RealKcat

Enzyme turnover number (kcat) and Michaelis constant (KM) are the two numbers that determine how fast an enzyme works and how tightly it holds its substrate, and they gate everything from metabolic model construction to enzyme engineering campaigns. Measuring them is slow and condition-dependent; predicting them from sequence has become a small sub-field. The recurring failure mode of sequence-based predictors is that they are largely insensitive to mutations — substituting a catalytic residue that abolishes activity often barely moves the prediction away from the wild-type value, which is exactly the regime an enzyme engineer cares about.

RealKcat, developed by groups at the University of Nebraska-Lincoln, Iowa State University, and the University of South Florida, attacks that failure directly. It reframes kinetics prediction as classification over order-of-magnitude bins rather than regression onto a continuous value, on the argument that the underlying measurements carry more than an order of magnitude of inter-laboratory spread anyway. That formulation makes the model tolerant of assay noise while leaving it free to register the large, real drops in activity that catalytic mutations cause.

RealKcat is a distinct effort from KcatNet, the other enzyme-kinetics predictor in the catalog: different institutions, a classification rather than a graph-cluster regression formulation, and a second output — KM — that KcatNet does not predict.

#Key Features

• Mutation sensitivity by construction: The training set is augmented with 5,278 synthetic catalytic variants generated by alanine substitution at annotated catalytic residues, so the model sees what a knocked-out active site looks like.
• Two kinetic parameters: Separate classifiers predict catalytic turnover and substrate affinity from the same sequence-plus-substrate feature space.
• Order-of-magnitude classification: Collapsing kinetic values into bins makes the task robust to the experimental variability that dominates aggregated kinetics databases.
• Hand-curated training data: KinHub-27k was assembled by reconciling BRENDA, SABIO-RK, and UniProt entries against 2,158 primary papers, rather than accepting database records at face value.
• Structure-free by design: Structural descriptors were deliberately excluded after the authors found that naive structural features degraded generalisation, so only a sequence and a SMILES string are needed at inference.

#Technical Details

RealKcat represents each enzyme-substrate pair by concatenating ESM protein language model embeddings of the enzyme sequence with ChemBERTa embeddings of the substrate SMILES, producing a unified feature space that is then classified by gradient-boosted trees (XGBoost). Separate models are trained for kcat and KM. The training corpus, KinHub-27k, contains 27,176 experimentally reported enzyme-substrate entries consolidated from BRENDA, SABIO-RK, and UniProt and verified across 2,158 primary sources, plus the alanine-scan augmentation set. Evaluation spans cross-validation, a held-out split, out-of-distribution enzymes, and few-shot settings, including a dense mutational landscape of the alkaline phosphatase PafA, on which the model tracks both the direction and the magnitude of mutation-induced changes while retaining discrimination between wild-type and mutant contexts. Performance is reported as e-accuracy, a metric that counts a prediction correct when it falls within one order of magnitude of the measured value.

The repository ships trained kcat and KM checkpoints under the MIT licence, along with the KinHub-27k data, PafA and out-of-distribution test sets, and two inference notebooks — one Colab-hosted for zero-setup use and one for local runs.

#Applications

The intended users are enzyme engineers and metabolic modellers. In directed evolution and rational design, RealKcat can pre-screen variant libraries for the substitutions that preserve or improve turnover, concentrating assay effort on the survivors. In genome-scale metabolic modelling, enzyme-constrained models need kinetic parameters for reactions that have never been assayed, and a mutation-aware predictor lets those constraints be updated when a strain carries engineered variants. The same predictions support biomanufacturing route selection and mechanistic questions about which residues carry catalysis.

#Impact

RealKcat's argument — that the task formulation and the curation matter more than the architecture — is a useful corrective in a field where kinetics predictors have often been benchmarked on the same noisy database dumps. Its principal caveat is that the reported gains are in silico, validated on curated held-out data and a published mutational landscape rather than on new wet-lab measurements, and the paper remains a preprint. Order-of-magnitude output is also coarser than a point estimate, which suits triage better than it suits quantitative model parameterisation. Public checkpoints, data, and a one-click notebook under MIT make the claims easy for other groups to test on their own enzymes.