Skip to content
Google DeepMind's AlphaGenome 2, Announced on 11 November 2026, Predicts How DNA Variants Act in Specific Cell Types and, by DeepMind's Own Tests, Finds Nearly Twice as Many Causal Variants as Its Predecessor
AI & Technology

Google DeepMind's AlphaGenome 2, Announced on 11 November 2026, Predicts How DNA Variants Act in Specific Cell Types and, by DeepMind's Own Tests, Finds Nearly Twice as Many Causal Variants as Its Predecessor

Illustration by tuput

English

The new model takes a million letters of DNA plus a chosen tissue, cell type or cell state, and predicts gene activity for that one condition. On an immune-cell test set DeepMind reports 30 percent of causal variants found against 16 percent for its earlier model. The figures are the maker's own, from a preprint that has not been peer reviewed.

· · 9 min read

Google DeepMind announced AlphaGenome 2 on 11 November 2026, an AI model that predicts how a change in DNA affects gene activity in one chosen cell type, such as a gut macrophage or a resting T cell, instead of averaging across a whole tissue. On a test built from immune-cell data, DeepMind reports that the model picked out 30 percent of causal variants against 16 percent for its earlier model, which the company describes as nearly twice as many. The paper behind it is a preprint dated 17 November 2026, meaning it was posted before peer review, and every performance figure below is DeepMind’s own. As of 19 November 2026 none of the pages consulted reports an independent replication.

The model is offered to researchers, not clinicians. DeepMind says it is for research use only and has not been validated or approved for clinical or diagnostic purposes.

Same DNA, different reading

Every cell in the body carries the same DNA, and each cell type switches a different set of genes on or off. DeepMind’s announcement describes the aim as predicting how small DNA differences change that switching in specific cells, which it calls a key driver of human traits and disease. The Human Cell Atlas, a global consortium founded in 2016, is mapping gene activity in cells across the body, and its site lists more than 3,900 members from over 1,700 institutes in more than 100 countries.

Most of the DNA does not code for proteins. DeepMind’s June 2025 post on the first AlphaGenome says coding regions make up 2 percent of the genome, and the other 98 percent, the non-coding part, directs gene activity and holds many variants linked to disease. A variant effect predictor is a model that reads a DNA sequence with and without one letter changed and reports what the change does. Researchers use the scores to decide which variants to test first in the laboratory.

Cell type matters because the same variant can act in one cell type and be silent in the next. The TenK10K study in Nature, whose data DeepMind used as a benchmark, reports that about half of the genetic effects on gene expression it found were seen in only one or a few cell types.

What the first AlphaGenome did, and what its authors left open

The first AlphaGenome, published in Nature on 28 January 2026 (volume 649, pages 1206 to 1218), reads 1 million DNA letters at a time and predicts thousands of readouts, from gene expression to chromatin accessibility (how open the DNA is to the cell’s machinery). The paper says it covers 5,930 human and 1,128 mouse tracks across 11 kinds of measurement. Its abstract says it matched or exceeded the strongest available external models in 25 of 26 variant-effect evaluations. DeepMind’s 2025 post, written when the work was still a preprint, said 24 of 26, so the Nature count is the one that applies. The model was later paired with AlphaGenome Atlas, a 1-petabyte catalogue that, according to DeepMind’s product page, scores all nine billion possible single-letter changes in the human genome.

The Nature paper also lists what the model could not yet do. Its authors wrote that accurately recapitulating tissue-specific patterns across cellular contexts, and predicting condition-specific variant effects, remain challenging. Among the future directions they named were integrating single-cell data and adding DNA methylation, a chemical mark on DNA that affects how genes are read. AlphaGenome 2 takes up both.

One condition at a time

DeepMind says the first model produced fixed, parallel outputs for many tissues and cell lines from one run, which spread its capacity thin and made it hard to separate closely related cell types. AlphaGenome 2 is a conditional model. It takes the 1-million-letter sequence together with a condition, defined as a tissue, cell type or cell state, and feeds that condition into every stage of the network. The same stretch of DNA can then be read differently for each condition.

