IMPLICATOR
Google DeepMind released AlphaGenome 2 on Wednesday, a model that predicts how DNA variants act in specific cell types and states rather than averages across whole tissues. On DeepMind’s own TenK10K immune-cell benchmark, it recovered 30% of putatively causal variants at 90% precision, versus 16% for the original model fine-tuned on single-cell data.
“When you develop drugs, you’re targeting specific proteins and specific cell types,” said Žiga Avsec, DeepMind’s genomics lead.
What Changed
- Google DeepMind released AlphaGenome 2 on November 11, 2026, a model that predicts how DNA variants act in specific cell types and states rather than tissue averages.
- On DeepMind's own TenK10K immune-cell benchmark it recovered 30% of putatively causal variants at 90% precision, versus 16% for the original model fine-tuned on single-cell data.
- The model reproduced key aspects of a 2024 inflammatory bowel disease finding involving the ETS2 gene in gut macrophages.
- The benchmark comes from a DeepMind preprint that has not been peer reviewed; a Charité preprint found flaws in the original AlphaGenome's expression predictions.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Predictions for individual cells
The model was trained on more than 100 million individual cell profiles across 68 tissues, using human and mouse data. Nearly every cell carries the same DNA, but different cells switch different genes on and off, with patterns that also change in response to infection.
AlphaGenome 2 takes a million-letter DNA sequence alongside a chosen biological condition and predicts effects for that condition alone. Its predecessor produced fixed outputs across many tissues simultaneously. The new model also adds predictions of DNA methylation to its experimental readouts.
FREE WEEKDAY MORNING BRIEFING
Follow AI's push into biology and medicine.
The Implicator Morning Briefing filters the AI news cycle to the stories worth your attention and explains their consequences. From San Francisco, every weekday at 4:45 a.m. Pacific, 7:45 a.m. Eastern.
About five minutes. No hype. No spam.
Researchers can supply a snapshot of gene activity from a few dozen cells to define a previously unseen cell state without retraining the model. Given gene-activity measurements, it can also predict missing experimental data, such as whether stretches of DNA are accessible to regulation.
Reproducing an inflammatory disease finding
In one retrospective example DeepMind presented involving inflammatory bowel disease, AlphaGenome 2 reproduced key aspects of a mechanism established experimentally in 2024. It identified a DNA variant that selectively increases activity of the ETS2 gene in gut macrophages, a type of immune cell, while leaving other cell types and genes largely unaffected.
“They’ve managed to figure out the answer in 30 minutes, and it took us a year and a half of experiments,” said James Lee, an immunologist and geneticist at the Francis Crick Institute in London who led the earlier study. Lee said the original model’s lack of cell-type specificity meant it “hasn’t been hugely useful for us.”
DeepMind's numbers, and an outside test of version 1
The headline benchmark figures come from DeepMind’s own paper, a preprint on its website that has not been peer reviewed, and independent evaluations of version 2 have not yet been reported.
Know someone who'd find this useful? ✉️ Email it to a friend in one click, or they can subscribe free here.
A Charité-led preprint posted November 9 tested the original AlphaGenome, not version 2, on 2,111 promoter-gene pairs disrupted in computer simulations. The authors found that the model registered the local loss of promoter activity but often failed to carry that effect through to the rest of the gene’s transcript. The authors concluded that systematic biological benchmarking and experimental validation remain essential before such predictions are used to prioritize or classify noncoding variants.
Access and computing costs
AlphaGenome 2, intended for research only and not validated or approved for clinical or diagnostic use, is available through an API free for non-commercial research, with commercial research in private preview.
Avsec said its more detailed predictions cost more to compute than those of the original AlphaGenome, which remains in service. He said those demands would have made it impractical to use the new model for the genome-wide atlas DeepMind released in September, covering every possible single-letter change in human DNA.
DeepMind is working to generate more training data for poorly represented cell states. Avsec said future versions could learn from data “across multiple species” and, longer term, contribute to efforts to build virtual cells, digital simulations of biological systems.
Frequently Asked Questions
What is AlphaGenome 2?
It is Google DeepMind's DNA sequence model, released November 11, 2026, that predicts how genetic variants affect gene activity in specific human cell types and states. It was trained on more than 100 million individual cell profiles across 68 tissues, using human and mouse data.
How does it differ from the original AlphaGenome?
The original model produced fixed outputs across many tissues at once. AlphaGenome 2 takes a DNA sequence plus a chosen biological condition and predicts effects for that condition alone. It also adds DNA methylation predictions and can handle new cell states from a few dozen cells without retraining.
How well does it perform?
On DeepMind's own TenK10K immune-cell benchmark, it recovered 30% of putatively causal variants at 90% precision, versus 16% for the original model fine-tuned on single-cell data. The figures come from a DeepMind preprint that has not been peer reviewed.
Have outside scientists tested it?
Independent evaluations of version 2 have not yet been reported. A Charité-led preprint posted November 9 tested the original AlphaGenome and found it often failed to carry the effect of promoter disruption through the rest of a gene's transcript.
Who can use AlphaGenome 2?
It is available through an API free for non-commercial research, with commercial research in private preview. It is intended for research only and has not been validated or approved for clinical or diagnostic use.
AI-generated summary, reviewed by an editor. More on our AI guidelines.
Free AI briefing · Weekdays
The AI stories that matter, sourced and explained.
Join the Morning Briefing. It goes out every weekday at 4:45 a.m. Pacific, with later sends for the East Coast, Berlin and Tokyo.
Free when you sign up: The Paperclip Compendium, our tested guide to running AI agents.
Free. Unsubscribe in one click.
Marcus Schuler
Editor-in-Chief and founder of Implicator.ai. Former ARD correspondent and senior broadcast journalist with 10+ years covering tech. Writes daily briefings on policy and market developments. Based in San Francisco. E-mail: editor@implicator.ai
