Given only an enzyme, a generator writes the full reaction: 25.0% of metabolite-recombination pairs recovered exactly within 400 candidates
Synopsis
Enzyme function prediction has relied on homology search or on mapping enzymes to known reactions, so it cannot output a reaction absent from its training data. A model conditioned on enzyme sequence and AlphaFold structure and generating a reaction metabolite by metabolite recovers the complete correct reaction for 25.0% of metabolite-recombination test pairs within 400 sampled candidates on the reaction-disjoint EnzymeMap_x benchmark, and for only 1.6% of pairs containing a metabolite unseen in training. A matched ablation finds no measurable effect of the structural channel on exact recovery; vocabulary coverage and chemical initialisation account for most of the gain.
Interpretation
A generative model can write a complete reaction, reactants and products, from the enzyme alone rather than selecting from a supplied reaction library. BLASTp, DIAMOND, CLIPZyme and Horizyn by design return only reactions already in training or supplied at inference, and the closest prior model, RxnPred, reports fingerprint similarity on held-out enzymes but fails to recover the correct answer on unseen reactions. On the 2,349 test pairs of the reaction-disjoint EnzymeMap_x benchmark, full-reaction recovery within 400 samples is 14.30% (336 pairs) and product-set recovery is 27.93%; shuffling the enzyme input drops these to 0.38% and 0.77%, against 0.81% for ranking training product sets by frequency alone.
The model encodes enzyme sequence and AlphaFold structure into one residue-level representation, and an autoregressive decoder writes the reaction one metabolite at a time. The structural stream is an ESM-2 sequence stream concatenated with a GVP-GNN graph network run over a Cα nearest-neighbour graph, and the decoder reads the resulting 2L+2-length sequence through cross-attention. Vocabulary embeddings are initialised from ECFP4 fingerprints plus 40-dimensional RDKit descriptors reduced by PCA to 512 dimensions; within 25 samples, product-set recovery is 8.39%, against 6.90% for ChemBERTa initialisation and 6.13% for random initialisation.
Split by whether metabolites appear in training, the model's generalisation is mainly compositional. Earlier generative work reported emitting recombinant reactions but at low frequency, without stratifying evaluation by metabolite novelty. Within 400 samples, 319 of 1,275 recombination pairs (25.02%) are fully recovered; 17 of 795 pairs needing an unseen product metabolite (2.14%) are recovered; none of the 279 pairs whose unseen metabolite is a reactant is generated.
Perspective
Evaluation is confined to the 2,349 test pairs of the reaction-disjoint EnzymeMap_x benchmark under stereo-respecting exact multiset matching; 76.7% of test pairs join to a reaction rule, and 99.6% of those rules also occur among training reactions, so the benchmark tests recombination of known reaction types over new substrate and co-substrate combinations. The authors propose attention analysis, a leave-one-organism-out split and EC-based or rule-based splits to separate enzyme-level from reaction-type generalisation.
The vocabulary is transductive: metabolites that appear only in test are initialised from their own chemical structure before training, so novel-metabolite recovery is selection from a supplied vocabulary rather than open-vocabulary generation. Of the 1,277 test enzymes, 818 also occur in training, and enzyme-level scoring would raise recovery from 14.3% to 71.2%, with 98.5% of the newly credited pairs resting on memorisation. The top-1 re-ranking result depends heavily on assuming the answer is a new reaction; without that assumption top-1 is 0.21%. The authors state it is unclear whether the performance limits come from the architecture or from the training information available.