KcatNet

The turnover number kcat — how many substrate molecules one enzyme molecule converts per second at saturation — is the parameter that turns a metabolic network diagram into a quantitative model. Enzyme-constrained genome-scale metabolic models need one kcat per reaction, and almost no organism has them. Even in Escherichia coli, the most measured organism in biology, in vitro kcat values exist for only about 12% of enzymatic reactions, and the values that do exist span orders of magnitude depending on assay conditions. There is no high-throughput assay, so the gap has to be filled computationally.

KcatNet, from Jiangning Song's group at Monash University with collaborators at Beijing University of Chemical Technology, Virginia Commonwealth University, and the University of Melbourne, predicts kcat from an enzyme sequence paired with one substrate SMILES string. Its design point is geometric: rather than pooling a whole-protein embedding, it builds a residue graph from predicted contacts, partitions that graph into spatially contiguous clusters, and uses cross-attention to ask which clusters interact with which parts of the substrate. The clusters that light up correspond to catalytic pockets, so the prediction comes with a residue-level explanation of where the model thinks catalysis happens.

Earlier predictors — DLKcat, TurNuP, UniKP — work from sequence and substrate features without this structural decomposition, and generalize poorly to high-kcat enzymes and to sequences unlike their training data. KcatNet was published in Genome Biology in 2026 after a bioRxiv preprint the previous year, with code released on GitHub under GPL-3.0.

#Key Features

• Structure without structures: Residue contact maps are generated by a protein language model, so the geometric machinery applies to any sequence, including enzymes with no experimentally resolved structure.
• Residue partition module: Graph convolution plus graph partitioning groups spatially proximal residues into clusters, giving the model substructures rather than individual residues to reason over.
• Two-level substrate encoding: Substrates are represented both atomically — nine element groups plus degree, implicit hydrogens, hybridization, and chirality — and molecularly via a pretrained SMILES transformer, with message passing between the two levels.
• Interpretable catalytic pockets: Cross-attention weights between residue clusters and substrate produce importance scores that localize the residues driving predicted activity.
• Mutation sensitivity: The model distinguishes kcat between wild-type and point mutants of the same enzyme, which sequence-pooled predictors typically cannot.

#Technical Details

Residue features come from two pretrained protein language models — ProtT5, producing 1024-dimensional embeddings, and ESM-2, producing 1280-dimensional embeddings — with ESM-2 also supplying the residue contact map that defines graph adjacency. Substrate atoms carry 43-dimensional feature vectors; the molecular-level SMILES transformer embedding is 1024-dimensional. Graph convolution updates residue embeddings, Atoms-to-Molecule and Residues-to-Clusters pooling produces the two interacting representations, and a cross-attention module refines both before unpooling and a final MLP emits the log-scale prediction. Training data comprise 11,757 curated kcat entries covering 11,288 unique reactions and 7,441 unique enzymes, assembled from BRENDA, SABIO-RK, and UniProt with substrate SMILES resolved through PubChem, split 80/20. On the held-out test set KcatNet reached a Pearson correlation of 0.84, R² of 0.69, RMSE of 0.78, and MAE of 0.55 — an 18% R² improvement over UniKP, and significantly better than DLKcat on both RMSE and MAE. On mutant enzymes it achieved R² of 0.71 against 0.59 for wild types on the same test split.

#Applications

The immediate use is parameterizing enzyme-constrained genome-scale metabolic models for organisms with no measured kinetics. The authors rebuilt ecGEMs for yeast and fungal species using KcatNet predictions in place of the EC-number-based value transfer the original pipeline relied on, and improved proteome and growth-rate predictions across four yeasts — Saccharomyces cerevisiae, Kluyveromyces lactis, K. marxianus, and Yarrowia lipolytica — reducing mean squared error against published data in 16 of 22 species-condition combinations relative to DLKcat-parameterized models. The second use is enzyme engineering triage: screening a saturation-mutagenesis library in silico and shortlisting variants worth expressing. Applied to 10,545 single-point missense variants of a bacterial GH13_31 α-glucosidase, nine predicted-beneficial mutations were selected for expression and assay, and seven showed higher kcat than wild type.

#Impact

The α-glucosidase result is the part that matters most: an in-silico kinetic screen with a 7-of-9 experimental hit rate is directly usable in directed-evolution campaigns, where cost is dominated by how many variants have to be made. More broadly, KcatNet moves kcat prediction from a sequence-similarity problem toward a structural one, and the interpretability that falls out — residue clusters ranked by contribution — makes predictions inspectable rather than only scored. Limits remain. Training data omit pH and temperature because the source databases largely lack them, so predictions are condition-agnostic. Under the stringent 95%-identity-centroid holdout proposed by Kroll and Lercher, every model tested including KcatNet fell below 0.7 Pearson correlation, with TurNuP best — generalization to unfamiliar enzyme families remains this field's open problem.