Glonce

DeepMind launches AlphaGenome 2 to predict DNA variant effects by cell type

modelsThe Glonce editorsGoogle DeepMind🔥75Glonce rating75Impact70Novelty75Relevance95Surprise45Players85Credibility85
DeepMind launches AlphaGenome 2 to predict DNA variant effects by cell type

Google DeepMind has announced AlphaGenome 2, a DNA sequence model that predicts how DNA sequences and their variants affect specific human cell types. The pitch is that every cell carries the same DNA, yet a heart muscle cell or an immune cell fighting an infection switches different genes on and off. Predicting how subtle DNA variations change that gene expression in cells is, in DeepMind's words, still an unsolved challenge, and it is a key driver of human traits and disease.

The first AlphaGenome, launched last year, predicted how variants affect molecular activity across broad tissues. DeepMind later added AlphaGenome Atlas, a platform with predictions for all 9 billion possible single-letter changes in the human genome. AlphaGenome 2 zooms in from tissue averages to individual cell types and cell states, and DeepMind says the redesigned architecture is a step-change improvement in variant predictive performance. Its headline claim: AlphaGenome 2 identifies nearly twice as many variants that change gene expression in cells as DeepMind's previous state-of-the-art model.

The design change is the core of it. The original model took a million-letter DNA sequence and predicted thousands of experiments from many tissues in parallel, which spread its capacity thin and made it hard to separate closely related cell types. AlphaGenome 2 is instead a conditional model: it predicts effects for one condition (a tissue, cell type or state) at a time, and feeds the target condition in at every stage of its architecture alongside the million-letter sequence. This lets it learn how the same stretch of DNA is read differently across cell types. It also enables cross-modality imputation: give it one measurement, such as single-cell gene activity, and it can predict unmeasured properties such as chromatin accessibility, which DeepMind says can help researchers choose which experiments to run next.

The model was trained on bulk and single-cell data from human and mouse samples. The intro cites more than 100 million individual cell profiles across 68 tissues; the methods section says over 107 million. It predicts the same core readouts as AlphaGenome (gene expression, histone modifications, chromatin accessibility) and adds DNA methylation, which DeepMind calls a widely requested readout.

DeepMind reports three kinds of results. First, on genome regions outside its training data, the model explains significantly more of how gene expression differs across cell types within a tissue, including continuous changes such as blood stem cells maturing into red blood cells and immune lineages. Second, on the TenK10K dataset, a large single-cell map of 28 primary immune cell types, it reliably identified nearly twice as many causal variants (30% vs. 16%) and directional gene effects (43% vs. 23%) as the original AlphaGenome fine-tuned on single-cell data and other specialized tools. Third, it reproduced a recent landmark finding that a variant in a distant regulatory switch drives inflammatory bowel disease (IBD) by overactivating the ETS2 gene in gut macrophages: the model singled out the causal variant and showed it selectively boosts ETS2 in gut macrophages, leaving other cell types and genes largely unaffected.

AlphaGenome 2 can also handle new experiments without retraining. A scientist uploads a single-cell RNA-seq snapshot from just a few dozen cells; the gene activity is compressed into a cell state representation that conditions the model on unseen states. DeepMind says this can match or outperform models trained or fine-tuned directly on those datasets. In an example using unseen single-cell data from a COVID-19 mRNA vaccine study, the model distinguished 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 unactivated cells.

To spare researchers heavy computation, DeepMind precomputed AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions. It says this addresses two problems. One is mapping enhancer-gene links in real human cell types: CRISPR screens mostly use lab-grown cell lines, so such maps are often missing for cells in the body, and the model extends the mapping to primary human cell types such as brain astrocytes. The other is pinpointing causal variants from genome-wide association studies (GWAS) and the cells where they act. Across a number of complex traits, from height to type 2 diabetes, the precomputed scores help identify the variants and cell types most likely involved, and when confident the model also accurately predicts whether a variant turns a gene's activity up or down.

