# AlphaGenome 2, November 2026: DeepMind AI reads DNA cell by cell

- URL: https://jemiboo.com/en/tech/2026-11-13/
- Topics: Tech
- Published: 2026-11-13
- Publisher: Jemiboo Magazine (JEMIBOO)
- Tags: AlphaGenome 2, AlphaGenome, Google DeepMind, genomics AI, single-cell genomics, DNA variants, non-coding DNA, DNA methylation, precision medicine, Korea bio big data, AI drug discovery

![A glass sequencing chip and sample tubes on a laboratory bench](https://jemiboo.com/img/2026-11-13-tech-cover.jpg)

Among DNA variants linked to disease, the share that a model correctly pinpointed as the likely cause rose from 16% to 30%. That is the headline test result for AlphaGenome 2, the genomics AI model Google DeepMind released on November 11. The first AlphaGenome, released a year and a half earlier, answered at the level of tissue averages. The new model lets researchers pick a single cell type and state, such as an immune cell or a gut cell, and predicts what a variant does inside that cell. The same variant can be silent in one cell and drive disease in another. The weight of this release is that the question genomics AI asks has moved from "which variant is it?" to "in which cell does it cause trouble?"

## Key takeaways (as of November 13, 2026)
- Google DeepMind released AlphaGenome 2 on November 11, 2026, a genomics AI model trained on more than 107 million human and mouse cell profiles.
- DeepMind says AlphaGenome 2 found 30% of likely causal variants in a 28-immune-cell-type benchmark (TenK10K), nearly double the 16% reached by the first AlphaGenome fine-tuned on single-cell data.
- AlphaGenome 2 is available now through an API for non-commercial research and is not approved for clinical diagnosis, while Korea's National Integrated Bio Big Data project is preparing its first data release in November 2026.

## AlphaGenome 2 asks about one cell at a time

Only 2% of the human genome codes for proteins. The remaining 98% is non-coding, and it is where the switches sit that decide when, where and how strongly genes are turned on. The first AlphaGenome, released by DeepMind on June 25, 2025, read up to one million DNA letters at once and predicted the effect of those switches down to a single letter. The work was published in Nature in January 2026. AlphaGenome 2 adds one input. Alongside the one-million-letter DNA sequence, the user specifies a tissue, a cell type or a cell state. The training data are bulk and single-cell datasets from 68 human and mouse tissues, and DNA methylation has been added to the outputs. DNA methylation is a chemical mark that keeps a gene switched off without changing the underlying sequence.

| Item | AlphaGenome (June 25, 2025) | AlphaGenome 2 (Nov. 11, 2026) |
|---|---|---|
| DNA read in one pass | Up to 1 million letters | 1 million letters |
| Prediction unit | Fixed outputs per tissue and cell line | A chosen tissue, cell type or state |
| Training data | Hundreds of human and mouse cell types and tissues | 68 human and mouse tissues, more than 107 million cell profiles |
| New prediction | — | DNA methylation |
| Access | API for non-commercial research | API for non-commercial research; private preview for commercial research |

Source: DeepMind blog posts of June 25, 2025 and November 11, 2026.

## 30% of causal variants across 28 immune cell types

The performance comparison used TenK10K, a single-cell immune dataset covering 28 primary immune cell types. The baselines were the first AlphaGenome fine-tuned on single-cell data and a set of other specialist tools.

| Benchmark (TenK10K, %) | AlphaGenome 2 | Comparison model |
|---|---|---|
| Share of likely causal variants identified | 30 | 16 |
| Share with correct direction of gene effect (up or down) | 43 | 23 |

Source: DeepMind announcement of November 11, 2026. The announcement does not give the detailed conditions under which the two figures were measured. The way the model is used has also changed. A researcher uploads single-cell RNA data from a few dozen cells; the model summarizes the state of those cells and produces predictions even for states it has not seen before, with no retraining. In practice, a lab can assess variants against its own cells even when that cell state is missing from public datasets. Sarah Teichmann, a professor at the University of Cambridge and co-founder of the Human Cell Atlas, said in DeepMind's November 11 announcement: "What is exciting about AlphaGenome 2 is that it adapts straight away to new single-cell data without retraining."

