Google DeepMind Launches AlphaGenome 2: What It Means for Disease Research
Google DeepMind unveiled AlphaGenome 2, an AI model that predicts how DNA variants affect cell types, a step toward tracing the cellular roots of disease.
Image: Wikimedia Commons / National Human Genome Research Institute
LONDON — Google DeepMind on Wednesday unveiled AlphaGenome 2, a new artificial intelligence model that predicts how changes in DNA affect individual human cell types, giving researchers a sharper tool for tracing the cellular origins of disease.
The model, announced on November 11, moves beyond the tissue-wide averages of the first AlphaGenome. It conditions each prediction on a single tissue, cell type or cell state, so one DNA sequence can behave differently across cell types. DeepMind trained it on more than 107 million individual cell profiles from human and mouse data across 68 tissues.
Key Takeaways
- AlphaGenome 2 predicts how DNA variants change gene activity in specific cell types, rather than in tissue averages.
- On a benchmark of 28 immune cell types, it recovered 30 percent of likely causal variants, nearly double the first AlphaGenome’s 16 percent.
- It is available now for non-commercial research through the AlphaGenome API; commercial research access is in private preview.
- DeepMind stresses the model is for research only and has not been validated or approved for clinical or diagnostic use.
How AlphaGenome 2 Predicts DNA Effects in Individual Cell Types
AlphaGenome 2 takes a DNA sequence of up to one million letters together with a chosen biological condition and predicts gene expression, chromatin accessibility, histone marks and DNA methylation for that condition. It can also adapt to a researcher’s own single-cell data without retraining. Most DNA variants linked to disease sit outside protein-coding genes, in regions that switch genes on or off, and their effect often shows up in only one cell type.
AlphaGenome 2 vs AlphaGenome 1: What the Benchmarks Show
DeepMind’s main test used TenK10K, a single-cell map of 28 primary immune cell types. On that benchmark, AlphaGenome 2 recovered 30 percent of putatively causal expression-linked variants at 90 percent precision, against 16 percent for the first AlphaGenome. It also called the direction of gene effects correctly in 43 percent of cases, against 23 percent.
| Measure | AlphaGenome 1 | AlphaGenome 2 |
|---|---|---|
| Training data | Bulk tissue and cell-line data | 107M+ cell profiles, 68 tissues |
| Prediction scope | Broad tissues and cell lines | One tissue, cell type or state |
| Causal variants found on TenK10K | 16 percent | 30 percent |
| Gene-effect direction correct | 23 percent | 43 percent |
| Access | AlphaGenome API | Non-commercial API; commercial in private preview |
How Researchers Can Access AlphaGenome 2 and What It Cannot Do
AlphaGenome 2 is available from today for non-commercial research through the AlphaGenome API, free of charge. Commercial research access is in private preview, and DeepMind has not said when it will open more widely. The release follows Anthropic’s Fable 5.2 and Mythos 5.2, and Google has also been building out the hardware behind its AI work, including its latest tensor processing units.
The launch lands as the AI industry faces closer scrutiny over spending and returns. A Federal Reserve governor warned this week that heavy AI investment could trigger a market correction if financial returns fall short.
DeepMind is explicit about the model’s limits. Both AlphaGenome versions are intended for research use only and have not been validated or approved for clinical or diagnostic purposes. The model cannot diagnose disease or read a personal genome on its own.
Frequently Asked Questions
What is AlphaGenome 2 and what does it do?
AlphaGenome 2 is a DNA sequence model from Google DeepMind, announced on November 11, 2026. It predicts how a change in DNA affects gene activity in one cell type at a time, rather than in a tissue average.
Can AlphaGenome 2 diagnose diseases or predict my health risks?
No. Google DeepMind says both AlphaGenome models are for research use only and have not been validated or approved for clinical or diagnostic purposes. The model cannot diagnose disease or read a personal genome.
How can researchers access AlphaGenome 2?
Non-commercial researchers can use it now through the AlphaGenome API, free of charge. Commercial research access is in private preview, and DeepMind has not announced a date for wider commercial availability.
How is AlphaGenome 2 different from the first AlphaGenome?
The first AlphaGenome predicted molecular activity across broad tissues. AlphaGenome 2 conditions each prediction on one cell type or state, adapts to a researcher’s own single-cell data, and found nearly twice as many causal variants on DeepMind’s benchmark.
DeepMind said AlphaGenome 2 Skills for Google Antigravity will follow. It has not yet said when commercial access will move beyond private preview.
