AlphaGenome 2: DeepMind DNA Model Reads Cells

Every cell in your body carries the same DNA, yet a heart muscle cell reads that DNA differently from an immune cell fighting an infection. That difference is where much of human disease hides — and it is also where DNA models have been weakest. On 11 November 2026, Google DeepMind announced AlphaGenome 2, a model that predicts how a DNA variant behaves in one specific cell type at a time, instead of averaging across whole tissues. On its headline test it found nearly twice as many disease-linked variants as its predecessor.

Quick answer: DeepMind's AlphaGenome 2 predicts how DNA variants act in single cell types, finding nearly twice as many disease-linked variants as before. Research use only.

The short version: AlphaGenome 2 is a conditional DNA model. Give it a million-letter stretch of DNA plus one cell type or state, and it predicts how variants in that stretch change gene activity in that cell alone. It is available now for non-commercial research through an API. It is a research tool, not a clinical one — DeepMind says so plainly, and every performance figure in this article is DeepMind’s own, with no independent replication published yet.

What AlphaGenome 2 actually does

AlphaGenome 2 treats the cell type as an input, not an afterthought. The original AlphaGenome, launched last year, took a million-letter DNA sequence and predicted molecular activity across thousands of experiments from many tissues all at once. AlphaGenome 2 keeps the million-letter input but adds one biological condition — a tissue, a cell type, or a state — fed into every stage of the network. The condition can come from a reference catalogue of cell types, or be generated on the fly from a scientist’s own single-cell experiment through a cell encoder.

For each condition, the model predicts the same core readouts as its predecessor — gene expression, histone modifications, chromatin accessibility — plus one widely requested new readout: DNA methylation, the chemical tags that switch genes on or off. It can also fill in missing data. Feed it a single measurement, such as a cell’s gene activity, and it predicts unmeasured properties such as how tightly the DNA is packaged — which helps researchers decide which lab experiments are worth running next.

Why the first AlphaGenome was not enough

Predicting everything at once spread the first model too thin. DeepMind’s own account of the redesign is blunt: the parallel approach made it hard to capture subtle differences between closely related cell types. A variant that matters in one immune cell and not in its near-identical neighbour would blur into a tissue average.

The first model was still useful enough to earn a companion tool, the AlphaGenome Atlas, which holds predictions for all 9 billion possible single-letter changes in the human genome. But for the questions researchers actually ask — which variant, in which cell, drives this disease — an average was the wrong answer. AlphaGenome 2 is the attempt to give a specific one.

What the numbers say — and whose numbers they are

Every figure below is DeepMind’s own, from its 11 November announcement. No outside group has independently reproduced them yet, and this article labels them as the lab’s claims, not established facts.

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. On TenK10K, a single-cell map of 28 primary immune cell types, it identified 30% of causal variants, against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools. It also got the direction right — whether a variant turns a gene up or down — in 43% of cases, against 23% for the predecessor. That is the “nearly twice as many” in the headline: a real improvement on a real benchmark, measured by the people who built the model.

Measure	AlphaGenome 1 (fine-tuned)	AlphaGenome 2
Causal variants found (TenK10K, 28 immune cell types)	16%	30%
Direction of gene effect correct	23%	43%
Training data	Bulk tissue averages	107M+ single-cell profiles, 68 tissues
Model design	Parallel predictions, all tissues at once	Conditional: one cell type or state per run
New readout	—	DNA methylation

The model also tracks continuous change, not just static states. DeepMind reports it follows gene expression as blood stem cells mature into red blood cells and immune lineages with considerably greater accuracy than the fine-tuned predecessor — the kind of moving target that tissue averages cannot capture at all. Vendor benchmarks keep arriving with the same caveat attached: OpenAI’s BioBasicsBench scored GPT-6 Astra 66.7% on messy lab work, and there too the number is the vendor’s own, measured on its own test.

Two cases where the cell type changed the answer

The strongest evidence in the announcement is not a benchmark table. It is two cases where looking at the wrong cell would have given the wrong answer.

The first is inflammatory bowel disease. A landmark study showed that a variant in a distant regulatory switch drives the disease by overactivating the ETS2 gene in gut immune cells called macrophages. Working purely computationally, AlphaGenome 2 reproduced that finding: it singled out the causal variant and showed that it selectively boosts ETS2 in gut macrophages, while other cell types and genes stayed largely unaffected. That is a reproduction, not a lab confirmation — the model re-derived a known result from sequence alone, which is exactly what researchers want a prediction tool to do before they trust it on unknown variants.

The second is a COVID-19 mRNA vaccine study. Given single-cell data the model had never seen, AlphaGenome 2 distinguished vaccine-activated T cells from resting ones — and revealed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. In resting cells, that effect would have stayed invisible. A tissue-average model would have missed it entirely.

Notably, the model can do this for cell states it has never been trained on. Scientists can upload a single-cell RNA snapshot from just a few dozen cells, and AlphaGenome 2 conditions itself on that unseen state — DeepMind says it matches or outperforms models trained directly on those datasets. That matters because disease and vaccination create states no atlas contains yet. It is the second concrete biology result from an AI system this month: Claude recently mapped a 40-year-old protein vault mystery, also from sequence data and lab work combined.

Who can use it now, and how

AlphaGenome 2 is available today for non-commercial research through the AlphaGenome API, with interactive Colab tutorials for expression conditioning. Commercial research access is in a private preview. DeepMind is also collecting community use cases through a form until 30 November 2026, with selected projects eligible for extra quota and credits, and AlphaGenome 2 Skills are coming soon to Google Antigravity, its agentic development platform, where the model would pair with tools such as Gemini 4 for automated research workflows.

