On the TenK10K immune cell dataset, AlphaGenome 2 picked out 30% of causal variants. The original AlphaGenome found 16%. Google DeepMind announced the new DNA model on November 11. The gain comes from one design change: it now predicts DNA effects one cell type at a time.
The model is open now for non-commercial research through the AlphaGenome API. Commercial research access is in private preview. DeepMind detailed the release in its AlphaGenome 2 announcement blog post and an accompanying technical paper.
What AlphaGenome 2 does
Every cell in the body carries the same DNA. Different cell types switch different genes on and off. A heart muscle cell and an immune cell read the same code in very different ways.
Global cell atlas projects have mapped gene activity in millions of cells. Yet predicting how a single-letter DNA change alters that activity in a given cell type is still an open problem. AlphaGenome 2 targets that gap.
The model reads a DNA sequence up to a million letters long. It then predicts how that sequence, and changes to it, affect a specific tissue, cell type or cell state. DeepMind says it can also make predictions for cell and disease states the model has not seen before.
What changed from the first AlphaGenome
The first model predicted many tissues and thousands of experiments in one pass. That approach made it hard to capture subtle differences between closely related cell types.
AlphaGenome 2 is a “conditional” model. It predicts for one condition at a time. It also takes that condition as input at every stage of its architecture. That lets it learn how the same stretch of DNA reads differently across cell types.
| Feature | AlphaGenome (2025) | AlphaGenome 2 (2026) |
|---|---|---|
| Resolution | Broad tissue averages | Single cell types and cell states |
| Prediction style | Many tissues in parallel | One condition at a time |
| DNA methylation | No | Added |
| Causal variants found (TenK10K) | 16% | 30% |
| Correct effect direction (TenK10K) | 23% | 43% |
The baseline figures in the last two rows refer to the original AlphaGenome fine-tuned on single-cell data, along with other specialized tools. DeepMind says the new model finds nearly twice as many variants that change gene expression as the previous state-of-the-art model.
AlphaGenome 2 keeps the earlier readouts, including gene expression, histone modifications and chromatin accessibility. It adds DNA methylation predictions.
DeepMind trained the model on bulk and single-cell data from human and mouse samples. That data includes over 107 million individual cell profiles across 68 tissues.
Why it matters for disease research
Many disease-linked variants sit in enhancers. These are distant regulatory switches that control genes from far away. Their effects are hard to trace, and they often matter only in one kind of cell.
DeepMind tested this with a known case. An earlier study found that a variant in a distant switch drives inflammatory bowel disease (IBD). It does so by overactivating the ETS2 gene in gut immune cells called macrophages. AlphaGenome 2 reproduced this result computationally. It pinpointed the variant. It also showed the variant raises ETS2 activity mainly in gut macrophages, with little effect elsewhere.
The model also tracks gradual changes. It better captures how gene expression shifts as blood stem cells mature into red blood cells and immune cell lines.
Precomputed scores for GWAS variants
Genome-wide association studies (GWAS) usually flag broad DNA regions tied to a disease. They rarely name the exact causal variant, gene or cell type. To save scientists heavy compute work, DeepMind precomputed AlphaGenome 2 predictions for trait-linked variants across hundreds of human cell and tissue conditions.
DeepMind applied these scores to complex traits ranging from height to type 2 diabetes. When confident, the model also predicts whether a variant raises or lowers gene activity.
The scores also help with enhancer-gene maps. CRISPR screens mostly use stable, lab-grown cell lines, so maps for real human cells are often missing. AlphaGenome 2 extends this mapping to primary human cells, such as brain astrocytes.
Technical detail: new cell states without retraining
A researcher can upload a single-cell RNA snapshot from a few dozen cells. The model compresses it into a cell state and uses it as the condition. DeepMind says this can match or outperform models trained or fine-tuned directly on those datasets.
In one test, DeepMind used unseen single-cell data from COVID-19 mRNA vaccine recipients. AlphaGenome 2 told vaccine-activated T cells apart from resting ones. It also found that a variant linked to severe COVID-19 disrupts gene activity more strongly in activated cells. A model looking only at resting cells would miss that effect.
The conditional design also supports cross-modality imputation. Given one measurement, such as single-cell gene activity, the model can predict unmeasured properties like chromatin accessibility. DeepMind pitches this as a way to plan follow-up experiments.
Who can use it, and how
- Academic and non-commercial researchers: Access is open through the AlphaGenome API documentation site, with interactive Colab tutorials for expression conditioning.
- Companies: Commercial research use is in private preview only.
- Agent users: DeepMind plans to release AlphaGenome 2 Skills in Google Antigravity soon, alongside agentic tools such as Gemini 4.
- Project pitches: DeepMind invites researchers to submit use cases by November 30. Selected projects may receive extra quota and credits.
DeepMind stresses that AlphaGenome and AlphaGenome 2 are for research use only. Neither has been validated or approved for clinical or diagnostic use.
A short AlphaGenome timeline
- June 25, 2025: DeepMind introduced the first AlphaGenome as an API preview for non-commercial research.
- September 8, 2026: DeepMind released AlphaGenome Atlas, with predictions for all 9 billion possible single-letter changes in the human genome.
- November 11, 2026: AlphaGenome 2 launched with cell-type predictions.
Outside reaction and open questions
Sarah Teichmann, co-founder of the Human Cell Atlas consortium and Professor of Stem Cell Medicine at the University of Cambridge, welcomed one feature. She said the model “adapts directly to new single-cell data without retraining.” Peter Donnelly, CEO of Genomics PLC, said his company plans to add it to its Mystra platform.
What remains unclear is when commercial access will open beyond the private preview. DeepMind has also not given a date for the Antigravity skills. It says ongoing work with the Wellcome Sanger Institute will profile more cell types, developmental lineages and perturbed cell states. Independent checks of the benchmark gains, outside DeepMind’s own tests, are the next thing to watch.
