Confirmed 11 Nov 2026 · Science A one-minute read

Google released a DNA model that reads genes cell by cell, free for research.

On November 11 Google DeepMind published AlphaGenome 2, trained on over 107 million single cells across 68 tissues. Caveat: it is for research only and is not validated for clinical or diagnostic use.

Until now

AI reads the genome in bulk, averaged across a whole tissue.

Now

It can now answer which cell type a variant acts in.

Why it matters

Most disease-linked DNA changes sit outside genes, and the hard question is which cell they act in. On immune cell data the model found the causal variant in 30% of cases against 16% for the previous version. Google gave it to academics free and put commercial access behind an unpriced private preview.

Established

On TenK10K immune data it found causal variants in 30% of cases versus 16%.
It found directional gene effects in 43% of cases versus 23%.
Non-commercial research access is open now through the AlphaGenome interface.

Not established

Any clinical result. Google says it is not validated for diagnosis.
A new drug or treatment. Both disease examples reproduce published findings.
Commercial terms. Business access sits in an unpriced private preview.

The five-minute readThe record, the timeline, the weight and 3 sources

The record

Google DeepMind published "AlphaGenome 2: Predicting how DNA shapes cells and disease" at deepmind.google on November 11, 2026, which we read in full. AlphaGenome 2 is a conditional DNA sequence model: it takes a 1 million letter DNA sequence plus a target condition, a tissue, cell type or cell state, and predicts effects for that condition. Outputs include gene expression, histone modifications, chromatin accessibility and DNA methylation. It can predict a missing measurement, such as chromatin accessibility, from a single measured readout. Training used bulk and single-cell data from human and mouse samples: the post gives over 100 million individual cell profiles in one place and over 107 million in another, across 68 tissues. Results: on the TenK10K dataset of 28 primary immune cell types it identified causal variants in 30% of cases against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools, and found directional gene effects in 43% of cases against 23%. DeepMind says it identifies nearly twice as many variants that change gene expression in specific cells as the previous state of the art. Two disease examples are given, both reproducing published findings: a distant regulatory variant raising the ETS2 gene in gut macrophages in inflammatory bowel disease, and a severe COVID-19 linked variant disrupting gene activity more strongly in activated T cells. Access: free through the AlphaGenome application programming interface for non-commercial research, private preview for commercial research, precomputed predictions for trait-associated variants across hundreds of conditions, Colab tutorials, and a call for use cases by November 30. DeepMind says it is for research use only and has not been validated for clinical or diagnostic use, and names the Wellcome Sanger Institute as a collaborator. Sarah Teichmann of the University of Cambridge and Peter Donnelly, chief executive of Genomics PLC, are quoted; Donnelly says his company plans to incorporate it into its Mystra platform. Not published: the price of commercial access, the compute used, the model size, or any laboratory validation of a prediction the model made first.

How it got here

Jun 2025
Google DeepMind releases the first AlphaGenome, a model predicting how changes in a DNA sequence affect gene regulation, free to non-commercial researchers through an application programming interface.
Sep 2026
DeepMind publishes AlphaGenome Atlas, a precomputed map of molecular predictions for every possible single-letter change in the human genome.
Nov 6
DeepMind publishes a machine-checked result cutting the cost of factoring 1024-bit keys. Edition 080.
Nov 10
DeepMind publishes Gemini Robotics 2.5 with a 96% kitting result from active Mercedes-Benz lines in Hungary and Germany. Edition 086.
Nov 11
AlphaGenome 2 is published: conditional on cell type, trained on over 107 million cell profiles across 68 tissues, free for non-commercial research and in private preview for commercial use.

Who has to react

Genomics PLC, which says it will put the model inside its own platform; the Wellcome Sanger Institute, named as a collaborator; drug discovery teams that pay for variant interpretation today; medicines regulators, who have no approved route for a model like this; and the academic groups whose single-cell tools it just outscored.

What to watch

Would strengthen itAn independent group reproduces the 30% causal-variant result on data Google did not train on, or a peer-reviewed paper reports a variant found first by AlphaGenome 2 and later confirmed in the laboratory, by December 31, 2027.
Would weaken itA published benchmark finds AlphaGenome 2 no better than existing single-cell tools outside its training distribution, by December 31, 2027.
Would strengthen itGoogle prices commercial access to AlphaGenome 2, or names a paying pharmaceutical customer, by September 30, 2027.

Why it weighs this much

Signal strength

CostMedium

CommitmentMedium

ReversibilityModerate

PersistenceYears

Cost is medium: training on over 107 million cell profiles is expensive, and giving the result away free to academics forgoes revenue, but this is a research release rather than a capital commitment. Commitment is medium because published numbers invite replication and a named collaborator and a named customer have attached themselves to it. Reversibility is moderate: an application programming interface can be closed or priced, as the private preview for commercial users already signals. Persistence is years, because precomputed predictions and published benchmarks stay useful, and a free tool academics build on is hard to withdraw. We do not score it higher: both disease examples reproduce findings already in the literature.

The words

Variant: A difference between one person's DNA and the reference sequence. Most sit outside genes and act by changing how nearby genes are switched on.
Gene expression: How strongly a gene is switched on in a cell. The same DNA gives different expression in a nerve cell and a skin cell.
Single-cell data: Measurements taken from individual cells rather than averaged across a whole tissue sample.
Causal variant: The one DNA change actually responsible for an effect, among many nearby changes that are only correlated with it.

Sources

Your call

Which explanation do you believe?

A call is final. Other people's calls show after yours.