DeepMind’s AlphaGenome 2 predicts variant effects one cell state at a time, and the atlas stays on the older model
Wednesday’s post puts 30 percent against 16 percent for causal variants on the TenK10K immune-cell set. Robert Hart’s Verge story, at 3:00 p.m. UTC, quotes Žiga Avsec on why that detail is too expensive for last month’s genome-wide atlas.
Deep dive4 minIn-depth
Priya Shah4 min read
- DeepMind
- AlphaGenome
- Genetics

Photo: Google DeepMindSource page
Google DeepMind on Wednesday published AlphaGenome 2, a DNA model that predicts how a genetic variant lands in a particular cell type or cell state rather than in a tissue average. The post is by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec. Robert Hart’s story in The Verge is timestamped 3:00 p.m. UTC. The model is available the same day for non-commercial research through the AlphaGenome API, and in a private preview for commercial research. A line at the bottom of the post says AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic use.
The change from the first AlphaGenome is how a cell is handed to the model. Last year’s version took a million-letter DNA sequence and predicted properties across thousands of experiments from many tissues at once. The post says that parallel pass spread the model’s capacity thin and made close cell types hard to separate. AlphaGenome 2 takes one condition at a time, defined as a tissue, a cell type, or a state, and uses that condition at every stage. The condition can come from a reference catalog, or from a scientist’s own single-cell experiment, compressed by what the post calls a cell encoder. The same stretch of DNA can then be scored as it is read in one cell and not another.
Two training counts, and one table Hart does not print
The summary line both accounts carry is that the model identifies nearly twice as many variants that change gene expression in cells, compared with the lab’s previous state of the art. Hart writes that it was trained on more than 100 million individual cell profiles across 68 tissues, which is the sentence in the post’s opening. A later paragraph says the training set was bulk and single-cell data from human and mouse samples, including over 107 million individual cell profiles. Those are two counts in one post. Hart does not print the second.
The percentage pair is also only in the post. On the TenK10K dataset, which the post describes as a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified nearly twice as many causal variants, 30 percent against 16 percent, and nearly twice as many directional gene effects, 43 percent against 23 percent, as the original AlphaGenome fine-tuned on single-cell data and other specialized tools. Hart’s “nearly twice” is the summary. It is not those four numbers. The post says the model also tracks continuous changes, such as human blood stem cells maturing into red blood cells and immune lineages, and that it adds DNA methylation, which it calls a widely requested readout, to the gene expression, histone, and chromatin measurements the first model already predicted.
A disease variant the lab says it recovered, and a state it had not been trained on
The inflammatory-bowel example is the lab’s account of a published finding, not a new patient result. The post says a landmark study found that a variant in a distant regulatory switch drives the disease by overactivating the ETS2 gene in gut macrophages. AlphaGenome 2, making predictions across immune cell types in the gut, singled out that variant and showed a selective boost to ETS2 in those macrophages, leaving other cell types and genes largely unaffected. The post does not name the study’s authors or its journal.
The claim about data the model has not seen is more specific than a general promise of transfer. A scientist can upload a single-cell RNA-seq snapshot from a few dozen cells. The post says that activity is compressed into a cell-state representation and used to condition the model. On unseen data from a COVID-19 mRNA vaccine study, it says, the model distinguished vaccine-activated immune T cells from resting ones, and a variant linked to severe COVID-19 disrupted gene activity more strongly in the activated state than in the resting one. If the input is a single measurement, such as gene activity, the post says the model can predict an unmeasured property, such as chromatin accessibility, which is how it proposes experiments rather than only scoring variants.
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.
— Sarah Teichmann, University of Cambridge, in DeepMind’s post
A second outside quote in the post is about drug programs, not about the TenK10K percentages. Peter Donnelly, chief executive of Genomics Limited, says genome-wide association studies are hard to turn into what a drug program needs, and that the lab’s model is an advance his company expects to put into its Mystra platform. Hart’s interviews make the same point in the lab’s voice. Gayoso told him the breakthrough is “tying human genetics to cell biology.” Avsec told him that drug work targets specific proteins and specific cell types, which is why the finer grid matters, and that the more detailed predictions are “more expensive” to produce than the earlier model’s. Hart writes that those demands would have made AlphaGenome 2 impractical for the atlas of every possible DNA letter change that DeepMind released in September. The Wednesday post says the lab will soon release AlphaGenome 2 skills in Google Antigravity. It does not say the new model replaces the one that built that atlas.


