Google DeepMind's AlphaGenome 2 Moves Variant Prediction to Single Cell Types

Google DeepMind released AlphaGenome 2 on November 11, 2026, a model the company says can predict how a specific DNA change behaves inside a particular cell type or cell state instead of averaging that effect across a tissue sample. The announcement came from the lab's London team and was covered the same day by Nature, The Verge and MarkTechPost, with The Decoder following on November 12. The system is a direct successor to AlphaGenome, and DeepMind frames it as a move from reading DNA in bulk toward reading it the way individual cells actually do.

The central difficulty in genomics is not reading DNA but interpreting it. Around 98 percent of the human genome does not code for proteins, according to Nature, and that noncoding majority contains the regulatory switches that decide when, where and how strongly genes are switched on. Variants in those regions have been tied to disease by large population studies, yet working out which variant actually matters, and in which tissue, has remained slow, costly and often inconclusive.

Every cell in the body carries the same DNA blueprint, but heart muscle cells, skin cells and immune cells switch on different genes, and the pattern shifts with whatever the cell is doing at that moment. The same variant can therefore have sharply different consequences depending on cell type and cell state, a distinction that bulk tissue measurements blur. DeepMind first described AlphaGenome in 2024, and Nature describes AlphaGenome 2 as an upgrade to the 2025 version of the model.

Two months before the release, DeepMind published AlphaGenome Atlas, a precomputed map covering all 9 billion single letter changes in the human genome. That atlas was built on AlphaGenome 1, and James Lee, an immunologist and geneticist at the Francis Crick Institute, told Nature it was of limited use to his team because it could not say which cell type a variant acts in. AlphaGenome 2 was built to answer exactly that question.

Key Facts

AlphaGenome 2 is a conditional sequence to function model. It reads a window of 1 million letters of human or mouse DNA, specifically 1,048,576 base pairs, together with a biological condition such as a tissue, a cell type or a cell state. The condition is a 256 dimensional embedding injected at every stage of the network through FiLM layers, so the same stretch of DNA is interpreted differently depending on which cell is reading it. The architecture is U-Net style, with a convolutional encoder, 11 transformer blocks and a decoder. MarkTechPost reported on November 11 that conditioning only the decoder or the output layer produced minimal gains, while conditioning the encoder and the transformer blocks delivered most of the improvement.

The training set is the decisive change. DeepMind used human and mouse bulk and single cell data spanning 68 tissues and more than 107 million single cell profiles, where the first AlphaGenome relied on bulk samples of mixed cells. The single cell expression encoder was trained mostly on CELLxGENE data covering more than 100 million human single cell RNA sequencing profiles, and the full model was trained in four stages on 2,048 TPU v4 chips, including a distillation step that compressed eight teacher models into one student. DeepMind said the model identifies close to twice as many variants that alter gene expression as the previous state of the art.

On TenK10K, a benchmark containing 28 primary immune cell types, AlphaGenome 2 identified 30 percent of causal variants at 90 percent precision, compared with 16 percent for AlphaGenome 1 fine tuned on single cell data. It also got the direction of the effect right, whether a variant raises or lowers gene activity, in 43 percent of cases, against 23 percent for the earlier model.

Other benchmarks moved in the same direction. Across GTEx tissues, the median Pearson correlation for expression change rose from 0.61 to 0.71, and the share of cell type variance explained in within tissue single cell data rose from 23 percent to 34 percent. On held out ENCODE-rE2G CRISPRi data, enhancer to gene AUPRC reached 0.78, against 0.73 for AlphaGenome 1 and 0.61 for a four model Borzoi ensemble, with the widest gap at enhancers between 100 and 2,500 kilobases away, where the scores were 0.62 and 0.48. Assay imputation beat the Avocado baseline on 166 of 181 held out tracks, and on an unseen Parkinson cohort the model recalled 41 percent of variants, against 17 percent for the cohort tuned Scooby system.

The model was trained on 637 human cell types and 311 mouse cell types, but it is not confined to them, since researchers can supply new single cell activity data and receive predictions for states the model never saw during training. DeepMind also published precomputed atlases: enhancer to gene links for 640 primary cell conditions, and scores for about 7 million common and GWAS variants across 510 conditions. The API accepts up to 256 cells per call.

Analysis

Nature reported on November 11 that the DeepMind team reproduced a 2024 discovery about inflammatory bowel disease without running a single experiment. The original work, led by James Lee at the Francis Crick Institute, found that the gene ETS2 becomes overactive in immune cells because of a variant sitting in a gene desert far from any protein coding sequence. Lee said the model produced the answer in 30 minutes, while his team had spent a year and a half on experiments, and he called that outcome reassuring rather than threatening.

That result is the strongest evidence yet that the conditioning approach, not raw scale alone, is what changed. The earlier design asked one network to output thousands of experimental tracks across many tissues at once, which DeepMind says diluted model capacity and made subtle differences between closely related cell types hard to capture. Moving the cell identity into the input, and forcing the network to modulate its internal activations around that identity, turns a static atlas into something closer to a prediction engine.

What this really means is that the bottleneck in genomics AI is shifting from prediction accuracy toward experimental design. If a model can infer chromatin accessibility from gene activity alone, and can flag where existing measurements leave gaps, then the scarce resource becomes the next experiment rather than the next dataset. The Verge reported on November 11 that Ziga Avsec, who leads genomics at DeepMind, argued this specificity matters for medicine because when you make a drug you target a specific protein in a specific cell type.

The caveats are real. The Verge reported that Avsec described AlphaGenome 2 as more expensive to run than its predecessor, which is why DeepMind did not use it to regenerate the whole genome atlas released in September. The model is offered for research only and has not been validated or approved for clinical diagnosis, and its weights will only be released when the underlying paper is published.

Why It Matters

Most drugs fail for reasons that trace back to cell biology: a target sits in the wrong tissue, or a pathway behaves differently in the cell that actually drives the disease. A model that resolves variant effects by cell type gives drug hunters a way to check, before committing years of laboratory work, whether a candidate mechanism operates in the relevant cells. DeepMind has already precomputed predictions for hundreds of cell and tissue conditions to support AI driven drug discovery and the interpretation of GWAS results.

The immune system example is concrete. The Decoder reported on November 12 that when the model was applied to COVID-19 mRNA vaccine data, it distinguished vaccine activated T cells from resting T cells, and showed that variants linked to severe COVID-19 disrupt gene activity more strongly in the activated state. That is the kind of state dependent effect that bulk averages tend to erase.

For population genetics, the practical payoff is narrowing. A GWAS hit can implicate dozens of variants across a region without saying which one is causal or which cell type is affected. The new atlases score roughly 7 million variants across 510 conditions, giving researchers a way to prioritize candidates such as brain astrocytes or particular immune populations before any wet lab work begins.

Next Up

AlphaGenome 2 is available from November 11 through the AlphaGenome API for noncommercial research, with commercial use limited to a private preview. DeepMind said it plans to add dedicated features to its agent development platform, Google Antigravity, soon, and that future versions may be trained on data spanning multiple species, with a longer term role in virtual cell programs.

Publication of the paper will bring the model weights, and independent groups will then test whether the benchmark gains hold on their own data. Until then, the company's own framing sets the boundary: this is a research instrument, not a diagnostic, and its most immediate contribution may be telling scientists which experiment to run next rather than replacing the experiment altogether.