A scientist can choose the condition in two ways. One is a reference catalogue; the tutorial in the AlphaGenome documentation prints 2,799 supported human conditions. The other is to upload single-cell RNA data from the scientist’s own experiment, which a cell encoder compresses into a representation of the cell state. The blog says a few dozen cells are enough, and the documentation sets a limit of 256 cells per call. Because the model separates a cell’s state from its outputs, a single measurement such as gene activity lets it predict unmeasured ones such as chromatin accessibility, and DeepMind says that can help researchers decide which experiments to run next. DNA methylation, which the blog calls a widely requested readout, is new in this version.

For training, DeepMind says it used bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles across 68 tissues. The blog’s summary line rounds this to “more than 100 million”.

What the blog and the preprint report

All results below are the authors’ own. On TenK10K, a single-cell map of 28 immune cell types, the blog says AlphaGenome 2 identified 30 percent of causal variants against 16 percent, and 43 percent of directional gene effects (whether a variant turns a gene up or down) against 23 percent, compared with the original model fine-tuned on single-cell data and other specialised tools.

The preprint states a different measure. Its abstract says AlphaGenome 2 outperforms fine-tuned AlphaGenome 1 and specialised models trained on the target datasets, “doubling recall of causal single-cell eQTLs at 90% precision”. An eQTL is a variant tied to a change in how much a gene is expressed. Recall is the share of true causal effects the model finds, and precision is the share of its calls that are right, so the claim is about finding twice as many true effects while keeping nine in ten calls correct. The blog’s “nearly twice as many” and the preprint’s “doubling” are separate tests and should not be read as one number.

The abstract adds that the model achieves state-of-the-art enhancer-gene linking, meaning it matches distant stretches of regulatory DNA (enhancers) to the genes they control, and that the team generated an atlas of enhancer-gene links and scores for 7 million common variants across hundreds of human cell types. The blog says the precomputed scores cover trait-linked variants for traits from height to type 2 diabetes, and that when the model is confident it also predicts whether a variant raises or lowers a gene’s activity.

On unseen cell states, DeepMind describes a test with data from a COVID-19 mRNA vaccine study. Given single-cell data it had not seen, the model separated vaccine-activated T cells from resting ones, and predicted that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. The blog gives no numeric score for this example.

A test with a known answer: ETS2

One example has an answer from independent laboratory work. In 2024, Stankey and colleagues reported in Nature that a gene desert on chromosome 21 (a long stretch with no genes of its own) carries a haplotype, a block of variants inherited together, linked to inflammatory bowel disease and three other conditions, and that the causal gene, ETS2, is a central regulator of inflammatory macrophages. DeepMind says AlphaGenome 2, making predictions across immune cell types in the gut, singled out the causal variant and showed it boosting ETS2 in gut macrophages while leaving other cell types and genes largely unaffected.

The documentation’s vignette gives the detail. Statistical fine-mapping narrows the disease signal to seven non-coding variants spanning 1.2 kilobases. One of them, rs2836882, ranks second by that statistical method, yet the model predicts it strongly lowers ETS2 expression in macrophages from the ileum, the last part of the small intestine, and finds little effect from the top-ranked variant. According to the vignette, ETS2 sits about 289 kilobases away, while nearer genes show negligible change, and the effect is absent in B cells, T cells and natural killer cells. The vignette says this agrees with laboratory experiments on primary macrophages. It is a worked example written by DeepMind’s own team.

The data underneath

The immune-cell benchmark rests on public work. TenK10K phase 1, published in Nature on 28 October 2026 (volume 659, pages 417 to 428), combines single-cell RNA sequencing of 5.4 million cells from 1,925 people with whole-genome sequencing. It finds likely regulatory variants for 83 percent of the 21,404 genes tested. The paper says it found 17,674 genes with expression-linked variants against 16,987 in eQTLGen, the largest whole-blood study, with a sample 15 times smaller.

The participants are 1,925 individuals of European ancestry from two Australian cohorts, and the authors write that further growth in ancestry diversity, among other things, will be needed. The sources opened for this article name no Indian dataset used in AlphaGenome 2. India’s own reference collection is covered in tuput’s article on GenomeIndia, which sequenced 10,074 people from 83 population groups.

Access and what comes next

AlphaGenome 2 is available for non-commercial research through the AlphaGenome API, with interactive Colab tutorials for conditioning on single-cell data, and in private preview for commercial research through Google Cloud. DeepMind says AlphaGenome 2 Skills for its Antigravity workbench will follow, and it asked researchers to send use cases by 30 November, with selected projects eligible for extra quota and credits.

