AlphaGenome 2: a conditional DNA model trained on 107 million cell profiles finds nearly twice the expression-changing variants. What it does and does not mean for drug discovery
DeepMind's second genome model answers the question the first could not: what a variant does in this cell type. That is the question target-discovery teams actually ask, which is why the commercial preview and the research-only label both matter.

Google DeepMind released AlphaGenome 2 on November 11. Like its predecessor, it reads one million letters of DNA and predicts what that stretch of sequence does: which genes it switches on, how accessible the chromatin is, which histone marks are present. Unlike its predecessor, it does so for one condition at a time. You give it a sequence and a target tissue, cell type or cell state, and it tells you what a variant would do there. The model also gains a new output, DNA methylation, and the ability to take a user's own single-cell data, compress it into a representation of that cell state, and predict for it, including readouts the user never measured.
The numbers
Training data was more than 100 million individual cell profiles across 68 tissues from human and mouse samples; DeepMind quotes 107 million. DeepMind reports that the model identifies nearly twice as many variants that change gene expression as the previous state of the art. On TenK10K, a benchmark spanning 28 immune cell types, it found causal variants 30% of the time against 16% for the original AlphaGenome fine-tuned on single-cell data and other specialised tools, and predicted the direction of a variant's effect on a gene correctly 43% of the time against 23%. Precomputed predictions exist for hundreds of human cell and tissue conditions.
Why cell type is the whole point
Most disease-associated variants from genome-wide association studies sit in regulatory DNA, not in genes, and most of them act in a handful of cell types. A variant that raises the risk of inflammatory bowel disease may matter only in a particular subset of gut immune cells; averaged across all tissues its effect vanishes. The first AlphaGenome predicted many tissues at once and was strongest where the training atlases were deep. The conditional design means a team studying a rare cell population can bring a few dozen cells' worth of single-cell data and get a prediction tuned to them. DeepMind's case studies include a computational reproduction of the ETS2 mechanism in inflammatory bowel disease and a vaccine-response example; both are predictions, not experiments.
Access and limits
AlphaGenome 2 is available now through the AlphaGenome API for non-commercial research, with Colab tutorials for expression conditioning. Commercial research access is a private preview. DeepMind says AlphaGenome 2 skills for its Antigravity agent platform are coming, which would let a coding agent run variant queries as part of a larger workflow. The announcement does not say whether weights will be released or under what licence, and it is explicit that the model has not been validated or approved for clinical or diagnostic use. Every performance figure is DeepMind's own, on benchmarks it chose; independent evaluations on held-out cohorts are the next thing to look for.
The counter-argument
Computational variant-effect prediction has a long record of impressive benchmark gains that translate slowly into validated targets, because the bottleneck is wet-lab confirmation, not candidate generation. Doubling the number of plausible variants doubles the follow-up list. Skeptics will also note that single-cell atlases are uneven: cell types that are well represented in public data will be predicted well, and the rare populations where new biology hides may not be. And a model whose commercial terms are undisclosed is hard to build a programme on.
What to do
For discovery teams, the practical step is a bake-off: take variants you have already validated in a known cell type and test whether AlphaGenome 2, conditioned on your own single-cell data, recovers them and their direction. That is cheap on the research API and tells you more than any published benchmark. For clinical and diagnostics groups, the answer is simpler: nothing changes until there is validation on patient cohorts, and the research-only label should be read literally. For Google Cloud customers, ask the account team what the commercial preview will cost and where data sent to the API is processed.
What to watch
Independent benchmarks on cohorts DeepMind did not use; the commercial terms; whether weights are released, which would matter for hospital systems that cannot send genomic data to an external API; and the first published target found with AlphaGenome 2 that survives wet-lab confirmation.
The story so far
- Nov 15, 2026 Mirendil's $1 billion Series A at a reported $5 billion, eleven months after founding: what investors are buying when th…
- Nov 15, 2026 Microsoft's Lazarus: seven edited neurons out of 12,000 backdoor a coding agent 98.8% of the time and pass every standar…
- Nov 14, 2026 llama.cpp 1.0.0 after 3.5 years: semantic versioning, 35 deprecated functions gone, FP8 KV cache. What the stability pro…
- Nov 14, 2026 IBM's Granite 4.1 LLMON previews train the model to separate instructions from data, claiming near-zero indirect prompt-…
Key terms in this story
- Evaluation (eval) A test set with known-good answers used to measure a model or system on your own tasks.
- Benchmark A standard test used to compare models, such as a set of coding tasks or exam questions.
- Agent An AI system that pursues a goal by taking actions with tools, in a loop, rather than answering a single question.
More in Health & medicine
All →
GE HealthCare's Critical Care Suite 3.0 clears the FDA: feeding-tube detection on the x-ray machine at 99.2% sensitivity, offline, and for children for the first time. What hospitals should ask before switching it on
On-device imaging AI sidesteps the integration and data-sharing problems that stall most hospital deployments; the questions that remain are about the…

Isomorphic Labs raises $5 billion with no drug in a human: the largest private biotech round ever, $7.7 billion in twenty months, and what the money is actually buying
The DeepMind spinout has now raised more than most public biotechs are worth, on preclinical data it has not published. The bet is that an AI design e…

Apple and OpenEvidence open a 200,000-person health chatbot study: wearable and lab data in, no model training, no device claim. What health systems should watch
Apple has chosen the clinical search tool most US doctors already use as the engine for its first consumer health conversation, and is running it as r…
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.