DeepMind's new model predicts how the same DNA is read in different cells, flagging nearly twice as many disease-linked variants

The backstory: same DNA, different jobs

The human body has tens of trillions of cells, and almost all of them carry the same DNA, like identical copies of the same instruction manual. Yet heart cells contract, immune cells patrol, and liver cells process toxins. The difference is that each cell turns only some of those instructions on or off.

Scientists can already take a snapshot of which genes are active in a single cell, and projects such as the Human Cell Atlas are mapping that picture in fine detail. A harder question remains: if one letter of DNA changes, will it cause trouble in only one kind of cell? Many disease-linked variants don't sit inside a gene itself but in a distant "switch" region, and tracking where those switches act has been difficult.

Last year, DeepMind released the first AlphaGenome, which mainly predicted how DNA changes affect a whole tissue on average. A tissue is a mix of many cell types, so averages can drown out a signal that appears in only one of them.

What launched

On November 11, DeepMind announced AlphaGenome 2. Instead of predicting many tissues in parallel, it conditions on one setting at a time — a tissue, a cell type, or a cell's current state. Alongside a DNA sequence of about a million letters, the model is told at every stage which biological condition it should be reading for.

DeepMind says the model was trained on bulk and single-cell data from humans and mice, including more than 107 million individual cell profiles across 68 tissues. It predicts the same core readouts as the original model, such as gene activity and whether the DNA is in an "open" state that is easier to read, and it adds DNA methylation, a chemical mark often linked to genes being switched off, which researchers had asked for.

The headline numbers, from DeepMind's own evaluation on TenK10K, a large single-cell map of 28 primary immune cell types: AlphaGenome 2 identified about 30% of causal variants, versus about 16% for the previous model after it was fine-tuned on single-cell data. For the direction of the effect — whether a variant turns a gene up or down — the figures were about 43% versus 23%. In that test, the new model found roughly twice as many usable clues.

The company also walked through two concrete cases. A recent study found that a variant far from the gene itself overactivates ETS2 in gut macrophages and contributes to inflammatory bowel disease. DeepMind says AlphaGenome 2 reproduced that result in silico: it singled out the variant and showed the effect was largely limited to those gut immune cells. In a second test, given single-cell data from a COVID-19 mRNA vaccine study the model had not seen during training, it distinguished vaccine-activated T cells from resting ones, and found that a variant linked to severe COVID-19 disrupted gene activity more strongly in the activated state — a difference that would be missed in resting cells.

A practical feature for labs is that researchers can upload a snapshot from just a few dozen cells, without retraining, and ask the model to predict for that new state. Sarah Teichmann, a co-founder of the Human Cell Atlas and a professor at the University of Cambridge, said that ability to adapt to new single-cell data is what makes the model exciting.

What it can and cannot do

DeepMind also precomputed predictions for many variants linked to complex traits such as height and type 2 diabetes, across hundreds of human cell and tissue conditions, to help researchers narrow a broad "this stretch of DNA is associated with disease" signal down to more likely variants, genes, and cell types. Peter Donnelly, CEO of Genomics Limited, said sequence-prediction models of this kind help bridge the gap between a statistical association and the mechanistic detail a drug program needs.

The limits are explicit. DeepMind states that AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic use. They are not a test a doctor can order. The performance figures also come mainly from the company's own evaluations and case reconstructions; independent checks will take time.

What it means for you

In the near term, this will not show up on a hospital lab report. The more realistic near-term effect is in research and early drug discovery: scientists may get a better first guess about which cell type to test, instead of running every possibility.

If later independent studies back it up, tools like this could shorten some of the detours in finding causes and drug targets. For now it is a finer map, not a diagnostic device.

Original source published: Google DeepMind, Nov 11, 2026, 23:00 Beijing time (UTC+8)