DeepMind sells "nearly twice as many". In the paper, it's 30% vs 16%

In brief

• The change: the model takes a million letters of DNA plus the target cell type or state, and predicts DNA activity in that cell type[S2, p. 2].
• The "twice as many": out of 6,346 cases where a statistical analysis gives a variant over 90% probability of acting on a gene, across 28 blood immune cell types, 30% found versus 16% for the old AlphaGenome adapted to cells, with nine picks out of ten correct [S2, p. 4, 5 and 36].

From the tissue average to the cell type

All your cells carry the same DNA, yet a heart muscle cell does not switch on the same genes as an immune cell. The first AlphaGenome, launched last year, predicted a variant's effect across broad tissues. AlphaGenome 2 works one condition at a time: a tissue, a cell type or a cell state.

The detail that made me put my cup down: for a cell state it has never seen, you just provide the gene activity measured in a few dozen cells; the model compresses it into a fingerprint, then predicts variant effects in that state, with no retraining[S2, p. 2].

Twice as many of what, exactly

The post claims that on TenK10K (28 immune cell types), AlphaGenome 2 finds nearly twice as many causal variants (30% vs. 16%) and effect directions (43% vs. 23%) as the original AlphaGenome fine-tuned on single-cell data and other tools.

The baseline is in the paper [S2, p. 4, 5 and 36]. The reference: 6,346 cases, counted cell type by cell type, where a statistical analysis of TenK10K gives a variant over 90% probability of acting on a gene. An estimate, not a certainty. As many decoys are mixed in (under 1% probability), at a comparable distance from the gene. The model ranks everything from most to least suspicious; you keep the top of the list as long as nine picks out of ten are reference cases, and count those caught: 30% for AlphaGenome 2, 16% for AG1-sc, the first AlphaGenome fine-tuned on the same cell data [S2, p. 3]. Out of 6,346, about 1,900 versus 1,000 (our calculation, rounded).

For the 43% vs. 23%, same 6,346 cases, no decoys: the model has to say whether the variant raises or lowers the gene's activity, and the paper counts the share of right answers, without spelling out the reference used for the direction. Cases are ranked from strongest to weakest predicted effect, and you keep the top as long as nine answers out of ten are right [S2, p. 4 and 36]. These cases sit on chromosomes that were not used to tune the model, which was trained on no TenK10K data [S2, p. 4, 19 and 35]. And the post's "other specialized tools"? On TenK10K, the paper compares only one, Scooby, and its text gives no figure for it at this threshold [S2, p. 4, 5 and 36].

The gut example

According to a recent study, a variant in a distant regulatory switch of the DNA drives inflammatory bowel disease by overactivating the ETS2 gene in macrophages, immune cells of the gut. Out of seven candidate variants, AlphaGenome 2 predicted a substantial effect for only one, linked to ETS2 and confined to monocytes and macrophages. A test after the fact: the answer was already known [S2, p. 11].

107 million or more than 100 million cells?

The post says more than 100 million cell profiles across 68 tissues, then over 107 million. The paper ties the 107 million to the last training stage and the more than 100 million to the encoder (the module that produces a cell's fingerprint), without saying whether they are the same cells [S2, p. 2 and 3].

Who can use it

The API is free for non-commercial research; commercial use goes through a private preview, with no published price. AlphaGenome 2 Skills, ready-made modules for AI agents, are coming "soon" to Google Antigravity, with no date; the weights (the learned parameters, to run the model yourself) and the code, "upon final publication" of the paper[S2, p. 53]. Research use only, neither validated nor approved for clinical or diagnostic use[S2, p. 53].

My verdict: going from tissue to cell type is the right zoom, and the few-dozen-cells trick is clever. The "twice as many", though, is AlphaGenome 2 versus the first AlphaGenome, in a match set up and scored by DeepMind. For the rest: no commercial price, no weights yet, not one date. The shop window is ready; the back room can wait.

Words to remember

• Variant: a spot in the DNA where the letter differs from one person to another.
• Gene expression: how strongly a gene is switched on in a cell.
• Causal variant: the one that produces the effect; here, flagged by a statistical analysis, not proven.
• API: programmatic access to the model.

Source sheet

• Google DeepMind: post "AlphaGenome 2: Deciphering how cells read DNA to better understand disease", by Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec, dated 11/11/2026.
• Google DeepMind: paper "AlphaGenome2 predicts regulatory activity and variant effects across cell states", 78 pages, linked from the post.
• Google DeepMind: AlphaGenome API documentation.
• On Silicium Rose: AlphaProof Code, Google DeepMind's other announcement this week: four properties of Signal's code proved.
• Links: Google DeepMind post (11/11/2026) · paper (PDF) · API documentation

Source passages

Documents read on 11/11/2026 between 15:12 and 15:16 UTC. Post: text in a single block, the position is the index of the passage's first character. Paper: PDF page.

