Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolution with AlphaGenome 2

Google DeepMind has released AlphaGenome 2, a genomics model that predicts how genetic variants affect specific human cell types instead of aggregating effects across bulk tissue. The announcement covers the model, its benchmark results and its access arrangements. AlphaGenome 2 was trained on more than 107 million individual cell profiles drawn from 68 tissues in humans and mice, according to DeepMind, and is aimed at helping researchers find the cellular drivers of disease.
The motivating problem is that every cell in the body carries the same DNA, but different cell types switch different genes on and off. A heart muscle cell reads its genome differently from an immune cell, and predicting those differences is one of genomics' harder open problems. The first AlphaGenome took a DNA sequence a million letters long and predicted its properties across thousands of experiments spanning many tissues simultaneously. DeepMind says that parallel approach spread the model's capacity too thin and made subtle differences between closely related cell types hard to capture.
AlphaGenome 2 restructures the task as a conditional model. In addition to the DNA sequence, it accepts a specific biological state as an extra input and computes effects for that one cell type. Beyond predicting how strongly a gene is expressed, it estimates how DNA is physically packaged in the nucleus, showing which stretches of the genome are accessible. It also predicts DNA methylation, the chemical tags that act as molecular switches to silence or activate genes.
A second capability is inferring missing data from a single measurement. If a researcher has measured only a cell's gene activity, the model can supply predictions for other properties, such as how tightly or loosely certain DNA stretches are packaged. In practice this suggests a prioritization tool: labs could use it to decide which experiments are worth running before committing bench time and reagents.
The architecture also adapts to new experiments without retraining. Researchers can feed in gene activity data from just a few dozen cells, and the model computes effects for that previously unseen cell state. DeepMind says it often beats specialized models trained directly on that data, which, if it holds generally, would reduce the cost of applying the model to new cell systems.
On the TenK10K benchmark dataset covering 28 primary immune cell types, AlphaGenome 2 identified about 30 percent of causal variants. In 43 percent of cases it also correctly predicted whether a variant upregulates or downregulates a gene. A fine-tuned version of AlphaGenome 1 managed 16 and 23 percent respectively. The model also tracks continuous changes, such as blood stem cells maturing into red blood cells, with considerably greater accuracy than its predecessor.
Two case studies are offered. Working purely computationally, the model reproduced an inflammatory bowel disease mechanism previously identified in a study: it isolated the exact DNA variant in a regulatory switch far from the gene that selectively overactivates ETS2 in gut macrophages, while other cell types and genes remained largely unaffected. In data from a COVID-19 mRNA vaccine study, it distinguished vaccine-activated T cells from resting ones and showed that a gene variant linked to severe COVID-19 disrupts cell activity mainly in the activated state, an effect that would have stayed invisible in resting cells.
DeepMind has precomputed predictions for hundreds of cell and tissue conditions, partly to support AI-driven drug development. These are also intended to help with genome-wide association studies, which link regions of DNA to traits such as height or type 2 diabetes but typically localize signals only to broad stretches of the genome. AlphaGenome 2 narrows the likely causal variants and identifies affected cell types, such as astrocytes in the brain or specific immune cells. When its confidence is sufficient, it also predicts whether a mutation raises or lowers gene expression, which matters for drug development. For developers, the shift to a conditioned single-cell interface is the notable design change: the same sequence can be scored under many states without separate models per tissue.
Access is split. AlphaGenome 2 is available now for noncommercial research through an API, while commercial users get access through a closed preview. The system is not approved for clinical diagnostics. DeepMind also plans to roll out dedicated AlphaGenome 2 features on its agentic development platform Google Antigravity, where the model is set to be combined with tools such as Gemini 4 for automated research workflows. A likely trade-off is that per-cell conditioning trades breadth for depth: predictions are sharper for the queried state, but the model needs a relevant state input to be useful, and the reported gains are measured on specific benchmarks rather than demonstrated across all tissues. Antibody design, target discovery and functional genomics are the obvious beneficiaries; clinical use is explicitly out of scope for now.
Why it matters: Research teams working on target discovery, functional genomics and drug development get a tool that resolves variant effects at cell-type granularity and proposes follow-up experiments, with free API access for noncommercial work. Product and platform teams should note the planned Antigravity integration, which signals genomics moving into agentic research pipelines. The closed commercial preview and the diagnostic exclusion mean regulated and commercial deployments remain gated for now, an inference from the stated access terms.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Google Deepmind has unveiled AlphaGenome 2, an AI model that analyzes genetic variants for individual cell types rather than across bulk tissue. Trained on more than 107 million cell profiles, it identifies nearly twice as many disease-relevant gene changes as its predecessor. The model is available for noncommercial research through an API and is set to be integrated into Google's agentic platform Antigravity. The article Google Deepmind shifts genomics AI from bulk tissue predictions to single cell type resolutio