The protein budget behind AI's search for faster enzymes

A cell cannot solve every chemical bottleneck by making more enzyme. Proteins take up part of a finite cellular budget. Spend more of that budget on one reaction, and less remains available for others. Enzyme speed therefore has consequences well beyond the molecule doing the work.[3]

KcatNet tackles that speed question. The model was published in Genome Biology on 11 February; Monash University highlighted the research again on 5 October.[1][5]

Its target, the turnover number or kcat, counts substrate molecules converted per enzyme molecule per second when substrate supply is saturating.[1]

Putting a rate into the cell's accounts

A metabolic model maps the reactions a cell can carry out. Adding enzyme constraints asks what it takes to run them: each reaction needs enough functioning enzyme, and the combined demand must fit within the protein available. The GECKO modelling framework turns this idea into an explicit account of enzyme use across a network.[3]

The same production rate can therefore carry a different protein cost. Conversely, holding enzyme abundance fixed while increasing turnover raises that reaction's capacity. A whole cell must still supply the other reactions in the pathway, so the useful target is the enzyme whose improved performance relaxes a real constraint.[4]

This is why an incomplete catalogue of enzyme rates is more than a missing spreadsheet column. GECKO's developers describe filling kinetic gaps with measurements from other organisms or related reaction classes. Those substitutions make a model possible, but they also give a rate predictor a valuable job: supplying estimates tailored to the enzyme being modelled.[3]

From an amino-acid sequence to a shortlist

KcatNet's public implementation takes a pair of inputs: an enzyme sequence and a chemical description of one substrate. Protein language models turn the amino-acid sequence into numerical features; a separate molecular model represents the substrate. Pairing the two matters because a turnover estimate describes an enzyme acting on something, rather than an enzyme in isolation.[2]

The geometric model groups nearby amino acids and learns relationships between these regions and the substrate, helping it identify potentially relevant catalytic pockets.[1]

For a laboratory test, the team screened 10,545 α-glucosidase variants computationally. It selected nine residue positions, then randomly chose one predicted beneficial substitution at each. Seven of the nine tested variants had higher turnover than the original enzyme in assays using pNPG, a laboratory substrate, at 37°C.[1]

One tested variant reached 6.25 conversions per enzyme molecule per second, against 4.25 for the wild type. That is a calculated increase of about 47% under the assay conditions.[6]

That result supports a practical use: directing scarce experimental effort towards a manageable set of candidates. The model's public repository supplies prediction and training code, making the sequence–substrate workflow inspectable. The useful output is a numerical hypothesis that an enzyme experiment can accept or reject.[2]

KcatNet does not take pH or temperature as inputs. Its proposed rates still need testing under the conditions in which an enzyme will be used.[1]

The deeper engineering opportunity is to connect two decisions: which protein to build, and how a cell should spend its resources. A faster enzyme may be valuable because it leaves capacity elsewhere in the system. After a promising molecular design, the question becomes concrete: what work becomes possible when this reaction needs less of the cell?

Report an error or suggest a correction