Show the 3 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4↑5↑6+6	"November 11, 2026" ; "Every cell in our body carries the same DNA blueprint, yet each cell type and state—like a heart muscle cell or an immune cell fighting an infection—switches different genes on or off to do its job." ; "Global efforts like the Human Cell Atlas are mapping this gene expression for millions of cells" ; "Last year, we launched AlphaGenome to predict how DNA variants affect molecular activity across broad tissues." ; "AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells, compared to our previous state-of-the-art model." ; "Trained on more than 100 million individual cell profiles across 68 tissues" ; "AlphaGenome 2 is available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research." ; "AlphaGenome 2 is a conditional model that predicts DNA effects for one condition - defined as a tissue, cell type or state - at a time." ; "including over 107 million individual cell profiles" ; "On the TenK10K dataset, a large single-cell map of 28 primary immune cell types, AlphaGenome 2 reliably identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) as the original AlphaGenome fine-tuned on single-cell data and other specialized tools." ; "A recent landmark study showed that a variant in a distant regulatory switch drives IBD by overactivating the ETS2 gene in gut immune cells (macrophages)." ; "Scientists can upload a single-cell RNA-seq snapshot of gene activity from just a few dozen cells" ; "this gene activity is then compressed into a cell state representation, before being used to condition the model on unseen states." ; "This enables AlphaGenome 2 to predict variant effects and gene regulation for new cell states" ; "AlphaGenome 2 can also generalize to entirely new single-cell experiments without needing to be retrained." ; "And soon, we will release AlphaGenome 2 Skills in Google Antigravity, our agentic development platform." ; "Please note that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes."	https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/
S2↑	p. 2 : "AlphaGenome2 is a deep learning model that takes as input a 1-Mb DNA sequence, a target condition (tissue, cell line, cell type, or cell state)" ; "post-trains on single cells sampled from 107 million human scRNA-seq profiles" ; "Inferring condition embeddings directly from scRNA-seq extends AlphaGenome2 to newly profiled cohorts and rare cell states without fine-tuning" ; p. 3 : "its single-cell expression encoder was trained on over 100 million individual scRNA-seq profiles across 68 tissues" ; "We constructed a directly comparable baseline (AG1-sc) by fine-tuning AG1 on the single-cell reference conditions in our training corpus using the same genome split" ; p. 4 : "using fine-mapped QTLs on held-out test chromosomes" ; "TenK10K, which mapped variant effects on expression (sc-eQTLs; 28 cell types)" ; "Gains were again largest in the 90% precision and accuracy regime: AG2 increased recall from 16% for AG1-sc to 30% for TenK10K sc-eQTL causality (overall AUPRC 0.76 vs. 0.72 for AG1-sc and 0.69 for Scooby), from 23% to 43% for TenK10K sc-eQTL sign prediction (overall AUC 0.85 vs. 0.79 and 0.77)" ; p. 5 (caption of figure 2e) : "Cell-type-pooled precision–recall (causality, 1:1 negative-to-positive ratio) and sign accuracy–recall (sign, 47.8% positive) curves for 6346 causal sc-eQTLs across 28 cell types, comparing Scooby, AG1-sc, and AG2." ; "Point callouts indicate recall at 90% precision or 90% sign accuracy." ; p. 11 : "show that rs2836882 drives disease risk by upregulating the distal gene ETS2, located approximately 290kb away" ; "To retrospectively test this capability at an experimentally dissected disease locus" ; "Scoring the seven variants in the fine-mapped IBD credible set across primary immune cell types of the ileum, AG2 recovered all three components of this mechanism: it predicted a substantial effect for only rs2836882, linked the variant selectively to ETS2 despite its 290-kb distance, and confined the effect to monocytes and macrophages" ; p. 19 : "separating validation chromosomes used for model selection from held-out test chromosomes used for zero-shot variant effect reporting" ; p. 35 : "AG2 was not trained on data from either study" ; p. 36 : "We evaluated AG2’s ability to discriminate causal regulatory variant–gene associations (PIP > 0.9) from non-causal pairs (PIP < 0.01) within each tissue or cell type." ; "Each causal SNV was matched without replacement to the non-causal control in the same stratum with the closest distance to its own target eGene TSS at a 1:1 ratio" ; "TenK10K: 𝑛 = 12,692 across 6346 positives and 6346 controls" ; "For high-confidence causal variants (PIP > 0.9), we assessed classification accuracy in predicting whether a variant increases or decreases expression of its target eGene" ; "ranking variants according to their predicted effect magnitude, and computing sign accuracy as a function of recall (the fraction of top-ranked variants retained)" ; "Scooby baseline For TenK10K, we compared against Scooby" ; p. 53 : "AlphaGenome2 is available for non-commercial use via an API" ; "AlphaGenome2 and its outputs have not been validated for, and are not approved for, any clinical or diagnostic use. The model source code and weights will be provided upon final publication."	https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf
S3↑1↑2↑3	"AlphaGenome 2 Our latest model. Improved cell-type specificity and generalizes to your own scRNA-seq data." ; "The API is offered free of charge for non-commercial use. Query rates vary with demand."	https://www.alphagenomedocs.com/