When a mutation changes the enzyme, the prediction should change too.

An enzyme can look almost identical in sequence and behave very differently after a catalytic-site mutation. How can a model reflect that difference without implying more numerical precision than the evidence supports?

I developed CatRange with collaborators, connecting biochemical data curation, enzyme–substrate representations, model development, evaluation, and usable inference.

• Protein representations
• Substrate chemistry
• Gradient-boosted classification
• Mutation-aware evaluation
• Research inference

The PNAS Nexus publication and public repository provide the research and software record. The contribution is a range-based, mutation-sensitive modeling framework—not an experimental measurement of every predicted variant.

A predicted regime is not a measured kinetic constant. Neighboring-bin recovery is an evaluation criterion, not automatically a calibrated confidence interval. Assay context, sequence coverage, and the deployed bin definitions still matter.