How to Use RealKcat

RealKcat is a multimodal enzyme-kinetics predictor that classifies enzyme-substrate pairs into experimentally motivated kcat and Km ranges. The released model is designed for enzyme variant and substrate-panel screening, where researchers need fast comparative kinetics information before committing to wet-lab assays.

On Neurosnap, researchers provide one enzyme Input Sequence together with one or more Input Molecules. The service returns both kcat and Km range predictions in one run, which makes it useful for substrate triage, variant prioritization, and early enzyme-engineering decisions when exact steady-state parameters are not yet available.

The RealKcat inference workflow combines two pretrained representation models: ESM-2 for the enzyme sequence and ChemBERTa for the substrate SMILES string. Those embeddings are concatenated, globally standardized, and then scored by parameter-specific XGBoost classifiers that map each enzyme-substrate pair into discrete kinetic bins. This is a different objective from point-regression models such as CatPred: RealKcat is optimized to provide robust range-level classification rather than a single exact value.

That range-based formulation is useful when the practical question is comparative, such as whether a substrate is likely to sit in a low-, medium-, or high-activity regime or whether a variant plausibly shifts Michaelis behavior enough to warrant experimental follow-up. The preprint emphasizes robustness on enzyme variants and curated enzyme-kinetics data, which aligns well with screening workflows in biocatalysis and protein engineering.

On Neurosnap, RealKcat should be interpreted as a ranking and triage tool. Researchers can compare returned kcat and Km ranges across a substrate panel, identify pairs that look catalytically promising or weak, and then decide which combinations deserve purification, kinetic assays, or deeper structure-guided redesign.

• Joint kinetic readout: RealKcat predicts both kcat and Km ranges in one run, which is useful when catalytic efficiency and substrate handling both matter.
• Multimodal representation stack: ESM-2 encodes the enzyme sequence while ChemBERTa encodes the substrate, giving the model richer context than sequence-only or molecule-only baselines.
• Panel-friendly workflow: One enzyme can be screened against many substrates in a single job, matching common early-stage enzyme characterization studies.
• Range-based interpretation: Discrete kinetic bins are practical for experimental triage, especially when the goal is to prioritize measurements rather than trust a single exact predicted number.