The average was the problem
The same DNA variant can have different regulatory effects across cell types and states. A variant that changes expression in gut macrophages may be obscured in a tissue-level prediction. The first AlphaGenome, released last year, predicted variant effects across broad tissues in parallel from a one-million-letter sequence. That made it efficient, and it also spread the model thin across closely related cell types, in DeepMind's own account.
AlphaGenome 2 drops the parallel design. It is a conditional model: one prediction takes one condition, a tissue, a cell type or a state, fed into the architecture at every stage alongside the sequence. The condition can come from a reference catalogue, or from a researcher's own single-cell experiment, compressed by a cell encoder into a state representation, with no retraining. DeepMind says a snapshot from a few dozen cells is enough. The same separation of state from output lets the model fill in unmeasured readouts: give it gene expression alone and it will predict chromatin accessibility, which helps a lab decide which experiment to run next. It also adds DNA methylation to the original model's readouts.
One condition per prediction also means a panel of hundreds of conditions takes hundreds of passes, where the old model produced its tissue panel in one. What that costs in practice — batching, latency, total compute across a large panel — the material checked does not establish.
Read the doubling carefully
The headline number is real, and it needs its denominator. On TenK10K, a single-cell map of 28 primary immune cell types, at the reported 90%-precision operating point AlphaGenome 2 recovered 30% of the benchmark's nominated putatively causal expression-linked variants. The fine-tuned first model recovered 16%. Direction of effect, whether a variant turns a gene up or down, came in at 43% against 23%, measured in the 90%-accuracy regime. On genome regions held out of training, the paper reports the model explained a median 34% of cell-type-specific expression variance within a tissue, against 23% for the predecessor.
Two qualifications sit underneath those figures. First, 30% recall at high precision still means roughly seven in ten of the nominated variants were not recovered at that threshold; the recovered set is a minority of the nominated one, which is also the paper's own framing. Second, the comparisons are against DeepMind's own earlier model, fine-tuned, and other tools DeepMind selected, on a benchmark DeepMind ran. That is normal for a launch paper, and it is also why the numbers should be read as DeepMind's measurements until independent groups rerun them.
The worked examples illustrate the cell-state specificity behind the aggregate results, because both check the model against results established outside it. In the inflammatory bowel disease example, the model's predictions linked rs2836882 to ETS2 approximately 290 kilobases away and localized the effect to monocytes and macrophages, consistent with the cited experimental study (DeepMind's blog describes the effect in gut macrophages). In data from a COVID-19 mRNA vaccine study the model had not seen, it separated vaccine-activated T cells from resting ones and predicted a stronger effect for a severe-COVID-associated variant in the activated state — a distinction that tissue-level predictions may obscure.
What a researcher actually gets
Access, as of November 11, 2026, comes in two tiers. Non-commercial researchers get AlphaGenome 2 through the AlphaGenome API today, with expression-conditioning Colab tutorials available alongside it. Commercial research is in private preview, with no public pricing. DeepMind also precomputed two resources for groups that do not want to run the model at scale themselves: enhancer–gene links across 640 primary single-cell reference conditions covering 39,323 protein-coding and long non-coding RNA genes, and scores for more than 6.8 million unique common and trait-associated variants across 510 human conditions. AlphaGenome 2 Skills for Google Antigravity are promised, not yet released.
The limits are stated in the paper, and they are not boilerplate. The paper identifies limits arising from observational, steady-state profiles — training data that do not themselves establish intervention outcomes — and uncertainty about predicting responses to gene knockouts, overexpression or dynamic state transitions. Transient, stimulated and disease states are underrepresented in training. Both AlphaGenome versions are restricted to research use and have not been validated or approved for clinical or diagnostic purposes.
Sources
- [1] Google DeepMind, "AlphaGenome 2: Deciphering how cells read DNA to better understand disease," November 11, 2026Read source
- [2] AlphaGenome 2 technical paper, November 11, 2026Read source
- [3] Unite.AI, "Google DeepMind's AlphaGenome 2 Predicts Cell-State Variant Effects," November 11, 2026Read source
- [4] SiliconANGLE, "Google DeepMind releases new advanced robotics AI model and next-generation gene prediction model," November 11, 2026Read source