![A biobank rack of sample tubes pulled from a cryogenic freezer](https://jemiboo.com/img/2026-11-13-tech-inline.jpg)

## A distant switch and a clue to gut inflammation

DeepMind also tested whether the model could rediscover a known result. A 2024 Nature study traced variants linked to inflammatory bowel disease to a "gene desert," a long stretch of DNA with no genes, and showed that a variant there overactivates the distant ETS2 gene in gut macrophages. AlphaGenome 2 reproduced the finding by computation alone and picked out the causal variant. In its predictions, that variant raised ETS2 only in gut macrophages and had little effect on other cells or genes. The limits are clear. According to Dataconomy, Žiga Avsec, who leads genomics at DeepMind, explained that AlphaGenome 2 needs more compute per prediction and so will not immediately replace the earlier model. DeepMind states that both models are for research only and have not been validated or approved for clinical or diagnostic use.

## For researchers in Korea, a November data release

AlphaGenome 2 can be used now through the API for non-commercial research, and commercial research is open as a private preview. DeepMind is collecting usage proposals through an application form until November 30 and says selected projects may receive extra usage and credits. Because the model needs more compute than its predecessor, labs would do well to check their computing capacity first. In the same month, Korea's National Integrated Bio Big Data project is preparing its first data release. The government program gathers genomic, clinical and public data donated by citizens and makes it available for research. It is run jointly by the Ministry of Health and Welfare; the Ministry of Science and ICT, which oversees science and digital policy; the Ministry of Trade and Industry (formerly the Ministry of Trade, Industry and Energy); and the Korea Disease Control and Prevention Agency (KDCA), the national public health agency.

| Item | Details | Source |
|---|---|---|
| Final target | 1 million people by 2032 | KDCA and partner ministries (Oct. 15, 2026) |
| Phase 1 (2024–2028) | 772,000 people | KDCA and partner ministries (Oct. 15, 2026) |
| Recruitment mix | 585,000 general participants, 140,000 patients with serious diseases, 47,000 patients with rare diseases | KDCA and partner ministries (Oct. 15, 2026) |
| Participants enrolled | About 220,000 (48 medical institutions) | MTN report (Sept. 10, 2026) |
| Eligibility | From national health-screening participants to any citizen | KDCA and partner ministries (Oct. 15, 2026) |
| First data release | November 2026 (date not set) | KDCA and partner ministries (Oct. 15, 2026) |

Minister of Health and Welfare Jeong Eun-kyeong said in the joint press release from the four agencies on October 15 that, starting with the first release in November, the government would support safe and wide use of the data. Global models and Korean data still move on separate tracks. Access to the national data requires an application and a review, and before uploading personal genetic information to an overseas API, researchers should first check their institution's research ethics review standards. For ordinary consumers taking a genetic test, AlphaGenome 2 will not change a diagnosis any time soon.

Genomics AI is shifting from lists of variants toward reading the context of the cell. What fills that context is how much cell data, from how diverse a population, has been collected. Researchers in Korea should keep an eye on this month's bio big data release notice and its access conditions.

## Frequently asked questions
### What is new in AlphaGenome 2?
The first AlphaGenome predicted a fixed set of outputs per tissue and cell line all at once, while AlphaGenome 2 lets users pick one tissue, cell type or cell state and predicts the effect of DNA variants within it. It adds DNA methylation prediction and can work from single-cell data on just a few dozen cells without retraining.
### Can anyone use AlphaGenome 2?
For non-commercial research it is available now through the AlphaGenome API, and DeepMind is accepting usage proposals until November 30, with possible extra usage and credits for selected projects. Commercial research is in private preview, and the model has not been validated or approved for clinical or diagnostic use.
### When will Korea's National Integrated Bio Big Data open?
KDCA and three partner ministries have said the first data release will take place in November 2026, but the exact date has not been announced. Researchers can obtain human biological samples and bio big data after applying and passing a review.

## Sources
1. Google DeepMind, AlphaGenome 2: Deciphering how cells read DNA to better understand disease (Nov. 11, 2026) — https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
2. Google DeepMind, AlphaGenome: AI for better understanding the genome (June 25, 2025) — https://deepmind.google/blog/alphagenome-ai-for-better-understanding-the-genome/
3. Dataconomy, Google DeepMind releases AlphaGenome 2 to study disease (Nov. 12, 2026) — https://dataconomy.com/2026/11/12/google-deepmind-releases-alphagenome-2-to-study-disease/
4. KDCA, Ministry of Health and Welfare, Ministry of Science and ICT and Ministry of Trade and Industry, Any citizen, not only health-screening participants, can now join the National Integrated Bio Big Data project (Oct. 15, 2026) — https://kdca.go.kr/bbs/kdca/42/312967/artclView.do
5. MTN, Pooling scattered medical data on 1 million people: National Integrated Bio Big Data platform opens (Sept. 10, 2026) — https://news.mtn.co.kr/news-detail/2026091015410156064
