Google DeepMind Unveils AlphaGenome 2 to Read How DNA Variants Act in Specific Cell Types

Google DeepMind has announced AlphaGenome 2, a next-generation model that predicts how DNA sequences and their variants affect specific human cell types. It drops the tissue-level averages of the original model for individual cell types and cell states, which DeepMind says gives scientists a sharper tool for finding the cellular drivers of disease. The model is available for non-commercial research through the AlphaGenome API and in private preview for commercial research.

From tissue averages to single cells

Every cell in the body carries the same DNA, but heart muscle cells and infection-fighting immune cells switch different genes on and off. Efforts such as the Human Cell Atlas are mapping gene expression across millions of cells. Predicting how subtle DNA variants change that expression within cells, a key driver of traits and disease, is still unsolved, DeepMind said.

DeepMind launched AlphaGenome last year and extended it with AlphaGenome Atlas, which holds predictions for all 9 billion possible single-letter changes in the human genome. AlphaGenome 2 redesigns the architecture. DeepMind says it finds nearly twice as many variants that change gene expression in cells as its previous state-of-the-art model. It was trained on more than 100 million individual cell profiles across 68 tissues.

The key change is a conditional design. The first model took a 1-million-letter DNA sequence and predicted thousands of experiments across tissues in parallel. That was efficient but spread the model’s capacity thin, making it hard to capture differences between closely related cell types. AlphaGenome 2 takes one condition at a time, defined as a tissue, cell type or state, and feeds it in alongside the sequence at every stage. The model learns how the same stretch of DNA is read differently across cell types.

The design also allows cross-modality imputation. If a scientist enters only one measurement, such as single-cell gene activity, the model can predict unmeasured properties like chromatin accessibility, which can help prioritize the next experiments. It was trained on bulk and single-cell data from human and mouse samples, over 107 million individual cell profiles in all. It predicts gene expression, histone modifications and chromatin accessibility, and adds DNA methylation, a widely requested readout.

Nearly twice the causal variants in immune cells

On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants and 43% of directional gene effects. The original AlphaGenome fine-tuned on single-cell data and other specialized tools reached 16% and 23%. Outside its training regions, it also explained more of how gene expression differs across cell types within a tissue, including the gradual changes as blood stem cells mature.

DeepMind also showed it reproducing a landmark study on inflammatory bowel disease, in which a variant in a distant regulatory switch overactivates the ETS2 gene in gut macrophages. Using predictions across gut immune cell types, the model singled out the causal variant and showed it boosts ETS2 selectively in gut macrophages, leaving other cell types and genes largely unaffected.

New cell states without retraining

AlphaGenome 2 can also adapt to new single-cell experiments without retraining. A scientist uploads a single-cell RNA-seq snapshot from just a few dozen cells, which is compressed into a cell state representation used to condition the model. DeepMind says this can match or beat models trained or fine-tuned directly on those datasets. On unseen data from a COVID-19 mRNA vaccine study, it separated vaccine-activated T cells from resting ones, and showed that a variant linked to severe COVID-19 disrupts gene activity more strongly in the activated state, an effect that would be missed in resting cells.

DeepMind precomputed predictions for trait-associated variants from genome-wide association studies across hundreds of human cell and tissue conditions, so researchers need not run large computations themselves. It extends enhancer-gene mapping to primary human cell types such as brain astrocytes. Across traits from height to type 2 diabetes, the scores help narrow down likely causal variants and the cells where they act. When confident, the model also predicts whether a variant turns a gene up or down.

Reaction and what comes next

Sarah Teichmann, co-founder of the Human Cell Atlas consortium and professor at the University of Cambridge, said what makes AlphaGenome 2 exciting is that it adapts directly to new single-cell data without retraining. Peter Donnelly, CEO of Genomics Limited, said human genetics is one of the most powerful tools in drug discovery, but moving from an association signal to what a drug program needs is hard. He said sequence prediction models help bridge that gap and that the company looks forward to adding AlphaGenome 2 to its Mystra platform.

DeepMind plans to profile more human cell types, developmental lineages and perturbed states with collaborators including the Wellcome Sanger Institute. It will also release AlphaGenome 2 Skills in Google Antigravity, its agentic development platform, pairing the model with agentic science tools such as Gemini. The API is open for research use.