Did AI Discover an Enzyme? Start With What Was Tested

Key takeaways

• Finding a new sequence and demonstrating enzyme activity are distinct achievements.
• CRISPR-like repeats alone do not establish gene-editing capabilities.
• AI’s contribution can range from writing analysis code to identifying candidates and proposing hypotheses.
• Claims about biological function must stay within the conditions actually tested.

Suppose Claude scans genetic data and flags a candidate for a new enzyme system. That could be a meaningful scientific contribution, but the phrase “AI discovered an enzyme” leaves the most interesting part unexplained: what do we actually know it does?

A familiar pattern is a starting point

CRISPR takes its name from a distinctive arrangement of short, repeated DNA sequences. Examining those repeats alongside nearby genes can give researchers clues about a possible biological system.

But “CRISPR-like repeats” describes a resemblance. It does not establish that the system performs the same function, much less that it can edit genes.

An enzyme is a biological catalyst: it speeds up a chemical reaction. To establish how a proposed enzyme system works, researchers need to identify the relevant molecules and show what reactions they participate in.

At the sequence-analysis stage, the precise description is a candidate for functional testing. That is a useful result. It gives researchers something concrete to investigate.

Give the AI credit for the job it did

“AI discovered it” can compress several very different contributions into one headline.

Imagine a researcher asks Claude to write code for analyzing genetic data. If the code helps process the data, Claude has assisted with analysis. If it also helps identify promising candidates, its contribution extends to the search itself.

Now suppose Claude proposes a hypothesis about what a candidate does. That adds another contribution, but the hypothesis still needs supporting evidence. A fluent explanation is not a measurement of biological activity.

Computational work can also establish a previously unrecognized family of sequences. That can be a scientific discovery in its own right. There is no rule that every discovery must begin at a lab bench.

The distinction worth preserving is between sequence discovery and functional validation. Both can matter. They answer different questions.

The experiment sets the limits of the claim

If researchers claim a candidate cuts DNA, they need evidence that it actually cuts DNA. They also need controls that help establish whether the candidate caused the observed effect.

If the proposed system works with RNA, the evidence must support that relationship too. A plausible predicted protein structure can strengthen the case for running an experiment. It does not substitute for measuring activity.

Even a successful experiment has boundaries. A reaction demonstrated in a test tube establishes activity under those conditions. Whether the system works inside cells remains a separate question. So does whether it acts accurately on the intended target.

Calling it a gene-editing tool raises the bar further. Researchers must move beyond showing that a reaction occurs to showing that they can direct it to the desired location.

The distance between those milestones is where a headline can get ahead of the biology.

Ask what is new—and what the AI added

“New” needs a definition too.

It might mean a previously unknown sequence. It might mean an enzyme with different properties within an established family. A claim about an entirely new mechanism requires evidence for that mechanism.

AI’s contribution deserves the same precision. Did it reduce the time spent on analysis? Surface a candidate that researchers might otherwise have missed? Propose a hypothesis that proved worth testing?

Analysis records and experimental results help answer those questions. They can show which inputs produced a candidate, what researchers chose to pursue, and how those choices connect to the eventual findings.

That makes it possible to describe the contribution concretely, rather than assigning all the credit to either the model or the people using it.

A promising candidate can be a valuable scientific result before its function is established. The useful question is how far the evidence takes us: to a new sequence, a demonstrated reaction, or a tool that works as intended. Read “discovered” as the beginning of that question, not the answer.