inference. Its durable contributions are the framing, in which a catalytic transformation rather than a binding partner conditions protein generation, and the EnzymeFill corpus, which its successor reuses.