Teaching protein AI how fast an enzyme works

An enzyme can keep a familiar overall shape while its working speed changes enormously. That makes a difficult assignment for AI: recognising a protein family and predicting a member’s performance require different information.[2]

An NIH highlight on 8 October revisited research by Margaux Pinney’s team on adenylate kinase, an enzyme involved in cellular energy regulation. The underlying Science paper was published on 12 June 2025.[2][1]

Researchers selected 193 natural versions from bacteria and archaea. The 175 passing their measurement filters spanned turnover rates of 1–803 reactions per second under the assay conditions, despite similar predicted overall folds.[1]

A reaction needs a clock

A turnover rate, written kcat, describes the pace per enzyme when substrate is abundant. Measuring it requires more than observing a bright reaction: the amount of enzyme must also be known. The HT-MEK microfluidic platform, introduced in 2021, measures both. It produces and captures tagged proteins in tiny chambers, then tracks fluorescent signals as reactions proceed.[3]

The tags provide an estimate of enzyme concentration. Repeating measurements at different substrate concentrations lets researchers fit a kinetic curve. Valves isolate the chambers and control the start of reactions, allowing many variants to be examined in parallel with repeated measurements on the same device.[3]

Consider two hypothetical chambers producing product at the same measured rate. If one contains twice as much enzyme, its molecules are, on average, working at half the pace. Equal output would conceal unequal performance. Normalising by enzyme concentration is what turns a measurement of the whole chamber into information about the catalyst.

Adenylate kinase also moves as it works: opening two domains to release product is its rate-limiting step. A static structural resemblance therefore leaves out part of the job.[1]

Giving a protein map a performance dimension

The researchers examined ESM-2, a large protein language model. Its representations grouped some structural features well, but proximity to the nearest five sequences was essentially uninformative about turnover rate. Nearby proteins on that map could work at very different speeds.[1]

ProteinNPT offers a way to add measured properties to sequence knowledge. Its original NeurIPS paper describes a model that handles sequences and their labels together. Attention connects positions and property labels within each protein, then connects proteins across the batch. During training, it learns to recover hidden sequence elements and hidden property values.[4]

At prediction time, proteins with unknown properties are presented alongside labelled training examples. Those known measurements remain part of the context used to estimate the missing value. The architecture can therefore draw on both a protein’s sequence representation and its relationship to experimentally characterised examples.[4]

In the ADK tests, ProteinNPT combined pretrained representations with measured kinetic labels and outperformed the lightweight regressors compared. Prediction improved as more measurements were supplied; evaluation held out 20% of the dataset across repeated samplings.[1]

This changes what a map of proteins can be used for. Grouping by resemblance helps organise a family. Ranking by a measured activity helps choose among its members. The same sequence information can support both tasks, but the second needs a definition of performance—and observations that attach that definition to real examples.

The experiments supply that missing dimension. They compare natural solutions to the same catalytic task under a common assay, giving the learning system a quantitative account of how those solutions behave. A useful protein model can then learn from the difference between looking alike and working alike.[2]

Report an error or suggest a correction