On 11 November Google DeepMind released AlphaGenome 2, a conditional model that predicts how DNA variants affect gene regulation in one tissue, cell type or cell state at a time, trained on more than 107 million cell profiles from 68 tissues. On the TenK10K immune benchmark it identified 30% of causal variants (16% before) and recovered effect direction for 43% of variants (23% before). It reproduced a known IBD mechanism involving ETS2. Access is via API for non-commercial research; DeepMind says it is not validated for clinical use. No independent replication was found.
1. What Was Announced
On 11 November, Google DeepMind published AlphaGenome 2, a model that predicts how DNA variants alter gene regulation in specific tissues, cell types and cell states. [Established — Google DeepMind, blog post, 11 November 2026. Some outlets date the announcement 12 November; the DeepMind post itself carries 11 November.] The change from its predecessor is architectural. DeepMind says that instead of predicting many outputs in parallel, AlphaGenome 2 “is a conditional model that predicts DNA effects for one condition — defined as a tissue, cell type or state — at a time.” [Established — Google DeepMind, 11 November 2026.]
The original AlphaGenome, published in Nature, asked what a variant does on average across tissues. The new model asks what it does in this cell in this state. [Established — Nature, AlphaGenome paper; DeepMind blog 11 November 2026 for the contrast.] That matters because most disease-associated variants sit in non-coding DNA, and their effect depends on which cell reads them.
2. The Numbers, and What They Measure
DeepMind reports training on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. [Established — Google DeepMind, 11 November 2026. Dataconomy, 12 November 2026, gives “more than 100 million”; the lab’s own figure is used.] On TenK10K, a single-cell immune dataset covering 28 cell types, AlphaGenome 2 identified 30% of causal variants against 16% for its predecessor, and recovered the direction of the gene effect for 43% of variants against 23%. [Established — Google DeepMind, 11 November 2026. Secondary coverage describes the first figure as recall at 90% precision; the Leadsman could not confirm that framing from the DeepMind post alone.] DeepMind’s headline is that the model finds “nearly twice as many variants that change gene expression in cells.”
Read the numbers carefully. A doubling from 16% to 30% is a large relative gain, and an absolute recall of 30% still means that most variants in the benchmark’s causal set are not recovered. The label “causal” also deserves a pause: in statistical genetics, causal variants in a dataset like TenK10K are typically inferred by fine-mapping rather than demonstrated experimentally, so the benchmark measures agreement with a statistical inference. [Assessed with moderate confidence — general characterisation of how expression-linked causal variants are labelled; the Leadsman has not read the evaluation methods.] The same caution applies to the effect-direction figure, which rose from 23% to 43%: the sources reviewed describe it as a share of variants for which the direction was recovered, not as the model’s overall hit rate, and the exact threshold is given only in secondary coverage. [Assessed — the Leadsman could not confirm the metric definition from the DeepMind post alone.]
3. The Worked Examples
Two examples carry the argument. In inflammatory bowel disease, DeepMind says the model reproduced a landmark discovery that a variant drives the disease by over-activating the ETS2 gene in gut immune cells. [Established — Google DeepMind, 11 November 2026.] Secondary coverage specifies the variant as rs2836882, lying about 290 kilobases from ETS2, with the predicted effect localised to gut monocytes and macrophages. [Established as reported — Unite.ai and Implicator, via search results, 11–12 November 2026; not independently checked against the DeepMind post.] In COVID-19, the model suggested that a variant linked to severe disease disrupts gene activity more strongly in vaccine-activated T cells. [Established — Google DeepMind, 11 November 2026.]
Note the nature of these. The IBD result reproduces a known finding, which is the right test of a tool but not evidence of discovery. The COVID result is a prediction about activated cells that, in DeepMind’s telling, the model can make for states it was not retrained on. That is the ability that would, if it holds up, matter most to working scientists, because it lets a lab upload its own single-cell experiment and ask about variants in it. [Assessed — analytical reading; DeepMind says the model can condition on a researcher’s own single-cell data; the Leadsman has not seen independent tests of this.]
4. Access, Limits and the Gap Between Benchmark and Bench
The model is “available today for non-commercial research via the AlphaGenome API, and in private preview for commercial research.” DeepMind adds that AlphaGenome and AlphaGenome 2 “are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.” [Established — Google DeepMind, 11 November 2026.] Dataconomy adds that the new model has higher compute requirements and will not immediately replace the original. [Established — Dataconomy, 12 November 2026; Tier 3 on this specific point.]
Two structural points follow. First, the results available at the time of writing are the developer’s own, on its chosen benchmarks. The Leadsman found no independent replication, and a model released through an API rather than as open weights is harder to audit than one that can be run by third parties. [Assessed with high confidence — no independent evaluation was found in the sources reviewed; the API-access point follows from the stated terms.] Second, a prediction about a variant’s effect in a cell type is a hypothesis that still needs a wet-lab experiment, and the gap between a model’s predicted rank and a validated mechanism is where most such tools have historically lost value. [Assessed.]
5. Both Readings
The optimistic case. Genome-wide association studies have produced thousands of disease-linked regions in non-coding DNA, and the bottleneck has been identifying which variant, acting in which cell, does the work. A model that narrows a seven-variant credible set to one, and names the cell type, converts years of follow-up into a ranked shortlist. If the approach scales, it changes how target-discovery teams prioritise experiments. [Assessed — the strongest form of the case for the tool, consistent with DeepMind’s IBD example.]
The sceptical case. A tool that recovers 30% of benchmark causal variants, and 43% of effect directions, is a prioritisation aid, not an oracle; benchmark labels are statistical inferences; the model is not validated clinically; and the strongest evidence is the developer’s. The gain from one version to the next is real, but a larger number on an internal benchmark does not by itself tell a clinical geneticist whether to trust a prediction about a particular patient’s variant. [Assessed — the strongest form of the sceptical case.]
Bottom line: AlphaGenome 2 is a credible step from tissue-average to cell-specific variant prediction, documented by a developer with a record in this field, and available to academics now. What is established is the architecture, the reported benchmark gains and the access terms. What is not yet established is independent replication, performance on states the model was not trained on, and any clinical utility, which DeepMind itself disclaims. The Navigator’s test of the next six months is simple: whether laboratories outside DeepMind publish confirmations of predictions the model made for variants no one had yet validated.