Image: Google DeepMind — AlphaGenome 1 vs AlphaGenome 2 comparison (official blog visual)
Key Highlights
- Google DeepMind announced AlphaGenome 2 on 11 November 2026, its next-generation DNA sequence model.
- It predicts how DNA variants affect gene activity in specific cell types and states — a heart muscle cell, a gut immune cell, a vaccine-activated T cell — instead of averaging across whole tissues.
- On the TenK10K benchmark of 28 immune cell types, it recovered 30% of causal variants at 90% precision, against 16% for the original AlphaGenome.
- Trained on more than 107 million individual cell profiles across 68 tissues, from human and mouse data.
- Free for non-commercial research through the AlphaGenome API from launch day; commercial research is in private preview. Research use only — not validated for clinical or diagnostic use.
What happened?
Google DeepMind announced AlphaGenome 2 on 11 November 2026, in a blog post by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec. It is the successor to AlphaGenome, the DNA model DeepMind launched last year, and to the AlphaGenome Atlas, which holds predictions for all 9 billion possible single-letter changes in the human genome.
The problem it attacks is simple to state and hard to solve. Every cell in your body carries the same DNA, yet a heart muscle cell and an immune cell fighting an infection switch different genes on and off. A single-letter DNA variant can therefore be harmless in one cell type and harmful in another — and models that average across whole tissues wash that signal out. AlphaGenome 2 is built to predict variant effects cell by cell, including in cell states the model has never seen before.
How does Google DeepMind AlphaGenome 2 work?
The first AlphaGenome took a one-million-letter stretch of DNA and predicted thousands of experimental readouts across many tissues in a single pass. That was efficient, but it spread the model's capacity thin and made it hard to capture subtle differences between closely related cell types.
AlphaGenome 2 redesigns that approach. It is a conditional model: alongside the million-letter DNA sequence, it takes one target condition — a tissue, a cell type or a cell state — and feeds that condition into every stage of its architecture. It predicts for that condition only, which is how it learns that the same stretch of DNA is read differently by different cells.
Researchers can also build a condition from their own data. Upload a single-cell RNA snapshot from just a few dozen cells, and a cell encoder compresses it into a state the model can use — no retraining needed. The same design lets the model fill in missing lab data: given only gene-activity measurements, it can predict unmeasured properties such as chromatin accessibility, helping researchers decide which experiments to run next. Its readouts cover gene expression, histone modifications and chromatin accessibility, plus DNA methylation, a widely requested new addition.
Illustrative image: the structure of DNA and its bases (Wikimedia Commons, public domain)
AlphaGenome 1 vs AlphaGenome 2
| Feature | AlphaGenome 1 | AlphaGenome 2 |
|---|---|---|
| Prediction target | Broad tissues, all at once | One cell type or state at a time |
| Model design | Fixed parallel outputs | Conditional model, condition fed in at every stage |
| Training data | Bulk tissues and cell lines | Bulk + single-cell, 107M+ cell profiles, 68 tissues |
| Causal variants found (TenK10K) | 16% at 90% precision | 30% at 90% precision |
| Effect direction called | 23% | 43% |
| New readouts | Gene expression, histone marks, chromatin access | Adds DNA methylation |
| New cell states | Not supported | Works without retraining |
What the tests actually show
The headline result comes from TenK10K, a large single-cell map of 28 primary immune cell types. There, AlphaGenome 2 identified 30% of putatively causal variants that change gene expression, against 16% for the original AlphaGenome fine-tuned on single-cell data — and it called the direction of a variant's effect (whether it turns a gene up or down) 43% of the time, against 23%. On genome regions held out of training, it also explained a median 34% of cell-type-specific gene-expression variance within tissues, against 23% for its predecessor.
DeepMind also tested it against real disease biology. In 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. AlphaGenome 2 reproduced that finding computationally: it singled out the variant rs2836882 from a seven-variant set and showed it selectively boosts ETS2 in gut monocytes and macrophages, leaving other cell types largely unaffected.
A second test used single-cell data from a COVID-19 mRNA vaccine study that the model had never seen. AlphaGenome 2 distinguished vaccine-activated T cells from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly when cells are in the activated state — an effect invisible in resting cells.