For researchers who do not want to run the model at all, DeepMind has precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions. Two uses stand out. First, enhancer–gene mapping in primary human cell types — such as brain astrocytes, which support and protect neurons — instead of the lab-grown cell lines that most CRISPR screens rely on. Second, narrowing genome-wide association studies, which flag broad stretches of DNA linked to traits from height to type 2 diabetes, down to the specific variant and cell type most likely involved. When the model is confident, it also predicts whether a variant turns its gene up or down — a critical clue for drug development, since drugs target specific proteins in specific cell types.

Early industry interest is already visible: Genomics PLC chief executive Peter Donnelly said the company looks forward to incorporating the model into its Mystra platform, according to launch coverage.

What it is not

AlphaGenome 2 is not a diagnostic tool, and DeepMind draws that line in the announcement itself. Both AlphaGenome models are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. A prediction that a variant affects a gene in a cell type is a hypothesis for a lab to test, not a result to act on in a clinic.

There are two more honest limits. First, the conditional design has a cost: every condition needs its own prediction run, so a genome-wide atlas built with AlphaGenome 2 is impractical — the original model stays in service for broad scans. Second, all the headline figures are launch-day claims from the lab that built the model. Until outside groups rerun them, treat 30% versus 16% as DeepMind’s measurement of its own progress, not as a settled fact about the field.

Our take: a five-question test for genomics AI claims

AlphaGenome 2 is a genuine step, and the step is narrow in exactly the way that makes it believable. It does not claim to read genomes better in general. It claims to read them better in one cell at a time — and its two showcase cases are precisely the cases where that distinction is the whole story. The IBD reproduction and the vaccine-activated T cells are not party tricks; they are the failure mode of the old approach, fixed in public.

But the same narrowness is the test to apply to the next genomics announcement. Before sharing any “AI finds the gene for X” claim, ask these five questions:

Which cell type? A tissue average is not a cell answer. If the claim does not name the cell, it has not said anything yet.
Whose numbers? AlphaGenome 2’s figures are DeepMind’s own, unreplicated. A vendor benchmark is a starting point, not a verdict.
Computed or confirmed? Reproducing a known result in software is strong evidence for a prediction tool — and still not a lab result.
Prediction or diagnosis? Research-only models do not become clinical tools by being impressive. The line is a regulatory fact, not a modesty note.
Can you try it? An API for non-commercial research, a private preview for commercial use, and a 30 November community window are verifiable facts. A model nobody can touch is a press release.

On those five questions, AlphaGenome 2 scores well: it names its cells, labels its own figures, reproduces rather than asserts, states its research-only status, and ships an API on day one. The open question is the one no launch announcement can answer — whether outside labs, working on their own cells and their own variants, get the same doubling. That answer arrives over the next few months, one replication at a time.

If you are…	Use this route	What to expect
A genetics researcher with GWAS hits	Precomputed predictions across hundreds of cell conditions	Shortlist of causal variants + the cells they act in
A lab with single-cell data	AlphaGenome API + cell encoder, Colab tutorials	Predictions for your own unseen cell states, no retraining
A drug developer	Commercial research private preview	Access on DeepMind’s terms; no public timeline
A clinician or patient	Wait	Research use only — not validated for clinical or diagnostic use

Frequently Asked Questions

What is AlphaGenome 2?

AlphaGenome 2 is Google DeepMind’s next-generation DNA sequence model, announced on 11 November 2026. It predicts how DNA sequences and their variants affect specific human cell types, helping scientists find the cellular drivers of disease. It is available now for non-commercial research through the AlphaGenome API.

How is it different from the first AlphaGenome?

The first AlphaGenome predicted molecular activity across many tissues at once, which spread its capacity thin. AlphaGenome 2 is a conditional model: it takes a DNA sequence plus one cell type or state and predicts effects for that condition alone. On the TenK10K benchmark it found 30% of causal variants against 16% for its fine-tuned predecessor.

Can AlphaGenome 2 diagnose disease?

No. DeepMind states that both AlphaGenome models are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. Its predictions are hypotheses for laboratory research, not medical results.

Who can access AlphaGenome 2?

Non-commercial researchers can use it now through the AlphaGenome API, with interactive Colab tutorials. Commercial research access is in a private preview. DeepMind is collecting community use cases through a form until 30 November 2026, and AlphaGenome 2 Skills are coming to Google Antigravity.

What does “nearly twice as many variants” mean exactly?

On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of putatively causal gene-expression variants at 90% precision, against 16% for the original AlphaGenome fine-tuned on single-cell data. It also predicted the direction of gene effects correctly in 43% of cases, against 23%. Both figures are DeepMind’s own and have not been independently replicated yet.

Sources and methodology

Claims in this article rest on Google DeepMind’s own announcement, “AlphaGenome 2: Deciphering how cells read DNA to better understand disease” (11 November 2026), read in full, with independent coverage from The Decoder (12 November 2026) and Interesting Engineering (12 November 2026) used to corroborate availability, benchmarks and industry reaction. All performance figures are DeepMind’s own and are labelled as such; no independent replication had been published at the time of writing. This article was desk-researched on 12 November 2026 and was not hands-on tested — AlphaGenome 2 is accessed through an API for research use, and the author did not run the model.

Sources: Google DeepMind — AlphaGenome 2 announcement (11 Nov 2026) · The Decoder — AlphaGenome 2 cell-type resolution (12 Nov 2026) · Interesting Engineering — AlphaGenome 2 causal variants (12 Nov 2026) · Internal: Claude Mapped a 40-Year-Old Protein Vault Mystery · OpenAI BioBasicsBench: GPT-6 Astra Scores 66.7% on Messy Lab Work · Google DeepMind AlphaProof Code Formally Verified Signal Encryption Code

Arva Rangwala

Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.