The blog carries two endorsements. Sarah Teichmann, co-founder of the Human Cell Atlas consortium, is quoted praising the model’s ability to adapt to new single-cell data without retraining. Peter Donnelly, chief executive of Genomics Limited, is quoted saying the company plans to bring it into its Mystra platform. DeepMind’s product page also reports use of the first model: a Broad Institute researcher used its variant scores to find a variant in the DNM1 gene that laboratory experiments later supported, and a University of Exeter team narrowed one region from 526 candidate variants to 4 in a study of more than 54,000 UK Biobank participants. Both accounts are DeepMind’s.

YourStory, an Indian technology outlet, repeated the 30 percent and 16 percent figures, attributing them to DeepMind. None of the pages consulted for this article reports a test of the model by anyone outside DeepMind, and the preprint had not been peer reviewed as of 19 November 2026.

Share
Copied!

Sources & further reading

  1. Google DeepMind (Adam Gayoso, Natasha Latysheva, Pushmeet Kohli, Ziga Avsec), 11 November 2026: AlphaGenome 2, deciphering how cells read DNA to better understand disease
  2. Gayoso A et al., AlphaGenome 2 predicts regulatory activity and variant effects across cell states, preprint first published 17 November 2026 (Europe PMC record PPR1371267, not peer reviewed)
  3. Google DeepMind, 25 June 2025 (updated January 2026): AlphaGenome, AI for better understanding the genome
  4. Google DeepMind: AlphaGenome product page (AlphaGenome 2, AlphaGenome Atlas and case studies)
  5. Avsec Z et al., Advancing regulatory variant effect prediction with AlphaGenome, Nature 649, 1206 to 1218 (published 28 January 2026)
  6. PubMed record and abstract, Avsec et al., AlphaGenome, Nature 649:1206 (PMID 41606153)
  7. Cuomo ASE et al., Rare and common genetic variants shape single-cell immune gene expression (TenK10K phase 1), Nature 659, 417 to 428 (published 28 October 2026)
  8. PubMed record and abstract, Stankey CT et al., A disease-associated gene desert directs macrophage inflammation through ETS2, Nature 630:447 (2024) (PMID 38839969)
  9. AlphaGenome documentation: Fine-mapping the ETS2 locus (AlphaGenome 2 vignette)
  10. AlphaGenome documentation: Choosing a condition (AlphaGenome 2 tutorial)
  11. Human Cell Atlas: What is the Human Cell Atlas (mission and consortium figures)
  12. YourStory: Google DeepMind unveils AlphaGenome 2 to study DNA changes

Researched and written with the help of AI tools and edited for accuracy. Provided for general information and discussion only, not professional advice. See our editorial standards and disclaimer. Spotted an error? Tell us.

#alphagenome#google deepmind#genomics#single-cell biology#variant effect prediction#gene regulation#non-coding dna#tenk10k

Enjoyed this? Get the next one.

One good read at a time, straight to your inbox. No spam, unsubscribe anytime.

More in AI & Technology
Six Chinese AI Firms Got Caught Copying America's Best Models. The US Fix: Quietly Make Their Answers Worse.
A US security advisory accuses six Chinese AI companies of mass copying American models and recommends quietly serving them worse answers. Plus DeepMind's atlas of every possible human DNA mutation, Apple's AI-first iPhone launch, and India's biggest rail expansion in years.
Google Gave Gemini 4 Carbon to Cyber Defenders on 28 October 2026 and to Paying Developers on 10 November, and Anthropic and OpenAI Also Release Their Strongest Models in Tiers
Three labs now let verified security teams use their most capable models before, or with fewer limits than, everyone else. Here is what each one published, the stated reason, and which numbers are the makers' own.
Anthropic Trained a 34-Million-Feature Autoencoder on Claude 3 Sonnet, Google DeepMind Then Deprioritised Fundamental Research on That Method, and Anthropic's Attribution Graphs Gave Insight on About a Quarter of Prompts
Researchers can now swap one concept for another inside a language model and watch the answer change. The same papers say, in their own words, how little of a model that covers and where the tools fail.
← all articles