AlphaGenome 2 Reads DNA One Cell Type at a Time, and Finds Nearly Twice as Many Variants That Change Gene Activity


The same DNA, read differently in every cell

Every cell in a body carries the same DNA. A heart muscle cell and an immune cell use that blueprint to make different proteins, at different times, for different jobs. That is why a variant can matter in one cell type and do almost nothing in another, and why a model that averages across a tissue can miss the effect that drives a disease.

Google DeepMind announced AlphaGenome 2 on November 11, built to work at that finer resolution. It predicts how DNA sequences and their variations affect specific human cell types and cell states, rather than broad tissues. DeepMind describes it as a tool to uncover the cellular drivers of disease, and that framing is exact. This is a research model for prioritising experiments and interpreting variants. DeepMind states plainly that AlphaGenome and AlphaGenome 2 are intended for research use only and have not been validated or approved for clinical or diagnostic purposes.

The first AlphaGenome, launched last year, predicted molecular activity across broad tissues, and DeepMind later expanded it with AlphaGenome Atlas, which holds predictions for all 9 billion possible single letter changes in the human genome. AlphaGenome 2 changes the question, from what a variant does on average to what it does in this cell, in this state.

One condition at a time, fed into every layer

The redesign is architectural, not cosmetic. The first model took a DNA sequence of one million letters and predicted properties across thousands of experiments from many tissues all at once. DeepMind says that parallel approach spread the model’s capacity thin and made it hard to capture subtle differences between closely related cell types.

AlphaGenome 2 is a conditional model. It takes the same million letter sequence plus one specific biological condition, defined as a tissue, a cell type, or a cell state, and it uses that condition as an input at every stage of its architecture. It predicts for one condition at a time, so the condition can steer the whole prediction instead of competing with hundreds of other outputs.

That condition can come from two places. A researcher can choose it from a reference catalogue of cell types, or generate it on the fly from their own single cell experiment through what DeepMind calls the cell encoder. The model was trained on bulk and single cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues, drawing on global mapping work such as the Human Cell Atlas.

The outputs cover the same core readouts as the first model, including gene expression, histone modifications, and chromatin accessibility, which describe whether a stretch of DNA is open for use. AlphaGenome 2 adds one widely requested readout: DNA methylation, a chemical mark that helps control gene activity. Conditioning also enables cross modality imputation: input one measurement, such as gene activity, and predict unmeasured properties such as chromatin accessibility, which helps a lab decide which experiment to run next.

Nearly twice the variants, and still a minority

The headline result comes from TenK10K, a large single cell map of 28 primary immune cell types. On that dataset, DeepMind reports that AlphaGenome 2 identified 30 percent of causal variants that change gene expression, against 16 percent for the original AlphaGenome fine tuned on single cell data and other specialised tools. For directional gene effects, meaning whether a variant turns a gene’s activity up or down, the figures were 43 percent against 23 percent. DeepMind summarises this as nearly twice as many variants identified as its previous state of the art model.

Read both halves of that result. A near doubling is a real step for researchers ranking thousands of candidate variants. At the same time, 30 percent means most causal variants were still not recovered. This is a better filter, not a finished answer, and any short list still needs experimental validation.

DeepMind also reports better cell type specificity outside its training data, including tracking how gene expression changes as blood stem cells mature. Those are the company’s own results. On release day there is no independent replication of the TenK10K numbers, so hold them as strong, checkable claims until other groups run the model on their own data.

Two disease examples show what cell resolution buys

DeepMind gives two worked examples, and both turn on a variant whose effect only appears in the right cell state.

The first revisits a landmark study of inflammatory bowel disease. A variant in a distant regulatory switch drives the disease by overactivating the ETS2 gene in gut macrophages. Predicting across gut immune cell types, AlphaGenome 2 singled out that variant and showed it selectively boosting ETS2 in gut macrophages, leaving other cell types largely unaffected. A tissue level average could dilute that signal.

The second example tests generalisation to a state the model had not seen. Given single cell data from a COVID-19 mRNA vaccine study, AlphaGenome 2 distinguished vaccine activated immune T cells from resting ones. It then 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.

That second capability may be the most practical part of the release. A researcher can upload a single cell RNA sequencing snapshot from just a few dozen cells and condition the model on that unseen state without retraining it. DeepMind says it can match or outperform models trained directly on those datasets. Disease, treatment, and vaccine responses all differ from a resting reference catalogue, so a model that takes a new state as an input fits how the science actually arrives.

Help for the GWAS problem, precomputed

Genome wide association studies find broad stretches of DNA linked to a trait or disease, but they rarely name the exact causal variant, the gene it controls, and the cell where it acts. That gap is where drug targets are won or lost.

To lower the compute burden, DeepMind has precomputed predictions for trait associated variants across hundreds of human cell and tissue conditions. Two uses stand out:

  • Enhancer to gene links in real cell types. Many disease linked variants sit in enhancers, distant regulatory switches whose target gene is hard to trace. Screens mostly use lab grown cell lines, so wiring diagrams for primary human cells are scarce. AlphaGenome 2 extends that mapping to primary cells such as brain astrocytes.
  • Pinpointing causal variants and their cells. Across traits from height to type 2 diabetes, the scores help identify the variants and cell types most likely involved, and whether a variant raises or lowers a gene’s activity.

Neither use proves a target. Both narrow an expensive search, which is the honest value of a predictive model in a lab budget.

Access, terms, and what to do with it

AlphaGenome 2 and interactive Colab tutorials are available now for non commercial research through the AlphaGenome API. Commercial research access is in private preview, with no public price in the announcement. DeepMind says AlphaGenome 2 Skills will soon be released in Google Antigravity, its agentic development platform, so agent workflows can call the model during an analysis. Researchers can submit use cases by November 30, and selected projects may be eligible for additional quota and credits. DeepMind also names an ongoing collaboration with the Wellcome Sanger Institute to profile more cell types and states.

For teams considering it this week:

  • Start with a known result. Reproduce a variant effect your lab already trusts, such as the ETS2 case, before trusting the model on unknown variants.
  • Record the inputs exactly. Reference genome, variant normalisation, cell annotation, and the exact condition used decide whether a prediction can be compared later.
  • Treat the score as a ranking, not a finding. Budget for CRISPR or other experimental follow up on the top candidates.
  • Keep clinical work out of scope. The model is not validated or approved for diagnosis or clinical decisions.
  • Plan for access limits. Confirm whether your work falls under the non commercial API or the commercial private preview before building a pipeline on it.

The bottom line

AlphaGenome 2 is a step change in resolution, not just in scale. By conditioning a million letter DNA model on one cell type or state at a time, trained on more than 107 million cell profiles across 68 tissues, DeepMind reports nearly twice the causal variant recall of its predecessor, 30 percent against 16 percent on a 28 cell type immune benchmark, plus the ability to predict for a new state from a few dozen cells. The limits are clear: most causal variants in that benchmark are still missed, the numbers are the company’s own until others replicate them, and the tool is for research only. Within those limits, this is AI for science that earns its keep, not by replacing the lab, but by telling a good lab where to look first.