CatRange gives enzyme predictions a deliberately wider scale

CatRange predicts enzyme kinetics in ranges spaced by factors of ten. Published in September and typeset on 13 October, the model pairs protein and substrate representations with decision-tree classifiers. Its training combines measured records with 5,278 synthetic catalytic-site variants calibrated against experimental examples.[1]

Turnover describes how quickly each enzyme processes substrate when substrate is abundant. A small sequence change can have several consequences. A 2021 study examined 1,036 variants of the enzyme PafA and found substantial functional effects in 702. Some substitutions disturbed catalysis; others increased the fraction of protein trapped in an inactive, misfolded state.[2]

The researchers varied the conditions during protein production separately from the conditions during the reaction. This helped distinguish an enzyme made incorrectly from one whose correctly folded form worked differently. For some variants, producing the protein at a lower temperature restored activity. The history of its formation affected what the subsequent assay measured.[2]

CatRange’s PafA test withheld 310 variants while training on other variants of that enzyme. One class held 59.7% of the test cases, so always choosing it would beat the model’s 51.3% exact-class accuracy. Accepting neighbouring classes lifted CatRange to 95.2%.[1]

The original PafA experiments also found connected regions of consequential residues reaching from the active site to the protein surface. Changes far from the chemical reaction could alter its performance. Those measurements supply a demanding test of whether sequence-based predictions capture the consequences of a mutation, rather than merely recognise a familiar protein.[2]

CatRange still underestimated severe activity losses in this landscape. Its coarse classes discard distinctions within a class. The evaluation shows why a high score for neighbouring ranges needs a closer examination of which mutations the model can distinguish.[1]

Report an error or suggest a correction