DeepMind’s AlphaGenome 2 predicts mutation effects in single cell types

The latest:

Google DeepMind has released AlphaGenome 2, an AI model built to predict how genetic variants behave inside specific cell types and states rather than across the genome as a whole. The company said in a blog post published Wednesday that the system offers a far finer-grained view than its predecessor, and that this resolution could expose disease-relevant effects in particular cells that broader models overlook.

Details:

• The model: AlphaGenome 2 succeeds AlphaGenome, the genome-predicting model DeepMind released last year, according to The Verge. The new version is designed to predict the effects of genetic variants on individual cell types and states, a level of resolution the earlier model did not target.
• The training data: DeepMind said the model was trained on more than 100 million individual cell profiles drawn from 68 tissues. That scale is what gives the system its more detailed view of how a single variant plays out differently from one cell population to another.
• The biological problem: Nearly every cell in the human body carries identical DNA, but heart muscle cells, skin cells and immune cells fighting an infection switch different genes on and off. Those patterns also shift depending on what a cell is doing at a given moment.
• Why that matters: Because of those switching patterns, the same genetic variant can produce very different effects depending on the type and state of the cell. The Verge described the problem as notoriously difficult, and one holding important clues to the causes of disease.
• DeepMind’s framing: The company said the model effectively turns static cell atlases — the large repositories that map and catalog individual cells in detail — “into a predictive engine for human genetics.” The claim positions the system as a layer built on existing biological reference data rather than a replacement for it.
• The medical pitch: Google presented the model as a route toward identifying disease causes and pointing to new treatments. The company did not announce a specific therapeutic program, clinical partner or timeline tied to the release.
• Access and rollout: The announcement came through a Google blog post on Wednesday. Google did not set out pricing, availability terms or which research institutions would be able to use AlphaGenome 2 first.
• The wider push: The release follows Google’s earlier atlas of the human genome, which the company also framed as potentially paving the way for new treatments. Rival AI developers are moving into the same territory, with Anthropic signalling an intent to develop its own drugs.

Background:

AlphaGenome, launched last year, predicted how DNA sequence changes affect genome function at a broad level. Cell atlases, meanwhile, catalog individual cells across human tissues in detail but remain static reference maps rather than predictive tools.

Between the lines:

The pitch rests on resolution rather than scale alone: a variant that looks harmless when averaged across a genome may register strongly in one immune or muscle cell population. Stated plainly, DeepMind is selling prediction layered on top of existing atlases. The absence of any named clinical partner or timeline leaves the treatment claim as a direction of travel, not a result.

What’s next

Watch for peer-reviewed validation of AlphaGenome 2’s variant predictions, the terms on which researchers get access, and whether Google names a disease target or clinical collaborator tied to the model.