Finding an enzyme for the substrate in hand

Suppose the molecule on a chemist's desk is fixed, and the unanswered question is which enzyme might transform it. An enzyme name alone is not the evidence that question needs. A 2025 study approaches the search by connecting measured substrate–enzyme pairs with models that can recommend candidates in either direction.¹

Build the connections before predicting them

Paton and colleagues screen 111 substrates against 314 non-haem iron enzymes, reporting 215 new compatible reaction pairs.¹ The central object is therefore a matrix of tested combinations, rather than a list of enzymes associated with broad functional annotations. The work concerns a particular oxidative enzyme family, not arbitrary enzyme chemistry.¹

We find that design choice important for interpreting a recommendation. If a proposed match is wrong, the record should allow a reader to ask where the model's information came from. Was there a nearby tested substrate, a related enzyme sequence, or only a broad similarity argument? Those explanations would lead to different follow-up questions.

Navigate in either direction

The CATNIP interface accepts a molecular structure to recommend enzymes, or an enzyme sequence to recommend substrates.¹ Its models use relationships in chemical space and protein sequence space to rank possible partners.¹ A ranked list is a proposal about where to look next; it does not turn every suggested pair into an observed reaction.

Consider how we would report a prospective use. The story should retain the rank of every candidate tested, the reason that shortlist was chosen and the experimental outcome of each pairing. A later success would then have a denominator. Without that record, a striking example could obscure how much searching was necessary to find it.

The distinction also matters when a test is negative. Our preferred account would state what was actually measured and whether the experiment was informative enough to evaluate the proposed pairing. An uninformative test should remain distinguishable from a demonstrated failure under defined conditions. That is an editorial requirement we would bring to follow-up reporting, not a claim that this study resolves every such ambiguity.

What counts as useful transfer?

The paper reports experimentally observed compatibility and builds predictions from it.¹ We would keep the evidence for those two stages separate when assessing transfer to a new molecule. A promising neighbour is a rationale for an experiment, while the new experiment is what adds another actual connection to the map.

For our next question, we would ask how a recommendation changes a chemist's starting shortlist. Does it suggest an enzyme the researcher would otherwise have overlooked? Does it offer a clear reason to prefer one early test over another? These are practical questions that a comparison could answer without claiming a universal catalyst finder.

This is a launch retrospective on a published experimental and computational workflow. AiChemEx has not run CATNIP predictions, expressed these enzymes or repeated the reactions. The useful shift we see is from naming an enzyme family to documenting and testing a specific proposed partnership, while keeping the uncertainty attached to that partnership.