Accurate prediction of kinetic parameters is crucial for understanding known and tailoring novel enzymes for biocatalysis. Current models fail to capture mutation effects on catalytically essential residues, limiting their utility in enzyme design. We grid-searched through ten model architectures (25,671 hyperparameter combinations) to identify a gradient-based additive framework called RealKcat, trained on 27,176 experimental entries curated manually (KinHub-27k) by screening 2,158 articles. Clustering catalytic turnover (𝑘𝑐𝑎𝑡) and substrate affinity (KM) by rational orders of magnitude, RealKcat achieves >
85% test accuracy, demonstrating highest sensitivity to mutation-induced variability thus far, and is the first-of-its-kind-model to demonstrate complete loss of activity upon deletion of the catalytic apparatus. Finally, state-of-the-art 𝑘𝑐𝑎𝑡 validation accuracy (96%) on alkaline phosphatase (PafA) mutant industrial dataset confirms RealKcat's generalizability in learning per-residue catalytic relevance.