AlphaGenome 2 refines genome reading cell by cell

Google DeepMind launches AlphaGenome 2, an AI model that predicts the effect of genetic variants on specific cell types. The tool is already accessible for non-commercial research but not for clinical diagnosis.

Google DeepMind has presented AlphaGenome 2, an AI model for genomics that moves from tissue-scale predictions to cell-by-cell reading. According to the company, this version better identifies the effect of genetic variants on specific cell types and can help isolate possible causes of disease. The service is available for non-commercial research via API, while commercial access exists in a closed preview. No white coat on paper yet, but the tool is starting to make an impact in the lab.

De la moyenne du tissu au détail qui compte vraiment

The change is simple to say, harder to do in biology: AlphaGenome 2 no longer tries to encompass all tissues at once. The model receives a DNA sequence and a precise biological state as input, then calculates what happens in that specific cell type. Google DeepMind explains it trained the system on over 107 million individual cell profiles from 68 tissues, in humans and mice.

This approach serves a very concrete goal: better understanding why a mutation acts in one cell but not another. The model predicts gene activity, how DNA is packaged in the nucleus, and DNA methylation, i.e., chemical markers that can activate or silence a gene. It can also infer missing properties from a single experimental measurement, which helps choose which bench experiments to run next.

Des variantes mieux triées, des effets plus lisibles

On the TenK10K benchmark, covering 28 primary immune cell types, AlphaGenome 2 identified about 30% of causal variants, compared to 16% for a refined version of AlphaGenome 1. In 43% of cases, it also correctly predicted whether a variant increased or decreased gene expression, versus 23% for its predecessor. The figures remain limited to this test, but the gap gives an idea of the resolution gain.

DeepMind also says it reproduced, purely computationally, a mechanism already described in a study on inflammatory bowel disease. The model isolated the exact variant located in a regulatory switch far from the gene, capable of overactivating ETS2 in intestinal macrophages, without affecting other cell types in the same way. For researchers, this kind of result matters because it brings a broad genetic signal closer to an exploitable biological target.

Ce que l’IA voit dans les cellules que les autres ratent

AlphaGenome 2 also proves more useful when the cellular state changes. On data related to the maturation of blood stem cells into red blood cells, its predictions matched observed measurements better than the previous model. In a study of an mRNA vaccine against COVID-19, it distinguished activated T lymphocytes from resting cells and showed that a variant associated with severe forms mainly disrupted cellular activity in the activated state. In resting cells, the signal would have gone unnoticed.

The model was also applied to pan-genomic association studies, or GWAS, linking large DNA regions to traits like height or type 2 diabetes. AlphaGenome 2 then helps narrow the area, spot the most plausible variants, and identify the relevant cells, such as certain astrocytes in the brain or specific immune cells. This is the kind of work that doesn’t make headlines but avoids weeks of blind sorting.

Un outil de recherche, pas un diagnostic

Google DeepMind specifies that the precomputed predictions already cover hundreds of cellular and tissue conditions, notably to support drug discovery. The company also announces API access for non-commercial research and a closed preview for commercial uses. However, AlphaGenome 2 is not approved for clinical diagnosis.

In the same move, DeepMind wants to integrate dedicated functions for AlphaGenome 2 into Google Antigravity, its agent platform, with Gemini 4 to automate certain research workflows. It remains to be seen how these tools will be adopted by labs, balancing speed promises and methodological caution. In genomics, AI does not replace experience; it can mainly help avoid doing ten times too many experiments.

Key points

  • AlphaGenome 2 targets a specific cell type, not the average tissue.
  • 107 million cell profiles were used for training.
  • On TenK10K, it finds about 30% of causal variants.
  • DeepMind provides non-commercial API access.
  • The tool is not approved for clinical diagnosis.
  • Google Antigravity is expected to integrate dedicated functions soon.

By the numbers

  • 107 million — individual cell profiles, human and mouse, according to Google DeepMind.
  • 68 tissues — training scope announced for AlphaGenome 2.
  • 28 primary immune cell types — TenK10K benchmark used for evaluation.
  • 30% — causal variants identified on TenK10K, according to DeepMind.
  • 16% — score of a refined AlphaGenome 1 version on the same test.

To read