EnzymeCAGE

Metabolic databases are full of reactions with no enzyme attached. KEGG and MetaCyc report that 40–50% of known enzymatic reactions lack an associated sequence — orphan reactions — and of the roughly 190 million protein sequences in UniProt, fewer than 0.3% are expert-curated. Existing responses are unsatisfying: Enzyme Commission classes compress distinct chemistries into one four-digit code, while sequence-similarity tools such as BLASTp and Selenzyme fail where they are needed most, on enzymes with low homology to anything characterized.

EnzymeCAGE, developed by Shuangjia Zheng's group at the Global Institute of Future Technology, Shanghai Jiao Tong University, with collaborators at HKUST, Sun Yat-sen University, Mila and MIT, scores enzyme–reaction compatibility directly. It takes an enzyme structure and a reaction and returns a single catalytic score between 0 and 1, so the same model answers both directions of the problem: given an uncharacterized enzyme, which reaction does it run; given an orphan reaction, which sequences in a candidate set are plausible catalysts. Because scoring is similarity-free, an enzyme can rank highly despite substantial divergence from anything in training.

Posted as a preprint in December 2024, it was published in Nature Catalysis in 2026.

#Key Features

• Pocket-level geometry, not whole-chain sequence: Catalytic pockets are extracted with AlphaFill (or P2Rank as a lighter alternative) and encoded as graphs whose nodes carry residue type, backbone coordinates, dihedral angles and side-chain torsions, with edges encoding atomic distances and orientations.
• Reaction-center-aware chemistry encoding: Substrate and product conformations are generated and the reaction center — the atoms whose covalent bonding changes — is identified and up-weighted, so the molecular graph network attends to the bonds actually being made and broken.
• Geometry-enhanced interaction modelling: A pocket attention module injects the pocket's distance map as an attention bias, and a center-aware cross-attention module couples pocket residues to substrate atoms, producing explicit 3D interaction features rather than a dot product between independent embeddings.
• Emergent catalytic site identification: Attention weights between residues and substrate reaction-center atoms recover catalytic regions validated against the Catalytic Site Atlas, despite catalytic-site labels never appearing in the training objective.
• Runs on predicted structures: Experimental and predicted structures are both supported, and a one-command mining pipeline turns a reaction CSV plus a directory of PDB or CIF candidates into a ranked table.

#Technical Details

Training data was assembled from Rhea, MetaCyc and BRENDA into roughly 1.5 million structure-informed enzyme–reaction pairs spanning more than 3,000 species, with two purpose-built negative-sampling strategies drawn from enzyme functional similarity and reaction templates. The architecture runs the enzyme and the reaction through separate trunks — a graph network over the pocket structure and a molecular graph network over substrate and product conformations — before fusing them; pocket and reaction embeddings are then combined with ESM-2 sequence embeddings for global enzyme context and DRFP reaction fingerprints, and passed to an output network trained end-to-end with binary cross-entropy.

Two held-out sets probe different failure modes. Enzyme-405 measures functional prediction on enzymes unseen during training; Orphan-335 measures enzyme retrieval for orphan reactions, those carrying no recorded catalyst when the training data was assembled. Baselines across the two span sequence search (BLASTp), pathway-design tools (Selenzyme), substrate-specificity predictors (ESP), EC classifiers (CLEAN) and reaction-conditioned screening (CLIPZyme). Five external sets — P450, terpene synthase, phosphatase, glutarate and withanolide — test family-specific transfer after light fine-tuning. Checkpoints and the dataset are distributed through Google Drive under a CC BY-NC 4.0 licence with added restrictions barring commercial use and any use in drug discovery or development.

#Applications

The direct use is biocatalyst discovery: a metabolic engineer with a target transformation supplies the reaction SMILES and a candidate structure library, and receives a ranked shortlist to express and assay. The reverse query — annotating an uncharacterized enzyme by ranking reactions — supports genome mining and the functional annotation of metagenome-derived sequences that homology search leaves blank. The pathway case study reconstructs glutarate biosynthesis end-to-end, the shape of the problem in synthetic biology, where a route is only as good as its weakest enzyme assignment. The recovered catalytic regions are a useful secondary output for mutagenesis planning.

#Impact

EnzymeCAGE is one of the clearer demonstrations that enzyme function prediction benefits from treating catalysis as a structural interaction problem rather than a classification or retrieval problem over sequences. Its practical reach is narrowed by two things. The licence forbids commercial use and explicitly excludes drug discovery and development, ruling out much of the industrial biocatalysis audience the work addresses. And inference has a heavy front end: pocket extraction requires AlphaFill, or P2Rank 2.5.1 with a Java 17 runtime, before any model code runs, and weights arrive through Google Drive rather than a versioned registry.