AlphaGenome 2 and interactive Colab tutorials for expression conditioning are available today for non-commercial research through the AlphaGenome API, and in private preview for commercial research. DeepMind says AlphaGenome 2 Skills will arrive soon in Google Antigravity, its agentic development platform. It is working with collaborators including researchers at the Wellcome Sanger Institute to profile more cell types, developmental lineages and perturbed states. It invites use cases through a form by 30 November, and selected projects may get extra quota and credits. The model is for research only and has not been validated or approved for clinical or diagnostic purposes.

Key facts

  • Google DeepMind's AlphaGenome 2 is a conditional DNA model: it predicts variant effects for one tissue, cell type or state at a time, taking the target condition as input at every stage alongside a million-letter DNA sequence.
  • DeepMind says it identifies nearly twice as many gene-expression-changing variants as its previous state-of-the-art model; on the TenK10K immune-cell dataset it reports 30% vs. 16% causal variants and 43% vs. 23% directional gene effects against the original AlphaGenome fine-tuned on single-cell data and other specialized tools.
  • It was trained on bulk and single-cell human and mouse data (more than 100 million cell profiles across 68 tissues in the intro, over 107 million in the methods) and adds DNA methylation to the earlier readouts.
  • Given a few dozen cells of single-cell RNA-seq, it can condition on unseen states without retraining; examples include reproducing an IBD finding involving ETS2 in gut macrophages and a COVID-19 vaccine T-cell result.
  • Available now for non-commercial research via the AlphaGenome API and in private preview for commercial research; research use only, not validated for clinical or diagnostic use.

Why it matters

Many disease-linked variants sit in distant regulatory switches called enhancers, and some act only in particular cells, so tissue-level averages blur the signal. AlphaGenome 2 moves from broad tissues to individual cell types and states, which is where DeepMind says disease-relevant effects hide. Its headline gain is nearly twice as many expression-changing variants found compared with DeepMind's own previous best model. It also links human genetics (GWAS hits) to single-cell atlases such as the Human Cell Atlas, which DeepMind describes as mapping gene expression for millions of cells.

Who it affects

The audience is researchers in human genetics, single-cell biology and disease biology, in academia and, through the private preview, commercial research. Teams studying immune diseases, complex traits such as height or type 2 diabetes, and cell states created by disease or vaccination get the most direct examples. DeepMind also says it is collaborating with researchers at the Wellcome Sanger Institute.

How to use it

AlphaGenome 2 is available today for non-commercial research via the AlphaGenome API, with interactive Colab tutorials for expression conditioning. Commercial research use is in private preview. A scientist can upload a single-cell RNA-seq snapshot from a few dozen cells to condition the model on a new cell state, or use DeepMind's precomputed predictions for trait-associated variants across hundreds of human cell and tissue conditions. AlphaGenome 2 Skills in Google Antigravity are promised soon, with no date. DeepMind is collecting community use cases through a form by 30 November, and selected projects may be eligible for extra quota and credits. No API pricing is given.

How solid is it

This is a first-party announcement, and every performance claim is DeepMind's own. The 'nearly twice as many variants' comparison is against DeepMind's previous model, and no independent evaluation is mentioned. The text names no paper, preprint or DOI. The percentages (30% vs. 16%, 43% vs. 23%) are given without a defined metric or denominator. The ETS2 and COVID-19 vaccine cases are described as computational reproductions or examples, not new experimental validation. The training-data figure also appears as 'more than 100 million' in one place and 'over 107 million' in another.

Risks and caveats

DeepMind states that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The precomputed scores flag the variants and cell types most likely involved, and the direction of effect is described as accurate when the model is confident, which is a narrower claim than accurate everywhere. Commercial access is limited to a private preview, no commercial partners are named, and the Antigravity release has no date beyond 'soon'.

“AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.”

— Google DeepMind, AlphaGenome 2 announcement