ModelsNEWS BRIEFING

DeepMind releases AlphaGenome 2 to predict how cells read DNA variants

4 min read

88Z AI, 12 November — Google DeepMind announced AlphaGenome 2 on 11 November 2026. The model conditions predictions on a specific cell type or state rather than tissue averages. DeepMind says it identifies nearly twice as many variants that change gene expression as the previous model, and it is available to non-commercial researchers through an API.

At a glance

  • DeepMind announced AlphaGenome 2 on 11 November 2026, predicting DNA variant effects for a specific cell type or state.
  • The new model is conditional and was trained on more than 107 million cell profiles across 68 tissues, with DNA methylation added as a readout.
  • DeepMind says it identifies nearly twice as many expression-changing variants as the prior model, and reports 30% versus 16% causal-variant recall on TenK10K.
  • It is available now for non-commercial research through the AlphaGenome API, with Colab tutorials; commercial use is in private preview. Antigravity Skills are only previewed as coming soon.
Illustration of the DeepMind AlphaGenome 2 announcement: reading DNA variants by cell type
Illustration of the DeepMind AlphaGenome 2 announcement: reading DNA variants by cell type

From tissue averages to specific cells

On 11 November 2026, Google DeepMind published a blog post introducing AlphaGenome 2, its next-generation DNA sequence model. Officially, the model is meant to bridge single-cell biology and human genetics by predicting how genetic variants affect hundreds of human cell types. Last year's AlphaGenome predicted how variants affect molecular activity across broader tissues, and AlphaGenome Atlas offered predictions for about 9 billion possible single-letter changes in the human genome.

DeepMind's stated limitation is that tissue averages can hide variants that act only in particular cells. AlphaGenome 2 is designed to move from those averages to individual cell types and cell states, including disease-relevant cells that a broad-tissue model could miss, and to make predictions for cell and disease states the model has not seen directly.

A conditional architecture and finer training data

The post contrasts the two designs. The first AlphaGenome took a million-letter DNA sequence and predicted properties across many tissues and cell lines in parallel. That was efficient, but official authors say it spread the model's capacity thin and made subtle differences between closely related cell types harder to capture. AlphaGenome 2 is a conditional model: it predicts DNA effects for one condition at a time, defined as a tissue, cell type, or state, and uses that condition as input at every stage alongside the sequence. The condition can come from a reference catalog or be generated on the fly by a cell encoder from a scientist's own single-cell experiment.

DeepMind says AlphaGenome 2 was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. It predicts the same core readouts as the first model, such as gene expression, histone modifications, and chromatin accessibility, and adds DNA methylation. The company also describes cross-modality imputation: if a scientist provides one measurement, such as single-cell gene activity, the model can predict unmeasured properties such as chromatin accessibility, which may help prioritize the next experiment.

Official performance figures and worked examples

DeepMind says AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as the previous state-of-the-art model. On genome regions outside its training data, the company says it explains more of how gene expression differs across cell types within a tissue, including continuous changes as human blood stem cells mature into red blood cells and immune lineages.

On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, DeepMind reports that AlphaGenome 2 reliably identified 30% of causal variants versus 16% for the original AlphaGenome fine-tuned on single-cell data and other specialized tools, and 43% versus 23% for directional gene effects. Those figures come from DeepMind's own evaluation and should be read as official reported results, not independently audited conclusions.

The post also describes a computational reproduction of a published inflammatory bowel disease finding, in which a variant in a distant regulatory switch overactivates the ETS2 gene in gut macrophages. DeepMind says AlphaGenome 2 singled out the causal variant and showed that it selectively boosts ETS2 in those cells, leaving other cell types and genes largely unaffected. On unseen single-cell data from a COVID-19 mRNA vaccine study, the company says the model distinguished vaccine-activated immune T cells from resting ones and indicated that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state.

How researchers can use it now, and what remains a preview

DeepMind says AlphaGenome 2 is available today for non-commercial research via the AlphaGenome API, with interactive Colab tutorials for expression conditioning, and in private preview for commercial research. The post also says AlphaGenome 2 Skills will be released in Google Antigravity, its agentic development platform, but it only says soon, so that should not be treated as a feature that is already live.

For readers who follow drug discovery, the post quotes the CEO of Genomics Limited saying DNA sequence prediction models can help move a genome-wide association signal closer to the causal variant, target gene, and cell type a drug programme needs. This remains a research tool, not medical advice. Later updates should be taken from the DeepMind blog, the paper, and the API documentation, and predictions still need experimental follow-up.

Sources & editorial note

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.

  1. Google DeepMindSource published

Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.

All AI news

Related reading

All stories
88Z.ai