AlphaGenome 2 stopped averaging your cells

In 2024, a team at the Francis Crick Institute in London did something expensive. They spent a year and a half of gene editing and molecular biology to explain why a single letter of DNA, sitting in a stretch of genome with no genes in it, raises the risk of inflammatory bowel disease. The answer was a distant switch that overactivates a gene called ETS2 in immune cells in the gut.

This week, researchers at Google DeepMind reproduced that result computationally, without a pipette.

"They've managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments," James Lee, the immunologist who led the 2024 study, told Nature. "It's reassuring from our perspective."

The model is AlphaGenome 2, announced on 11 November. Its improvement is narrow, technical and consequential: it predicts what a DNA variant does inside a specific cell type or cell state, instead of averaging across a tissue.

That distinction is the whole argument. Your heart, brain and immune system carry the same DNA sequence. What separates them is which genes each cell switches on, and how much. A variant can wreck a gene in one cell and do nothing in the cell next to it, which is why a tissue average can hide the mechanism a drug programme needs.

Four steps in 2024 became three in 2026: the variant, the mechanism, and a model that can reach them without the eighteen months in between.

What actually changed

AlphaGenome 1, previewed in June 2025 and published in Nature in January 2026, read a million letters of DNA and produced predictions for thousands of experiments across many tissues at once. DeepMind now says that design "spread the model's capacity thin" and made it hard to separate closely related cell types.

AlphaGenome 2 is a conditional model. Alongside the DNA sequence it takes a biological condition — a tissue, a cell type or a cell state — and uses that condition at every stage of its architecture, predicting for one condition at a time. It was trained on more than 107 million individual cell profiles, covering 637 human and 311 mouse cell types, and it adds DNA methylation, a readout researchers had been asking for.

Two consequences matter more than the benchmarks. First, it can predict missing measurements: give it single-cell gene activity and it will estimate chromatin accessibility, which tells a lab which experiment is worth running next. Second, it adapts to cell states it has never seen, from a few dozen cells and without retraining. Given data from a COVID-19 mRNA vaccine study, it separated 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.

"If you want to understand variants associated with genetic diseases, then you have to ask yourself what are the cell types that are affected," Žiga Avsec, who leads the AlphaGenome team, told Nature. Lee put the earlier version's limitation more bluntly: a model that cannot predict cell type specificity "hasn't been hugely useful for us".

The stack Google is building

Read the releases in order and a structure appears. Each one solves the layer above the last, and each one is useless on its own.

Isomorphic Labs, the Alphabet spinout that runs DeepMind's drug-design work commercially, raised $2.1 billion in May and said its first clinical trials were expected before the end of this year — a target it had previously set for 2025. AlphaFold told us the shape of the lock. AlphaGenome tells us which door it is on and in which room. Isomorphic cuts the key. None of that decides whether the key reaches the room.

Delivery is where biology stops being an information problem. A drug that binds the right protein in a dish still has to survive the bloodstream, cross the right membrane and enter the intended cell. Targeted delivery exists — lipid nanoparticles, antibody-drug conjugates, engineered T cells — but it is chosen per organ and per molecule, not computed from a sequence. Trials remain the only test that counts, and they are slow for reasons no model has changed.

What three to five years would require

The case for solving biology soon is not that the models are magic. It is that the cost of the first step — forming a correct hypothesis about mechanism — has collapsed faster than anyone predicted. Thirty minutes against eighteen months is not a productivity gain; it is a different kind of work.

Three things would have to hold. Variant-to-cell-to-drug has to keep closing, which this model makes plausible for the first time. Wet-lab validation has to keep up, which is a robotics and reagent problem, not an AI one. And layer five has to be solved once, in one tissue, for one class of molecule — after which the pattern tends to generalise.

AlphaGenome 2 does not cure anything. It is research-use only, not validated for clinical work, and its predictions are hypotheses a lab still has to test. Many diseases involve dozens of variants; some enhancers regulate genes from further away than the model can see. The 30-minute result reproduced something already known.

But the stack now has a shape, and the missing piece has a name. That is a different situation from five years of promising demos.

Source: Google DeepMind, "AlphaGenome 2: Deciphering how cells read DNA to better understand disease" (11 November 2026); Nature news, "DeepMind's ultra-powerful AI for genome mapping wows scientists" (11 November 2026); Isomorphic Labs funding announcement (12 May 2026).

Contact: [email protected]