Accurately modeling enzyme–substrate interactions in the presence of mutations—particularly those disrupting catalytically essential residues—remains a major challenge in computational enzymology, hindering advances in metabolic engineering and rational enzyme design. To address this gap, we present RealKcat, a catalytically aware machine learning framework that predicts biochemical reaction kinetic parameters (kcat, KM) with high fidelity and mechanistic resolution. Trained on a curated dataset of over 27,000 enzyme–substrate pairs spanning broad enzymatic diversity, RealKcat reframes kinetic prediction as a classification task across biochemically meaningful magnitude bins. Critically, RealKcat captures functional quenching effects resulting from point mutations at catalytic residues—outperforming conventional predictors in identifying enzymatic inactivation. In a validation test case on 1,016 mutants/variants of PafA (an industrially relevant alkaline phosphatase), the model distinguished functional from dysfunctional mutants with 53% accuracy for kcat and 93% for KM, achieving 96% and 100% agreement in order-of-magnitude classification, respectively. Beyond single-enzyme resolution, we integrate RealKcat within a variational kinetic modeling framework to dissect flux redistribution in isozyme-rich networks—networks where multiple enzymes catalyze the same reaction but exhibit differential expression, localization, and kinetic behavior. These redundant architectures pose a central challenge in pathway modeling, as they obscure isozyme-specific contributions and regulatory control. A key case study is the multi-KCS (β-ketoacyl-CoA synthase) and ELO (elongation of very long-chain fatty acids) enzyme system involved in plant cuticle lipid biosynthesis, where overlapping enzymes co-regulate the synthesis of protective lipid polymers. To probe this system, we not only apply RealKcat for in silico resolution of isozyme-specific kinetics under genetic and environmental perturbations, but also perform experimental isolation of combinatorial KCS/ELO isoforms to validate model predictions and quantify functional divergence. This integrated approach enables mechanistic dissection of pathway flux distribution and the identification of specific isozyme sets that confer metabolic robustness or bottleneck formation. Through this multi-scale application, RealKcat bridges molecular-level kinetic inference with systems-scale metabolic modeling—enabling predictive enzyme engineering, high-throughput variant prioritization, and synthetic pathway design. Ultimately, RealKcat establishes a versatile platform for rationalizing enzyme behavior in complex biological contexts, with broad utility across synthetic biology, computational biochemistry, systems pharmacology, and biomanufacturing.
Reference:
Sajeevan, K.A., Osinuga, A et al. (2025) ‘Robust Prediction of Enzyme Variant Kinetics with RealKcat’, bioRxiv : the preprint server for biology [Preprint]. Available at: https://doi.org/10.1101/2025.02.10.637555.
Funding: NIH MIRA Grant number 5R35GM143009; NSF IOS Grant number: 2212801
