AlphaGenome 2: DeepMind’s Cell-Level AI Maps How DNA Variants Drive Disease

Google DeepMind has unveiled AlphaGenome 2, a next-generation genomic AI model that predicts how variations in DNA affect individual cell types and cell states rather than averaging signals across whole tissues. The model, a successor to the AlphaGenome system released last year, is positioned as a research tool that could sharpen the search for the cellular mechanisms behind complex diseases.

• Cell-level resolution: AlphaGenome 2 conditions its predictions on a specific tissue, cell type or cell state, capturing effects that tissue-wide averages obscure.
• Measurable gains: In a benchmark of 28 immune cell types, the model identified 30% of causal variants versus 16% for its predecessor, and 43% versus 23% for directional effects on genes.
• Scale of training: The system was trained on more than 107 million single-cell profiles drawn from 68 tissues, including data from the Human Cell Atlas.
• Access model: AlphaGenome 2 is available for non-commercial research through the AlphaGenome API, with commercial research in a private preliminary phase.

A conditional architecture built for cellular context

The core premise is simple to state and hard to model: nearly every cell in the body carries the same DNA, but a heart muscle cell, a skin cell and an immune cell fighting an infection switch different genes on and off. Those patterns shift with what the cell is doing, so the same genetic variant can produce very different effects depending on cell type and state.

The original AlphaGenome ingested a DNA sequence of one million letters and produced predictions across thousands of experiments and many tissues simultaneously. That approach was efficient but spread the model’s capacity across too many tasks, making subtle differences between closely related cell types harder to detect.

AlphaGenome 2 redesigns that pipeline. Instead of parallel predictions, it is a conditional model that predicts the effects of DNA for one biological condition at a time, defined by a tissue, cell type or cell state. The target condition is fed into every stage of the architecture alongside the million-letter DNA sequence, allowing the model to learn how the same stretch of DNA is interpreted differently depending on cellular context.

The model can also accept representations generated on the fly from a researcher’s own single-cell experiments through a cell encoder, letting it analyze conditions absent from its reference catalogue without full retraining. Training combined human and mouse samples, and the new version adds DNA methylation estimates to the gene expression, histone modification and chromatin accessibility predictions already offered by its predecessor.

Benchmarks and a vaccine-driven test case

DeepMind reports that on the TenK10K dataset, a map of 28 primary immune cell types, AlphaGenome 2 reliably identified 30% of causal variants, compared with 16% for the previous model tuned with single-cell data and other specialized tools. For directional effects on genes, the comparison was 43% versus 23%. The company argues these results can help prioritize candidate variants that other analyses would miss.

The evaluation also probed whether the model could track continuous shifts in gene expression as cells mature, such as blood stem cells advancing toward red blood cell or immune lineages. In genomic regions excluded from training data, DeepMind reported improved ability to explain differences between cell types within the same tissue.

A particularly telling test used data from a study of an mRNA COVID-19 vaccine. AlphaGenome 2 distinguished T cells activated by the vaccine from those remaining at rest, and found that a genetic variant linked to severe COVID-19 altered gene activity more strongly when cells were activated. Analyzing resting cells alone would not have surfaced that behavior in the same way.

“This new version identifies nearly twice as many variants that modify gene expression in cells compared with the previous model,” the company said in a statement, adding that the tool gives scientists a powerful means to uncover the cellular mechanisms behind disease.

Sarah Teichmann, co-founder of the Human Cell Atlas consortium and professor of stem cell medicine at the University of Cambridge, highlighted the model’s ability to adapt to new data without retraining.

What it means for disease research

Much of the difficulty in genomics lies in the gap between knowing a variant exists and understanding where and how it acts. Many disease-associated variants operate in specific cell types, and their effects often vanish when signals are averaged across whole tissues. Cell-level precision could help researchers pinpoint causal factors in conditions such as Alzheimer’s disease, type 2 diabetes, inflammatory bowel disease and rare disorders.

DeepMind frames the work as complementary to global efforts such as the Human Cell Atlas, turning cellular atlases from descriptive maps into predictive resources. The model can also attempt multimodal imputation: if a team supplies single-cell gene activity data, AlphaGenome 2 may infer unmeasured properties such as chromatin accessibility, helping guide which experiments to run next.

Important caveats remain. AlphaGenome 2 is intended exclusively for research and has not been validated or approved for clinical or diagnostic use, and more detailed predictions carry higher computational cost. The reported figures are evaluation results rather than clinical outcomes.

Still, the significance lies in resolution rather than hype. AlphaGenome 2 does not cure anything, but it sharpens the map that researchers use to navigate from variant to mechanism. If the community validates its predictions against real experiments, it becomes scientific infrastructure: the kind of quiet advance that, accumulated over time, changes the pace at which medicine understands disease.