Google DeepMind's AlphaGenome 2 maps how DNA variants behave differently in each cell type

Every cell in the human body carries the same DNA, yet a heart muscle cell and an immune cell fighting an infection behave nothing alike — because each switches different genes on or off depending on what it needs to do. That gap between a DNA blueprint and what actually happens inside a specific cell has long been one of biology's hardest problems. On Wednesday, Google DeepMind announced AlphaGenome 2, an AI model built to close it.

AlphaGenome 2 predicts how genetic variants affect gene regulation in individual cell types and states — not just averaged across broad tissues, as its predecessor AlphaGenome did. Drawing on profiles of more than 100 million individual cells spanning 68 tissues, the model detects close to double the number of expression-altering variants that DeepMind's previous leading model could find, the company says.

When you develop drugs, you're targeting specific proteins and specific cell types.— Žiga Avsec, genomics lead, Google DeepMind

That specificity is the core argument for why cell-level resolution matters. A genetic variant might have a negligible effect in most tissues but a significant one in, say, gut macrophages — the immune cells implicated in inflammatory bowel disease. Tissue-averaged models can miss those signals entirely.

The architectural change that makes this possible is a fundamental redesign. Unlike the original AlphaGenome, which processed a million-letter DNA sequence while simultaneously predicting properties across thousands of experiments — diluting its focus — AlphaGenome 2 works as a conditional model. DeepMind says the system receives both the DNA sequence and a specified biological condition — a particular tissue, cell type, or cell state — at every processing stage, enabling it to capture how identical DNA is interpreted differently depending on the cell.

Tying human genetics to cell biology.— Adam Gayoso, research scientist, Google DeepMind

Gayoso described that connection as the project's central breakthrough. Genome-wide association studies (GWAS) have identified thousands of variants linked to traits and diseases, but connecting those variants to what they actually do inside specific cells has remained a bottleneck. AlphaGenome 2 is designed to let researchers drill down from population-level genetic signals to cellular-level effects, and potentially from there to disease mechanisms.

One notable capability: the model can generalize to entirely new cell types and biological states without being retrained. Researchers can provide a single-cell RNA-seq snapshot of gene activity drawn from as few as several dozen cells, which AlphaGenome 2 condenses into a cell-state representation before generating predictions. According to DeepMind, the model can match or outperform models trained directly on those datasets. As one demonstration, when fed previously unseen single-cell data from a COVID-19 mRNA vaccine study, AlphaGenome 2 told apart vaccine-activated immune T cells from resting ones and found that a variant associated with severe COVID-19 suppresses gene activity more sharply in the activated state — a difference invisible in unactivated cells.

The model also adds DNA methylation to its list of predicted readouts — a widely requested addition beyond the gene expression, histone modifications, and chromatin accessibility that AlphaGenome already covered. And because it is a conditional model, it can perform cross-modality imputation: given one type of measurement, it can predict unmeasured properties, potentially helping researchers decide what experiments to run next.

DeepMind frames AlphaGenome 2 as turning static cell atlases — large data repositories that catalog individual cells in detail, such as the Human Cell Atlas — into a predictive engine for human genetics. To spare researchers the burden of large-scale computation, the company has already run AlphaGenome 2 predictions for trait-associated variants across hundreds of human cell and tissue conditions and made the results available for direct access.

The added resolution comes at a cost. Avsec told The Verge that AlphaGenome 2's predictions are computationally more expensive than those of its predecessor — expensive enough that it would have been impractical to use it to produce the genome-wide AlphaGenome Atlas, which DeepMind released in September and covers every possible single-letter DNA change in the human genome. AlphaGenome is not being retired.

Building such a model has only recently become feasible, Avsec said, because new technologies can now capture individual cells and their states at high resolution and generate the vast datasets needed for training. DeepMind is working with researchers at the Wellcome Sanger Institute to profile a wider spectrum of human cell types, developmental lineages, and perturbed states. Looking further ahead, Avsec said that subsequent versions might incorporate data from multiple species, and that models like AlphaGenome 2 could ultimately contribute to the development of "virtual cells" — digital simulations designed to mimic biological systems.

Non-commercial researchers can begin using AlphaGenome 2 right away through the AlphaGenome API, while commercial researchers may apply for a limited private preview. The announcement is part of a broader push by major AI labs into science and medicine: Anthropic has been experimenting with drug design and given its Claude model access to a wet lab, while OpenAI has been applying its systems to longstanding problems in mathematics, according to The Verge.

Why it matters — By predicting how the same genetic variant behaves differently in specific cell types — including disease states never seen during training — AlphaGenome 2 gives researchers a more precise tool to connect genetic risk factors to the cells where disease actually begins, which is a prerequisite for designing targeted treatments.

Reported by theverge.com, deepmind.google, marktechpost.com