AlphaGenome 2 stops averaging: DeepMind's DNA model predicts one cell type at a time

Google DeepMind announced AlphaGenome 2 on November 11, a successor to the DNA sequence model it released last year. The input is unchanged: a million-letter stretch of DNA. What follows is not. Where the first model predicted activity across thousands of experiments from many tissues in parallel, the new one is conditional — it takes a single biological condition, defined as a tissue, cell type or state, and predicts for that condition alone. DeepMind trained it on more than 107 million individual cell profiles across 68 tissues, from human and mouse samples, bulk and single-cell. The readouts now include DNA methylation, a measurement researchers had been asking for.

The claim attached to the release is about prediction quality, not medicine. DeepMind says AlphaGenome 2 "identifies nearly twice as many variants that change gene expression in cells, compared to our previous state-of-the-art model." That comparison deserves taking apart. First, the architecture.

Why the parallel design was the wrong shape

The first AlphaGenome was efficient the way a wide lens is: it saw a lot at once and lost detail. It produced fixed parallel outputs for bulk tissues and cell lines from that million-letter sequence, an approach DeepMind now describes as having "spread the model's capacity thin" and made it hard to capture subtle differences between closely related cell types.

Bulk tissue is an average over many cell types, and averages hide exactly the signal disease genetics needs. A variant that acts in gut macrophages and nowhere else averages toward nothing. The Verge's Robert Hart put the pitch as a more detailed view of a problem that has resisted that view for two decades.

AlphaGenome 2 feeds the target condition into the model at every stage of its architecture rather than as a label bolted on at the end. That is what lets one set of weights learn how the same stretch of DNA is read differently across cell types. Conceptually it is a smaller change than adding a new modality. Practically it is larger, because the same model now serves hundreds of conditions instead of emitting all of them at once.

Nearly twice as many variants, on whose benchmark

The published numbers are specific. On TenK10K, a single-cell map of 28 primary immune cell types, the model identified 30 percent of causal variants against 16 percent for the original AlphaGenome fine-tuned on single-cell data, and 43 percent of directional gene effects against 23 percent. The first pair is the "nearly twice" in the headline. Both pairs sit inside DeepMind's technical report, which Nature's news desk describes as a preprint.

Here is the honest limit. These are DeepMind's benchmarks, run by DeepMind, against DeepMind's own prior model and a few specialised tools, and the document has not been through journal review. Nobody outside the company has replicated the result on cell types the authors did not choose. That does not make the numbers wrong. It makes them unconfirmed, which is a different status and should be reported as one.

Two capabilities that are not about scale

The conditional design buys two things DeepMind treats as features rather than side effects.

Cross-modality imputation comes first. Because the model separates a cell's biological state from its measured outputs, feeding it one measurement — single-cell gene activity, say — lets it predict properties nobody measured, such as chromatin accessibility. The use is triage: deciding which experiment is worth the money. It is also a claim that has to be checked per modality, since the accuracy of an unmeasured prediction is precisely what you cannot verify without doing the measurement.

The second is a cell encoder that generates a condition on the fly. A scientist uploads a single-cell RNA-seq snapshot from a few dozen cells and the model conditions on a state it was never trained for. DeepMind reports it can match or outperform models fine-tuned directly on those datasets. Nature puts the training scope at 637 human and 311 mouse cell types, so the encoder is the mechanism for everything past that boundary. In one example, given unseen data from a COVID-19 mRNA vaccine study, it separated vaccine-activated T cells from resting ones and found that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state — an effect the resting cells would have hidden. Sarah Teichmann of the University of Cambridge, a co-founder of the Human Cell Atlas, is quoted on the blog praising how the model adapts to new single-cell data "without retraining."

A replication that lands harder than the headline

The most persuasive item is not a percentage. A 2024 study in Nature traced inflammatory bowel disease to a variant in a distant enhancer that overactivates ETS2 in gut macrophages. AlphaGenome 2, run across immune cell types in the gut, singled out the same variant and showed the effect was selective for macrophages rather than spread across the tissue.

James Lee, the geneticist at the Francis Crick Institute who led that original work, told Nature that the model reached the answer in 30 minutes, against a year and a half of experiments in his lab. He also said the older model's inability to specify cell type was "a bit of a problem," which is the same criticism DeepMind makes of its own first version, arriving from the outside.

That is a reproduction of a known result, not a discovery. But it is a reproduction by a model built without that result in mind, in the cell type where the biology actually happens. Checks of that shape carry more weight than a benchmark delta, because the answer was fixed before the model ran.

What would settle it

Two things would move this from promising to established. The first is a benchmark run by someone with no stake, on cell types and states the model has not seen. TenK10K is a strong result from a dataset DeepMind picked and scored; the field needs the version of it that DeepMind cannot control. The second is peer review. Until the technical report clears a journal, the methodology has been read by editors at DeepMind, not by referees.

The release terms deserve a plain description too. Non-commercial research gets API access now. Commercial research sits in private preview, with no date and no published pricing. That is a normal pattern for a model trained on data with murky licensing, and it means the near-term users are academic labs rather than drug programmes.

None of this establishes clinical utility, a validated target or a treatment. DeepMind's own claim is narrow: better predictions of which variants do something in which cells, measured against its previous model. If that survives contact with independent groups, the precomputed enhancer-gene maps and trait-variant predictions are where the value will show up. DeepMind's Pushmeet Kohli has spent the past week arguing that AI for science is entering a phase of its own; this release is a better argument than the post that promised it, and it is still an argument about a model, not a result about a disease.