AlphaGenome 2 Predicts Gene Activity One Cell Type at a Time

The most interesting change in AlphaGenome 2 is what the model is asked to answer. Its predecessor predicted DNA properties across many tissues at once. The new version takes a specific cell state as an input and works out what a genetic variant does inside that one cell type. That shift addresses a problem that has limited genomics models for a while: the same DNA sequence produces very different gene activity depending on the cell it sits in.

Google DeepMind describes AlphaGenome 2 as a model for predicting how DNA sequences and their variations affect specific human cell types. It was trained on more than 107 million individual cell profiles from 68 tissues in humans and mice, and the stated goal is to help researchers find the cellular drivers of disease. You can read the technical document or DeepMind's blog post for the full details.

Why conditioning on a cell state changes the problem

The original AlphaGenome took in a DNA sequence one million letters long and predicted its properties across thousands of experiments spanning many tissues simultaneously. DeepMind says that approach spread the model's capacity thin. Closely related cell types often differ in subtle ways, and a broad model can average those differences away.

AlphaGenome 2 works as a conditional model. Alongside the DNA sequence, it receives a biological state and calculates effects for that single cell type. Think of it as the difference between asking a general question about a text and asking the same question with the reader's reading level specified. The model has less to juggle at once, and the answer can reflect the particular context.

More than gene expression

Predicting how strongly a gene is expressed is only part of the picture. AlphaGenome 2 also estimates how DNA is physically packaged inside the nucleus, showing which stretches of the genome are accessible. It can now predict DNA methylation as well. Methylation tags act as chemical switches that silence or activate genes, so a model that covers expression, packaging and methylation is looking at several layers of the same regulatory system.

The model can also fill in gaps. If a lab measures only a cell's gene activity, AlphaGenome 2 can predict other properties, such as how tightly certain DNA regions are packaged. For researchers working with limited measurements and a long list of possible experiments, that kind of inference helps decide which experiments are worth running next. It's a practical feature, and probably one of the more useful ones for wet-lab teams.

Benchmark results and a disease mechanism reproduced in silico

On the TenK10K benchmark, which covers 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 reached 16 and 23 percent on the same measures. DeepMind also reports considerably better accuracy on continuous processes, such as blood stem cells maturing into red blood cells.

The more striking result is a case study. Working purely computationally, the model reproduced an inflammatory bowel disease mechanism that was previously identified in a published study. It isolated the specific DNA variant in a regulatory switch located far from the gene it affects. That switch selectively overactivates ETS2 in gut macrophages, while other cell types and genes stayed largely unaffected. Recovering a known mechanism from sequence and cell-state data alone is a useful sanity check, though it doesn't show the model will find new ones.

Adapting to new experiments without retraining

A key feature of the architecture is that it can adapt to entirely new experiments without being retrained. Researchers can supply gene activity data from just a few dozen cells, and the model calculates effects for that previously unseen cell state. DeepMind says it often outperforms specialized models trained directly on that same data. That matters for groups studying rare cell populations, where collecting large datasets is hard.

In an application to a COVID-19 mRNA vaccine study, the model reliably separated vaccine-activated T cells from resting ones. It also showed that a gene variant linked to severe COVID-19 disrupts cell activity mainly in the activated state. In resting cells, that effect would have stayed hidden. This is a good example of why cell-state context matters: the same variant can look harmless in one state and consequential in another.

Connecting genetic association studies to cell biology

Genome-wide association studies, or GWAS, link regions of DNA to traits such as height or type 2 diabetes. They often identify only broad stretches of the genome, leaving researchers with many candidate variants and little sense of which cells are involved. DeepMind says AlphaGenome 2 narrows the likely causal variants and points to affected cell types, such as astrocytes in the brain or specific immune cells. When the model is confident, it also predicts the direction of a mutation's effect on a gene, which is useful for thinking about drug targets.

DeepMind has already precomputed predictions for hundreds of cell and tissue conditions. Those precomputed results are meant to make it easier to link human genetics with cell biology, and they support the company's broader interest in AI-driven drug development. Teams working on target identification could use them as a starting point for hypotheses, though any candidate would still need experimental validation.

Access, limits and what to watch

AlphaGenome 2 is available now for noncommercial research through an API. Commercial users can apply for a closed preview. DeepMind is clear that the system is not approved for clinical diagnostics, which is the right boundary for a model at this stage. Predictions from a computational model are hypotheses, and the gap between a benchmark gain and a validated clinical result is wide.

DeepMind also plans to roll out dedicated AlphaGenome 2 features on its agentic development platform, Google Antigravity. There, the model is set to be combined with tools like Gemini 4 for automated research workflows. If that integration works as intended, the interesting question will be how much of the experimental loop, from variant prioritization to next-experiment selection, can run with human review at the checkpoints rather than human execution at every step. My read is that the cell-type conditioning is the more durable contribution, because it gives every downstream tool a more specific question to answer.