Measuring enzyme kinetics takes experiments, but researchers are building tools to help decide which experiments to run. In a March 2024 paper in BMC Bioinformatics, Xiao He and Ming Yan introduced GraphKM, a model that estimates the Michaelis constant from an enzyme’s amino-acid sequence and its substrate’s chemical structure.
The model covers both wild-type enzymes and variants with mutations. Its results suggest a useful starting point for screening, with a clear limitation: a predicted value cannot stand in for a measurement under specified laboratory conditions.
What the Michaelis constant tells us
For a reaction that follows Michaelis–Menten kinetics, the Michaelis constant, KM, is the substrate concentration at which the reaction rate reaches half its limiting value. It helps describe how reaction rate changes as more substrate becomes available.
KM is not generally the same quantity as a binding dissociation constant. It reflects the reaction mechanism, so a smaller value does not automatically mean tighter substrate binding. The IUBMB’s enzyme-kinetics recommendations make this distinction explicit.
Connecting protein sequences and chemical graphs
GraphKM represents a substrate as a molecular graph: atoms form the nodes and bonds connect them. A pretrained protein language model supplies a numerical representation of the enzyme sequence. Neural networks combine these inputs, and gradient boosting is used in the prediction pipeline.
The researchers assembled 19,754 cleaned enzyme–substrate entries from BRENDA and SABIO-RK. These included 11,314 wild-type entries and 8,440 mutant entries. They reserved 20% of this dataset for testing and compared four graph-network architectures.
The combined GAT–GCN architecture performed best overall in that comparison, with R² of 0.622 and a root-mean-square error of 0.792 on log-transformed KM values. Those are dataset-level performance measures, not a guarantee that every prediction is close to its experimental counterpart.
Testing beyond the original dataset
The authors also collected an independent literature dataset, HXKm, with 443 entries after removing exact enzyme-sequence/substrate pairs that overlapped the training set. GraphKM reached a Pearson correlation of 0.589 there, compared with 0.227 for the earlier KM_prediction model tested on the same dataset.
This additional test is valuable, but a correlation of 0.589 shows only moderate agreement. Removing identical pairs also does not necessarily remove closely related enzymes. Performance on unfamiliar protein families or chemistry needs to be established separately.
The four panels in the figure compare predicted and recorded log-transformed values for the original test set. Points nearer the diagonal indicate closer agreement. They should not be confused with the separate HXKm validation results.
Why experimental conditions still matter
The same enzyme and substrate can have different measured KM values at different temperatures or pH values. When several measurements described one sequence–substrate pair, the researchers retained the maximum value. Their model uses sequence and substrate inputs, rather than predicting how the kinetic parameter changes with assay conditions.
For researchers, the practical use is to help prioritise candidates and generate hypotheses. Promising estimates still need enzyme assays, with the relevant conditions recorded. The authors provide research code and data; using that implementation requires a computational workflow.
Related research: DPI_CDF examines protein sequences for drug-target clues.
Original research
He, X. and Yan, M. (2024). GraphKM: machine and deep learning for KM prediction of wildtype and mutant enzymes. BMC Bioinformatics, 25, 135.
Update note, 23 September 2026: This archived report has been revised to clarify the meaning of KM, the validation results, assay-condition limits and the figure attribution.
