Google DeepMind Unveils AlphaGenome 2, Which Predicts DNA Variant Effects Cell Type by Cell Type and Recovers 30% of Causal Variants in Its Own Test
The successor to AlphaGenome conditions its predictions on a chosen cell type or state, but the benchmark gains come from DeepMind's own evaluations and the model is for research use only.

Key facts
- Google DeepMind announced AlphaGenome 2 on 11 November 2026, a conditional DNA model that predicts variant effects for one chosen cell type, tissue or cell state at a time.[1][2]
- On the TenK10K benchmark of 28 primary immune cell types, AlphaGenome 2 identified causal variants in 30% of cases, against 16% for the original AlphaGenome fine-tuned on single-cell data.[1][2][3]
- AlphaGenome 2 predicted the direction of a variant's effect on a gene correctly in 43% of cases on TenK10K, versus 23% for the fine-tuned original model, according to DeepMind.[1][2]
- AlphaGenome 2 is available for non-commercial research through the AlphaGenome API, commercial access is in private preview, and DeepMind says it is not validated for clinical or diagnostic use.[1][2]
- The Decoder noted that most AlphaGenome 2 performance claims come from DeepMind's own benchmarks and that its coverage included no independent evaluation.[2]
Google DeepMind announced AlphaGenome 2 on 11 November 2026, a DNA-sequence model that predicts how a genetic variant changes gene activity in one chosen human cell type rather than across blended tissues. In DeepMind's own test on 28 primary immune cell types, the model recovered about 30% of causal variants, against 16% for the original AlphaGenome after fine-tuning on single-cell data, according to DeepMind's announcement.
What is new in AlphaGenome 2?
The first AlphaGenome produced fixed, parallel predictions for bulk tissues and cell lines. AlphaGenome 2 is a conditional model: it takes a 1-million-letter DNA sequence plus one target condition, such as a tissue, cell type or cell state, and predicts effects for that condition alone, DeepMind says. The condition feeds into every stage of the architecture. Outputs include gene expression, histone modifications and chromatin accessibility, plus DNA methylation, which is new in this version.
DeepMind's post gives two figures for training data. The introduction cites more than 100 million single-cell profiles across 68 tissues, while the body text says "over 107 million" cell profiles from human and mouse. The Decoder and The Neuron both repeat the 107 million figure. The two numbers differ slightly, and DeepMind does not explain why.
The model can also fill gaps. Given one measurement, such as single-cell gene activity, it can predict unmeasured properties like chromatin accessibility, which DeepMind says can help researchers choose which experiments to run first.
How well does it perform?
The headline benchmark is TenK10K, a dataset covering 28 primary immune cell types. AlphaGenome 2 identified causal variants in 30% of cases, compared with 16% for the original model fine-tuned on single-cell data and other specialised tools. For the direction of a variant's effect on a gene, the figures were 43% against 23%. DeepMind also says the model finds nearly twice as many gene-expression-changing variants as the previous state-of-the-art model, and that it tracks continuous changes such as blood stem cells maturing into red blood cells more closely than the fine-tuned original.
Two case studies illustrate the claims. In inflammatory bowel disease, the model computationally reproduced a known mechanism: a variant in a distant regulatory switch that boosts the gene *ETS2* in gut macrophages, with other cell types largely unaffected. The Neuron said that result comes from a 2024 Nature study, which makes it a useful check but, in its view, not the decisive test. In a COVID-19 mRNA vaccine study, the model separated vaccine-activated T cells from resting ones and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect that would be invisible in resting cells.
Can researchers use their own data?
DeepMind says researchers can upload single-cell RNA-sequencing data from as few as a few dozen cells. The data is compressed into a cell-state representation that conditions the model, with no retraining. According to DeepMind, the approach matched or outperformed models trained or fine-tuned directly on those datasets. The Decoder noted that this claim is DeepMind's own assertion.
The company has also precomputed predictions for trait-associated variants across hundreds of cell and tissue conditions. These are meant to help genome-wide association studies, which find regions of the genome linked to traits such as height or type 2 diabetes but often struggle to say which variant, gene or cell type is responsible. The model can additionally extend enhancer-to-gene mapping to primary human cell types, such as brain astrocytes, where lab cell lines fall short.
Who has access, and who is commenting?
AlphaGenome 2 is available now for non-commercial research through the AlphaGenome API, with Colab tutorials. Commercial research access is in private preview. DeepMind says researchers can submit use cases by 30 November for possible extra quota and credits, and that "AlphaGenome 2 Skills" for Google Antigravity is planned soon. The announcement lists Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec as authors.
The post carries two outside endorsements. Sarah Teichmann of the University of Cambridge and co-founder of the Human Cell Atlas said: "What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining." Peter Donnelly, chief executive of Genomics Limited, said GWAS is central to drug discovery but hard to translate into what a drug programme needs, and that his company plans to incorporate the model into its Mystra platform. Both comments were published by DeepMind, not gathered independently.
Why the benchmark numbers need care
The Decoder pointed out that most performance claims come from DeepMind's own benchmarks and comparisons, and that its article included no independent evaluation. It also said the inflammatory bowel disease result was computational only, with no wet-lab validation reported in the piece. The Neuron stressed that the 30% figure comes from one evaluation, not from a share of diseases solved or patients diagnosed, and that even on that test there is "considerable room" for improvement.
DeepMind itself says the model is for research use only and is not validated or approved for clinical or diagnostic use. It also says directional predictions are made only when the model is confident enough. The post reports no detailed error analysis beyond the benchmarks, pointing readers to the linked paper.
What happens next?
The question The Neuron raised is the sharpest one: whether outside groups can use AlphaGenome 2 to find mechanisms they would otherwise have missed, and then confirm them at the bench. Reproducing a known inflammatory bowel disease effect shows the model can recover published biology. Surfacing a new, experimentally confirmed mechanism would show it can guide discovery.
The practical value for drug developers may lie in narrowing a long list of candidate variants into a short list worth testing, which is how The Neuron framed the model's role. Whether that saves time depends on how often the shortlists hold up against independent data, and none of the sources read for this article reports such tests yet. The 30 November use-case deadline and the commercial preview will show how quickly academic and industry groups start publishing results.