For researchers who do not want to run large comparisons themselves, DeepMind precomputed two resources: enhancer–gene links across 640 primary single-cell reference conditions and 39,323 protein-coding and long non-coding RNA genes, and scores for more than 6.8 million trait-associated variants across 510 human conditions, covering traits from height to type 2 diabetes.
What it means for researchers in India
There is no India-specific announcement here, but the access terms matter for Indian readers. AlphaGenome 2 is available from 11 November for non-commercial research through the AlphaGenome API, and DeepMind says that use is free, with query rates varying by demand. Interactive Colab tutorials for expression conditioning are available at launch, which lowers the barrier for university labs and genomics researchers in India who want to test a variant hypothesis before committing to wet-lab experiments.
That is the practical value: population-genetics and disease-genetics groups can narrow a long list of candidate variants to the ones most likely to matter in a specific cell type, then design experiments around those leads. Genomics PLC, a UK genomics company, has already said it plans to incorporate AlphaGenome 2 into its Mystra platform. DeepMind also plans AlphaGenome 2 Skills for Google Antigravity, its agentic development platform, pairing the model with tools like Gemini 4.
The hard limit is equally clear. Both AlphaGenome versions are for research use only and have not been validated or approved for clinical or diagnostic purposes. This is not a tool that reads your personal genome, and no Indian hospital or clinic should be using it to make patient decisions.
What AlphaGenome 2 cannot do yet
DeepMind and its technical paper are candid about the limits, and they matter. Even at high precision, the model recovers only a minority of the variants that statistical analyses flag as likely causal — 30% is nearly double 16%, but 70% are still missed. It can underestimate effects outside accessible chromatin, at very distant regulatory elements, or through mechanisms that act after transcription. Transient, stimulated and disease states remain underrepresented in training, and fine-grained neuronal subpopulations are still hard to tell apart.
There is also a conceptual limit: the model's condition representations come from observational, steady-state profiles, so they capture correlations rather than proven cause and effect. It remains uncertain whether it can predict responses to gene knockouts, overexpression or dynamic cell-state transitions. And there is a compute trade-off — every condition needs a separate prediction, making it more intensive to run than the original model's one-pass design. The results are hypotheses that need experimental follow-up, not conclusions.
Should you care?
For context on how DeepMind's AI is being applied elsewhere, read our explainers on DeepMind's AlphaProof verification of Signal's encryption and MBZUAI's PAN 2 open-world model, and on WikiTrace's approach to LLM agent memory.
FAQs
What is AlphaGenome 2?
AlphaGenome 2 is Google DeepMind's next-generation DNA sequence model, announced on 11 November 2026. It predicts how DNA variants affect gene regulation in specific human cell types and states, helping scientists find the cellular drivers of disease.
How is AlphaGenome 2 different from AlphaGenome 1?
The original model predicted variant effects across broad tissues all at once. AlphaGenome 2 is a conditional model that predicts for one cell type or state at a time, trained on more than 107 million single-cell profiles — and it finds nearly twice as many causal variants on the TenK10K benchmark (30% vs 16%).
Is AlphaGenome 2 free to use?
Yes, for non-commercial research: it is available through the AlphaGenome API free of charge, with Colab tutorials provided. Commercial research is in a limited private preview, and query rates vary with demand.
Can AlphaGenome 2 diagnose disease?
No. DeepMind states explicitly that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. It produces research hypotheses that need experimental confirmation.
Takeaway
AlphaGenome 2 moves DNA modelling from tissue averages to individual cell types, and the numbers back the redesign: nearly twice as many causal variants found, new cell states handled without retraining, and free access for non-commercial research. The honest caveat is that it still recovers only a minority of candidate variants and its predictions are correlations, not proof. For genomics researchers, it is a sharper starting point for the next experiment — not a shortcut past it.
If you work in genomics, start with the free AlphaGenome API and the expression-conditioning Colab tutorials linked in the Sources below.
Sources
- Google DeepMind — AlphaGenome 2: Deciphering how cells read DNA to better understand disease (primary source, 11 November 2026)
- Unite.AI — Google DeepMind's AlphaGenome 2 Predicts Cell-State Variant Effects
- The Verge — DeepMind unveils AlphaGenome 2, a DNA model that predicts variant effects in specific cell types
- AlphaGenome documentation — API access and tutorials
Comments
Post a Comment