Sources checked
What AlphaGenome 2 does differently
Google DeepMind released AlphaGenome 2 on November 11, a DNA model that predicts how a genetic variant changes gene activity in a specific human cell type or state rather than in a tissue average.
The company says it identifies nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model.
Almost every cell carries the same DNA blueprint, but heart muscle cells, skin cells, and immune cells switch different genes on and off, and those patterns shift as a cell changes state.
Most variants that population studies link to disease sit outside genes, in regulatory stretches of DNA that act like dimmer switches. The same variant can do nothing in one cell type and change a gene's output sharply in another.
AlphaGenome 2 is designed to capture those differences.
For Adam Gayoso, a research scientist at DeepMind involved in the project, the big breakthrough is "tying human genetics to cell biology," he told The Verge.
The model gives researchers a way to drill down on genetic variants identified in large populations to see what effects they may be having on a cellular level.
From broad tissues to individual cell types
The first AlphaGenome, previewed in 2025 and published in Nature in January 2026, took a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues at once.
DeepMind says that parallel approach "spread the model's capacity thin" and made it hard to capture subtle differences between closely related cell types.
AlphaGenome 2 fundamentally redesigned how cell types are treated. Instead of predicting in parallel, it is a conditional model that predicts DNA effects for one condition at a time.
Alongside the million-letter DNA sequence, it uses the target biological condition as input at every stage of its architecture. This lets it learn how the same stretch of DNA is read differently across cell types.
The model predicts the same core experimental readouts as the original AlphaGenome — gene expression, histone modifications, and chromatin accessibility — and adds DNA methylation, a widely requested readout.
A cell encoder also lets scientists build a condition from their own single-cell experiment. Given one measurement like single-cell gene activity, the model can predict unmeasured properties such as chromatin accessibility.
Trained on over 100 million cells
DeepMind trained AlphaGenome 2 on over 107 million individual cell profiles from human and mouse bulk and single-cell data, covering 68 tissues.
The company says the model turns static cell atlases — huge data repositories that map individual cells in detail — into a predictive engine for human genetics.
On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 identified causal variants 30 percent of the time versus 16 percent for the best comparison model.
It got the direction of the gene effect right 43 percent versus 23 percent. Those are DeepMind's own evaluations reported in a paper linked from its announcement.
DeepMind's genomics lead Ziga Avsec told The Verge that this specificity matters when the goal is treating a particular disease. "When you develop drugs, you're targeting specific proteins and specific cell types," he said, which makes this granularity vital.
Pinpointing disease mechanisms and adapting to new data
AlphaGenome 2 can also generalize to entirely new single-cell experiments without being retrained.
Scientists can upload a single-cell RNA-seq snapshot of gene activity from just a few dozen cells, and the model compresses it into a cell state representation before conditioning its predictions.
DeepMind says it can match or outperform models trained or fine-tuned directly on those datasets.
In a practical demonstration, DeepMind used a landmark study showing that a variant in a distant regulatory switch drives inflammatory bowel disease by overactivating the ETS2 gene in gut immune cells called macrophages.
AlphaGenome 2 reproduced this finding computationally, singling out the causal variant and showing it selectively boosts ETS2 in those macrophages while leaving other cell types and genes largely unaffected.
Using single-cell data from a COVID-19 mRNA vaccine study that the model had not seen during training, it 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.
DeepMind also says its precomputed scores help pinpoint causal variants and cell types for traits from height to type 2 diabetes.
Cost, limits, and what comes next
AlphaGenome 2 does not replace its predecessor.
Avsec told The Verge its more detailed predictions are "more expensive" to produce, which made it impractical to use for the genome-wide atlas covering every possible single-letter change in the human genome that DeepMind released in September.
The original AlphaGenome will remain in service.
Sarah Teichmann, co-founder of the Human Cell Atlas consortium and professor at the University of Cambridge, said in a statement: "What makes AlphaGenome 2 exciting is how it adapts directly to new single-cell data without retraining, predicting how DNA variants act in newly profiled cell states from even a small sample of cells."
The model is available from today for non-commercial research through the AlphaGenome API. Commercial research access is in a limited private preview.
DeepMind says the model has not been validated or approved for clinical or diagnostic purposes. The announcement does not mention releasing model weights.
Avsec said DeepMind is generating more data for poorly covered cell states.
Future versions could train on data across multiple species, and such models could feed into longer-term efforts to build virtual cells — digital simulations designed to mimic biological systems.
DeepMind is collaborating with the Wellcome Sanger Institute to profile more cell types, lineages, and perturbed states.
AlphaGenome 2 is the latest in a growing push by Google and its rivals to apply AI to fundamental problems in biology and medicine. Pushmeet Kohli, who heads DeepMind's AI for science research, has hinted at more results ahead.
The organization is expected to open an automated materials science lab in the UK soon. Anthropic has been experimenting with drug design and given Claude a wet lab, while OpenAI has been tackling longstanding problems in mathematics and related fields.
The race to build AI that can navigate biology at cellular resolution is only accelerating.
Common questions
How is AlphaGenome 2 different from the original AlphaGenome?
The first AlphaGenome predicted variant effects across broad tissue averages in parallel.
AlphaGenome 2 is a conditional model that takes a specific biological condition alongside the DNA sequence and predicts for one cell type at a time, using that condition at every stage of its architecture.
This captures subtle differences between closely related cell types that the previous model would miss.
What data was AlphaGenome 2 trained on?
The model was trained on over 107 million individual cell profiles from human and mouse bulk and single-cell data, covering 68 tissues. DeepMind says the training data includes recent technologies that capture individual cells and their states at high resolution.
Is AlphaGenome 2 available for commercial use?
AlphaGenome 2 is available for non-commercial research via the AlphaGenome API from today. Commercial research access is in a limited private preview.
The model is not validated or approved for clinical or diagnostic purposes. DeepMind says its predictions cost more compute than the original AlphaGenome, which will remain in service.
Sources & further reading
How this story was made
Written by Kristian Kostov with AI assistance and checked against the linked sources. Company performance claims are attributed to the company. Analysis reflects AiLookout’s interpretation; we have not independently tested the products discussed. The cover is an AI-generated editorial illustration, not a photograph or evidence of an actual event.
Our editorial standards





