DeepMind's AlphaGenome 2 finds nearly twice as many disease-linked DNA variants

Your DNA is a long string of letters. A variant is a spot where your letter differs from the usual one. Most variants do nothing. A few change how strongly a gene is switched on, and some of those raise the risk of disease.

AlphaGenome 2 takes a stretch of DNA one million letters long and predicts what a cell will do with it. It predicts gene expression (how active a gene is), chromatin accessibility (how open the DNA is for the cell to read) and histone modifications (chemical tags on the proteins that DNA wraps around).

It adds one readout the first model lacked: DNA methylation, a chemical tag that helps switch genes up or down. DeepMind calls it a widely requested feature.

The goal is simple. Give scientists a way to guess which variants matter before they spend months in a lab.

DeepMind launched the first AlphaGenome last year. It predicted properties across thousands of experiments from many tissues all at once, from a million-letter input.

That was efficient, but it had a cost. The company says spreading the model's capacity across so many outputs made it hard to catch small differences between closely related cell types.

DeepMind then built the AlphaGenome Atlas, a platform with predictions for all 9 billion possible single-letter changes in the human genome. Researchers use these tools to rank variants and make sense of rare diseases, complex traits and basic biology.

A tissue is a blend of many cell types. A prediction for "blood" is an average, and an average can hide a variant that matters in only one kind of cell.

DeepMind changed the design. AlphaGenome 2 is a conditional model. That means it predicts for one condition at a time, where a condition is a tissue, a cell type or a cell state.

The model gets the DNA and the target condition together, and it uses the condition at every stage. This lets it learn how the same stretch of DNA is read differently in different cells.

In real life Think of one recipe card used by a thousand kitchens. A bakery, a sushi bar and a pizzeria each read it differently. The old model guessed the average dish. The new one asks which kitchen you mean.

The condition can come from a catalog of known cell types. It can also be made on the fly from a scientist's own single-cell experiment, through what DeepMind calls a cell encoder.

This design also allows cross-modality imputation, which is a fancy name for filling in missing data. If a lab measured only gene activity, the model can predict other properties, such as how open the DNA is. That can help decide which experiment to run next.

The training set is large:

• Cells: more than 107 million individual cell profiles.
• Tissues: 68.
• Species: human and mouse samples.
• Data types: both bulk data (an average over many cells) and single-cell data (one cell at a time).

Projects such as the Human Cell Atlas are mapping gene activity in millions of cells. Predicting how small DNA changes alter that activity is still an unsolved problem, and DeepMind says this model turns those static maps into a predictive engine.

DeepMind tested the model on TenK10K, a large single-cell map of 28 primary immune cell types. The company compared it with the first AlphaGenome fine-tuned on single-cell data (fine-tuning means giving a finished model extra training on one kind of data) and with other specialized tools.

By the numbers:

"Direction" means whether a variant turns a gene up or down. That is a key clue for understanding how the gene feeds into disease.

DeepMind also says the model tracks continuous changes better, such as blood stem cells maturing into red blood cells and immune cells. On DNA regions outside its training data, it explains more of how gene activity differs across cell types in the same tissue.

AlphaGenome 2 found nearly twice as many variants that change gene expression in cells as DeepMind's previous state-of-the-art model.

The catch: these are DeepMind's own results on its chosen benchmark. Percentages of 30 and 43 also mean the model still misses most variants, so it narrows the search and does not end it.

DeepMind gave two examples.

The first is inflammatory bowel disease (IBD). A recent study showed that a variant in a distant regulatory switch drives IBD by overactivating a gene called ETS2 in gut immune cells called macrophages. AlphaGenome 2 reproduced this finding on its own. It singled out the causal variant and showed that it boosts ETS2 in gut macrophages, leaving other cell types and genes largely unaffected.

The second is COVID-19 vaccination. Given unseen single-cell data from an mRNA vaccine study, the model told apart vaccine-activated T cells from resting ones. It also showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. That effect would be missed in resting cells.

A co-founder of the Human Cell Atlas consortium, who is a professor of stem cell medicine at the University of Cambridge, said in DeepMind's post that what makes the model exciting is how it adapts to new single-cell data without retraining.

This is the part DeepMind stresses most. A scientist can upload a snapshot of gene activity from just a few dozen cells. The model squeezes that into a "cell state" summary and uses it as the condition.

DeepMind says the model can then match or beat models trained or fine-tuned directly on those datasets. That means a lab does not need to train a new model for every new experiment.

DeepMind precomputed predictions for variants tied to traits across hundreds of human cell and tissue conditions, so researchers do not need to run huge computations themselves. It names two problems.

• Enhancer-gene links: many disease-linked variants sit in enhancers, distant switches that turn genes on. Lab screens mostly use lab-grown cell lines, so the wiring in real human cells is often unknown. The model extends this mapping to primary human cell types, such as brain astrocytes, which support and protect neurons.
• Causal GWAS variants: genome-wide association studies point to broad stretches of DNA linked to a trait. The model helps pick the exact variant, the target gene and the cell type, for traits from height to type 2 diabetes.

AlphaGenome 2 and interactive Colab tutorials are available now for non-commercial research through the AlphaGenome API (a doorway that lets software talk to the model). Commercial researchers can join a private preview.

DeepMind says AlphaGenome 2 Skills are coming to Google Antigravity, its platform for building AI agents. It also names collaborations with researchers at the Wellcome Sanger Institute to profile more cell types, developmental lineages and perturbed cell states.

Teams can send use cases through a form by 30 November. Selected projects may get extra quota and credits.

DeepMind adds a clear limit: AlphaGenome and AlphaGenome 2 are for research only. They have not been validated or approved for clinical or diagnostic use.

• Not a test you can take: this is a research tool. It will not diagnose you, and no clinic can use it that way.
• Maybe faster science: if researchers can rank suspect variants better, they can pick better targets for studies and, later, for drugs.
• Cell-level detail: future findings may say not only which gene matters but in which cell it matters.
• Still early: the results are DeepMind's own on one benchmark, so wait for independent labs to test it.