Google DeepMind announced AlphaGenome 2 on Wednesday, November 11, a DNA model that reads genetic variants the way they actually work in the body: one cell type at a time. Instead of averaging a variant’s effect across a whole tissue, the new model predicts how a change in DNA alters gene activity inside specific human cells. DeepMind trained it on more than 107 million individual cell profiles across 68 tissues, and it is available now for non-commercial research through the AlphaGenome API.

On one headline test, the improvement is close to double. AlphaGenome 2 found 30% of the variants scientists believe cause changes in gene expression on TenK10K, a single-cell map of 28 primary immune cell types. Its predecessor, fine-tuned on the same single-cell data, found 16%. It also called the direction of a gene’s effect correctly 43% of the time, against 23% before. Those numbers come from DeepMind’s own preprint, which has not been peer-reviewed yet, so they carry the usual caveat: the company is grading its own homework for now.

Then came the demo that made scientists sit up. In 2024, researchers at the Francis Crick Institute in London spent a year and a half of lab work tracing a driver of inflammatory bowel disease to a variant near the ETS2 gene, hidden in a stretch of DNA that codes for no protein at all. AlphaGenome 2 reproduced the finding from data alone, singling out the same variant and showing that it boosts ETS2 only in gut macrophages, leaving other cell types alone.

Thirty minutes. That is how long the model took.

“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 and geneticist at the Francis Crick Institute who led the 2024 study, told Nature. “It’s reassuring from our perspective.”

How DeepMind AlphaGenome 2 works

Every cell in the body carries the same DNA, yet a heart muscle cell and an immune cell fighting an infection switch very different genes on and off. AlphaGenome 2 takes a DNA sequence of up to a million letters plus one biological condition — a tissue, a cell type, or a cell state — and predicts how that stretch is read in that cell: which genes turn on, how the DNA is packaged, and how a single-letter change shifts the outcome.

That one-condition-at-a-time design is the core change from the first AlphaGenome, which DeepMind published in Nature in January 2026. The original model predicted effects across thousands of tissue samples in parallel, spreading its capacity thin and blurring the differences between closely related cells. AlphaGenome 2 was trained on 637 types of human cell and 311 kinds of mouse cell, and researchers can also feed it their own single-cell data — from a disease state, for example — and get predictions for cells the model has never seen.

It does not just repeat what it was trained on. Given single-cell data from a COVID-19 mRNA vaccine study it had not seen before, the model distinguished vaccine-activated immune T cells from resting ones, and flagged a variant linked to severe COVID-19 that disrupts gene activity far more strongly when cells are activated — an effect invisible in unactivated cells. DeepMind also added DNA methylation, a widely requested readout, to the gene expression, histone modification, and chromatin accessibility predictions the first model offered.

Why cell type was the missing piece

About 98% of the human genome does not code for proteins, and much of the DNA linked to disease sits in those non-coding stretches. Many of those variants sit in enhancers, distant regulatory switches that tend to be active only in specific cells. A variant that drives disease in gut immune cells may do nothing at all in skin or brain cells, so a model that averages across a tissue can dilute the signal until it disappears.

“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, the DeepMind researcher who leads the AlphaGenome team, told Nature.

That gap is why the first model, useful as it was, had limits. In September, DeepMind released the AlphaGenome Atlas, predictions for all 9 billion possible single-letter changes in the human genome, built with the first model. Impressive in scale, but the predictions still did not say which cell a variant would act in. “Having something that cannot predict cell-type specificity is a bit of a problem, which is why it hasn’t been hugely useful for us,” Lee told Nature. AlphaGenome 2 is built to close exactly that gap.

What AlphaGenome 2 can’t do yet

For all the praise, the limits are plain. Even at 90% precision on its headline test, the model recovered only 30% of the variants scientists believe are causal — meaning it still misses most of what it is looking for. Performance outside immune cells is less established, since the headline benchmark was an immune-cell dataset. And the technical paper is a preprint, so no independent group has replicated the results yet.

Access is narrow, too. Non-commercial researchers get the API for free, but commercial research is invite-only, and the model itself cannot be downloaded and run locally. DeepMind is blunt about the other boundary: AlphaGenome 2 is for research only and has not been validated or approved for clinical or diagnostic use. Nobody should expect a doctor to use it on a patient anytime soon.

The bigger picture

Drug discovery is the obvious first stop. Genome-wide association studies flag broad stretches of DNA linked to disease, but finding the exact causal variant, the gene it controls, and the cell where it acts has been slow, expensive work. DeepMind precomputed AlphaGenome 2 predictions for trait-linked variants across hundreds of human cell and tissue conditions, covering traits from height to type 2 diabetes, so researchers can start from a shortlist instead of a haystack.

Industry is already moving. Genomics Limited says it plans to build AlphaGenome 2 into Mystra, its platform for judging drug targets from human genetics. DeepMind is also working with the Wellcome Sanger Institute to profile more cell types and disease states, and plans to release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, pairing the DNA model with Gemini 4 for automated science workflows.

None of that makes AlphaGenome 2 a shortcut to cures. What it does is shrink the distance between a genetic clue and the cell where it matters, from eighteen months of bench work to an afternoon at a keyboard. For researchers staring at a list of thousands of candidate variants, that is the difference between a lead worth chasing and a lead that stays on the shelf.

FAQ

What is DeepMind AlphaGenome 2?

It is a Google DeepMind AI model, announced November 11, 2026, that predicts how DNA variants change gene activity in specific human cell types and states, rather than in broad tissue averages.

How accurate is AlphaGenome 2?

On DeepMind’s TenK10K immune-cell test, it identified 30% of causal variants versus 16% for the first AlphaGenome, and called gene-effect direction correctly 43% of the time versus 23%. The figures are DeepMind’s own, from a preprint that has not yet been peer-reviewed.

Can AlphaGenome 2 diagnose disease?

No. DeepMind says both AlphaGenome models are for research use only and have not been validated or approved for clinical or diagnostic purposes.

Who can use AlphaGenome 2?

Non-commercial researchers can use it now, for free, through the AlphaGenome API. Commercial research access is in private preview, and the model itself cannot be downloaded and run locally.

Sources: Google DeepMind, Nature, Unite.AI