What an AI model for genomes is actually doing

The phrase "AI model for biology" covers several different things, and the differences matter. It is worth separating them before asking what any of them can do.

Three different jobs

Structure prediction takes a protein's sequence and predicts the shape it folds into. This is the problem AlphaFold became famous for, and it is largely a solved problem for many single proteins — accurate enough that predicted structures are now a routine starting point rather than a curiosity. The 2024 Nobel Prize in Chemistry recognised this work alongside computational protein design.

Protein design runs the same logic backwards: describe the shape or the job you want, and get back a sequence that might fold that way. Designed proteins that had no natural ancestor have been made and tested in laboratories for several years now. Some work. Many do not, and the ones that do are validated at the bench, not in the model.

Sequence models for genomes are the newest of the three and the most often misdescribed. These are trained on very large collections of genomic sequence, and what they learn is statistical: given this stretch of DNA, what tends to come next, and which variants look ordinary versus strange. Recent models of this kind, such as those from the Arc Institute, work across whole genomes rather than single genes.

What "the model predicts" means

A genome language model does not know what a cell is. It has never seen a cell. It has seen an enormous amount of sequence, and it has learned the regularities in it — which arrangements of letters recur, which substitutions appear in living organisms and which never do.

That turns out to be genuinely useful, because evolution has already done a filtering pass. Sequences that break an organism tend not to be in the training data, so "unusual to the model" often correlates with "likely to cause a problem". This is the basis for using such models to flag variants of possible significance.

It is also the source of the main failure mode. The model is confident about what is typical. It has no access to what is true in a case it has not effectively seen before, and it cannot tell you the difference. A prediction is a hypothesis with a probability attached, not a result.

The part that gets skipped

In 2026 a group reported using genome-scale models to design viral genomes — bacteriophages, which infect bacteria, based on a well-studied laboratory virus. Of the designs tested, only a small fraction were functional. That ratio is the useful part of the story: the models proposed, and reality disqualified most of the proposals.

Work like this also prompted serious public discussion about safeguards, because the same class of tool points at organisms we very much do not want anyone designing. Researchers in this area have generally published alongside that discussion rather than after it, and screening of commercial DNA synthesis orders — a check on what actually gets manufactured — has become a more prominent part of the conversation as design tools improve.

A short list of things these models do not do

• They do not design an organism. No AI system has produced a novel free-living organism, and none is close.
• They do not replace experiments. Every claim above was settled in a laboratory.
• They do not explain their reasoning. A high score is not a mechanism.
• They do not know about anything absent from their training data, including most of the microbial world, which remains poorly sampled.

Why this is still good news

Set the hype aside and the actual capability is substantial: a way to rank a very large number of possibilities before spending bench time on them. Biology has always been limited by how many things you can try. A tool that reorders the queue — that moves the ten most promising candidates out of a list of ten thousand — changes what is affordable to investigate, even if it never tells you a single thing for certain.

That framing also sets the right expectation for anyone running these tools on their own data. The output is a prioritised list. What makes it valuable is the quality of the validation you put behind it.