AlphaGenome 2 moves genomics AI to single-cell resolution

By Marcus Bennett

AlphaGenome 2 identified about 30% of causal variants on the TenK10K benchmark, covering 28 primary immune cell types, according to Google DeepMind. In 43% of cases, the model also correctly predicted whether a variant upregulates or downregulates a gene, meaning whether it raises or lowers gene activity.

A fine-tuned version of AlphaGenome 1 reached 16% for causal-variant identification and 23% for direction-of-effect prediction in Google DeepMind's comparison. The company says AlphaGenome 2 finds nearly twice as many relevant gene changes as its predecessor. The gap amounts to 14 percentage points for causal-variant identification and 20 points for direction-of-effect prediction.

The improvement targets a basic problem in genomics: every human cell carries the same DNA, but cell types switch different genes on or off. Google DeepMind used the example of a heart muscle cell reading DNA differently from an immune cell. Predicting those cell-specific effects has remained one of the field's biggest unsolved problems, according to the company.

Google DeepMind released AlphaGenome 2 to predict how DNA sequences and their variations affect specific human cell types. The company says the model was trained on more than 107 million individual cell profiles from 68 tissues in humans and mice, with the aim of identifying cellular drivers of disease.

Google DeepMind's cell-state examples show why the distinction carries research stakes: a variant's effect can emerge only in an activated cell and remain invisible in a resting one. The company says AlphaGenome 2 can connect variant-level associations with the cellular environments in which they alter gene activity, helping researchers identify candidate cellular drivers of disease before choosing experiments.

What changes at single-cell resolution?

Google DeepMind said the original AlphaGenome accepted a DNA sequence one million letters long and predicted its properties across thousands of experiments spanning many tissues at once. The company said that parallel approach spread the model's capacity too thin and made subtle differences between closely related cell types harder to capture.

AlphaGenome 2 replaces that setup with a conditional model, meaning its output depends on an added biological-state input. Alongside the DNA sequence, the model takes the state of one specific cell type and calculates effects for that state, according to Google DeepMind.

According to Google DeepMind, the practical change is the unit of prediction. Instead of returning only a tissue-wide effect, AlphaGenome 2 calculates the effect for a specified cellular environment, so closely related cell types can be evaluated separately.

Google DeepMind says AlphaGenome 2 produces three molecular predictions:

• Gene expression: Google DeepMind says the model predicts how strongly a gene is switched on in the selected cell type.
• DNA packaging: Google DeepMind says it predicts how DNA is physically packaged in the nucleus and which stretches of the genome remain accessible.
• DNA methylation: Google DeepMind says the model predicts chemical tags that act as molecular switches to silence or activate genes.

AlphaGenome 2 can also infer missing data from a single measurement. If researchers measure only a cell's gene activity, for example, Google DeepMind says the model supplies predictions for additional properties, including how tightly or loosely stretches of DNA are packaged. The company says this inference helps researchers decide which laboratory experiments are worth running next.

How does the model handle disease mechanisms?

Google DeepMind says AlphaGenome 2 models continuous biological changes with considerably greater accuracy, including the maturation of blood stem cells into red blood cells. The company does not provide a numerical margin for that comparison.

Working purely computationally, AlphaGenome 2 reproduced an inflammatory bowel disease mechanism that a previous study had identified, Google DeepMind reported. The company said the system isolated the exact DNA variant in a regulatory switch far from the affected gene. Here, a regulatory switch is a DNA sequence that controls a gene's activity.

According to Google DeepMind, the variant selectively overactivates ETS2 in gut macrophages, while other cell types and genes remain largely unaffected. Google DeepMind's reproduced mechanism links a regulatory sequence, a specific immune-cell environment and a disease-associated gene without relying on a tissue-wide signal.

Can AlphaGenome 2 transfer to unseen cell states?

Google DeepMind says AlphaGenome 2 can adapt to entirely new experiments without retraining. Researchers can feed it gene activity data from just a few dozen cells, and the model calculates effects for a previously unseen cell state. According to Google DeepMind, it often beats specialized models trained directly on that data.

Applied to data from a COVID-19 mRNA vaccine study, Google DeepMind says the model reliably distinguished vaccine-activated T cells, a type of immune cell, from resting ones. The study data tied a gene variant linked to severe COVID-19 cases to disrupted activity mainly in the activated state.

Google DeepMind says the effect would have remained invisible in resting cells. The finding, as reported by Google DeepMind, gives activation state a concrete role in the experiment because an analysis limited to resting samples could miss the variant's function.

How will researchers use and access AlphaGenome 2?

Google DeepMind has already precomputed predictions for hundreds of cell and tissue conditions, partly to support AI-driven drug development. The same predictions are designed to connect human genetics with cell biology and help researchers analyze genome-wide association studies (GWAS).

These large-scale studies link DNA regions to traits such as height or type 2 diabetes but often locate signals only roughly within broad genomic stretches. AlphaGenome 2 narrows the likely causal variants and identifies affected cell types, according to Google DeepMind. The company cites astrocytes, a type of brain cell, and specific immune cells as examples.

When Google DeepMind is sufficiently confident, AlphaGenome 2 also predicts whether a mutation upregulates or downregulates a gene. The company calls that directional capability important for drug development because it specifies the sign of the predicted effect.

AlphaGenome 2 is available now for noncommercial research through an API (application programming interface), while commercial users get access through a closed preview. Google DeepMind says the system is not approved for clinical diagnostics.

Google DeepMind plans to roll out dedicated AlphaGenome 2 features soon on its agentic development platform, Google Antigravity. The company says the model will combine with tools such as Gemini 4 for automated research workflows.

Google DeepMind's next checkpoint is the planned Google Antigravity rollout with Gemini 4, which would extend the model from a research API into automated research workflows. The 30% causal-variant result gives the project a measurable baseline, but the lack of clinical diagnostic approval keeps its immediate role in research and drug-development support.