AlphaGenome 2 refines variant reading in each cell type
Google DeepMind unveils AlphaGenome 2, a model estimating the effect of genetic variants at the cellular type level. The tool, available for non-commercial research, aims to better link DNA, gene activity, and disease.
Google DeepMind has introduced AlphaGenome 2, an AI model designed to predict the effect of genetic variants on specific human cell types. The tool no longer provides just an overview of tissues: it drills down to the cellular level, a significant shift for genomics research and understanding disease mechanisms. According to the note published on November 12, 2026, it is already accessible for non-commercial research via API, without approval for clinical use.
De la vue d’ensemble à la cellule précise
The previous model, AlphaGenome, worked at the scale of tissues and very long DNA sequences, but this approach diluted signal precision. AlphaGenome 2 takes a different path: it functions as a conditional model, with a DNA sequence and a targeted biological state as input. As a result, it can estimate how a variant acts in a given cell type, rather than smoothing the effect across multiple tissues at once.
The model was trained on over 107 million individual cell profiles, from 68 human and mouse tissues. This base is used to predict not only gene expression but also DNA accessibility in the nucleus and methylation, chemical marks that can activate or silence gene expression. In short, the tool seeks to link three layers often analyzed separately in labs.
Pourquoi cette précision change la donne
DeepMind claims AlphaGenome 2 identifies nearly twice as many relevant variants as its predecessor. On the TenK10K dataset, covering 28 primary immune cell types, the model detected about 30% of causal variants and correctly predicted, in 43% of cases, whether a variant increases or decreases gene expression. The refined AlphaGenome 1 version reached 16% and 23%, respectively.
The gain is not just a benchmark score. In an example related to inflammatory bowel disease, the model reproduced a mechanism already observed experimentally and isolated a regulatory variant distant from the ETS2 gene, active in intestinal macrophages. It also better tracked the maturation of blood stem cells into red blood cells. This level of precision can help prioritize experiments rather than running the lab like a high-throughput laundry machine.
Ce que les chercheurs peuvent en faire dès maintenant
AlphaGenome 2 can also extrapolate to cellular states it has not seen during training. Google DeepMind explains that a researcher can provide a few dozen measured cells, then let the model predict other missing biological properties. This ability is particularly interesting for teams working with incomplete data or still limited cohorts.
The system was also tested on data related to an mRNA COVID-19 vaccine. It reliably distinguished activated T cells from resting cells and showed that a variant associated with severe COVID-19 forms mainly disrupted cellular activity in the activated state. In resting cells, the effect went unnoticed. This is precisely the kind of detail that broader approaches tend to smooth out.
Une brique pour la recherche, pas un outil clinique
DeepMind precomputed predictions for hundreds of cellular and tissue conditions to aid disease research, genome-phenotype association studies, and eventually drug discovery. GWAS, which link DNA regions to traits like height or type 2 diabetes, often locate signals in broad areas. AlphaGenome 2 helps narrow the target and identify the most plausible cells, whether astrocytes or specific immune cells.
The model is available for non-commercial research via an API, while commercial users access it through a closed early access. It is not approved for clinical diagnosis. DeepMind also plans to soon integrate dedicated AlphaGenome 2 functions into Google Antigravity, its agentic development platform. The ambition is clear: to accelerate dialogue between genetics, cell biology, and experimentation. It remains to be seen how many labs will move from testing to regular use.
By the numbers
- 107 million — cell profiles used for training, human and mouse, according to Google DeepMind.
- 68 tissues — source of training data mentioned by DeepMind.
- 28 cell types — scope of the TenK10K benchmark used for evaluation.
- 43% — cases where the model correctly predicted a variant’s effect on gene expression.
- November 12, 2026 — date of publication of the presentation by THE DECODER, citing Google DeepMind.