Google's AI can now read your DNA, cell by cell

On Wednesday, 11 November 2026, Google DeepMind announced AlphaGenome 2, an artificial intelligence model that predicts how changes in DNA affect gene regulation within specific human cell types, rather than across broad tissue averages. Where earlier tools could indicate only that a genetic variant might matter in, for example, the heart or the immune system, AlphaGenome 2 is designed to identify the precise cell type affected, and the state in which it is affected. A variant that is harmless in one cell type can be damaging in another, and the model is built to capture that distinction.

Background: regulation at single-cell resolution

Every cell in the human body carries the same DNA, but different cells read different parts of it. The genes switched on or off in a heart muscle cell differ substantially from those active in an immune cell. AlphaGenome 2 predicts those effects at the level of the individual cell type.

This distinction matters because most DNA variants linked to disease do not sit inside genes themselves. They sit in the regulatory regions around genes, the instructions that control when and where a gene is used. Determining which of those variants actually causes disease, and in which cells, has been one of the most difficult problems in genomics. Given a DNA sequence, AlphaGenome 2 predicts how a variant changes gene regulation in a specific cell type and state, providing researchers with a far more precise map of where a variant acts.

Empirical performance

The model was trained on more than 107 million individual cell profiles across 68 tissues, a dataset that provides a view of gene regulation at single-cell resolution across the human body.

On the TenK10K benchmark, which tests models across 28 primary immune cell types, AlphaGenome 2 identified 30 percent of causal variants, compared with 16 percent for the original AlphaGenome, an improvement of nearly double. It also determined the direction of gene effects more often: 43 percent of the time, versus 23 percent for its predecessor.

These are benchmark figures, not clinical results. The improvement is nevertheless substantial. Doubling the share of causal variants identified allows researchers to narrow the list of suspects in a disease study far faster, and knowing the direction of an effect, whether a variant turns a gene up or down, is precisely the information required to understand how a disease works.

Relation to the original AlphaGenome

AlphaGenome 2 is the successor to DeepMind's original AlphaGenome model, which predicted the functional effects of DNA variants at the level of broad tissue averages. That was already a useful tool for genomics research, but tissue averages blur the picture: a tissue is a mix of many different cell types, and a variant's effect can be strong in one of them and invisible in the rest.

The new model moves the resolution down one level, from tissue to cell type. It is part of a wider shift in genomics toward single-cell analysis, in which the objective is to understand biology cell by cell rather than in bulk.

Implications for research

For researchers studying genetic disease, the practical value is speed and precision. Genome-wide association studies routinely produce hundreds of candidate variants linked to a disease, but only a small number of them are actually causal. Sorting the real signals from the noise is slow, expensive work, and it usually requires laboratory experiments.

A model that can identify twice as many causal variants, and state which cell type they act in, allows researchers to focus their experiments on the most promising candidates first. Over time, that could accelerate the search for the mechanisms behind conditions in which genetics plays a role, from immune disorders to heart disease.

It also matters for drug discovery. Knowing which cell type a variant affects points researchers toward the biology to target, rather than a vague tissue-level association.

Access and limitations

AlphaGenome 2 is available free now for non-commercial research through the AlphaGenome API. Commercial research access is in private preview.

There is an important limit to note. The model is for research use only. It has not been validated or approved for clinical or diagnostic purposes, and it is not a medical tool. Its predictions are a starting point for research, not a diagnosis.

Outlook

The immediate test will be how widely researchers adopt the model and what it turns up in real studies. If the benchmark gains hold up on new datasets and new diseases, single-cell variant prediction could become a standard step in genomics research.

DeepMind has not said when, or whether, commercial access will open more broadly. For now, the model is a research tool, but one that gives scientists a sharper view of the genome than they had before.