Google DeepMind releases DNA AI that predicts genetic effects in specific cells

The same stretch of DNA can behave differently across cells. Google DeepMind released AlphaGenome 2 on November 11, 2026, to predict those differences more precisely. The research model targets individual cell types and states, with company-reported results showing nearly twice the causal variant detection of its predecessor on an immune-cell benchmark.

One DNA sequence, different cellular answers

The original AlphaGenome took a million-letter DNA sequence and predicted molecular properties across thousands of experiments and tissues at once. DeepMind says that design spread the model’s capacity thin, making subtle differences between closely related cells harder to capture. Broad tissue averages could also obscure variants whose effects appear only in a particular cell type.

AlphaGenome 2 instead pairs that DNA sequence with one biological condition: a tissue, cell type or cell state. It feeds that condition into every stage of the model. Researchers can select a condition from a reference catalog or supply their own single-cell experiment. The change lets the model predict how the same DNA is read differently across cells.

Training combined bulk measurements with single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. DeepMind says the model can also estimate missing experimental measurements. Given gene activity alone, for example, it can predict chromatin accessibility—another readout of DNA function—to help researchers choose experiments.

A stronger result on a specific immune-cell test

On TenK10K, a dataset covering 28 primary immune cell types, DeepMind reports that AlphaGenome 2 identified 30% of causal variants, versus 16% for the original model fine-tuned on single-cell data. It also reports a rise from 23% to 43% for identifying directional gene effects—whether a variant increases or decreases gene activity. These are company-reported benchmark results, not clinical detection rates.

The underlying TenK10K study, published in Nature on October 28, helps explain why cell specificity matters. Its first phase analyzed about 5.4 million cells from 1,925 people of European ancestry. About half the genetic effects on gene expression appeared in only one or a few cell types. That study supplies biological context for the benchmark; it does not independently verify AlphaGenome 2’s performance.

DeepMind also tested a known inflammatory bowel disease mechanism. It says AlphaGenome 2 computationally reproduced a finding that a distant DNA regulatory switch increases activity of the ETS2 gene in gut immune cells called macrophages. The model singled out the causal variant and its selective effect, leaving other cell types and genes largely unaffected. This reproduced an existing finding rather than discovering a new disease mechanism.

New cell states without another training run

Another difference concerns experiments the model has never seen. DeepMind says researchers can upload a snapshot of gene activity from a few dozen cells. A cell encoder compresses that information into a representation of the cells’ state, which then guides predictions without retraining. The company says this approach can match or outperform models trained or fine-tuned directly on those datasets.

In one demonstration using unseen data from a COVID-19 mRNA vaccine study, the model distinguished activated immune T cells from resting ones. DeepMind says it also predicted that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state. Looking only at unactivated cells would miss that state-dependent effect.

For researchers following up disease associations, DeepMind has precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions. Those scores are intended to help prioritize the variants, target genes and cell types most likely to matter, without requiring researchers to run large-scale computations themselves.

AlphaGenome 2 is available through the AlphaGenome API for non-commercial research, with interactive Colab tutorials. Commercial research access remains in private preview. DeepMind says the model is intended for research only and has not been validated or approved for clinical or diagnostic purposes.