Google DeepMind has introduced Google AlphaGenome 2, a research model designed to predict how a change in DNA could affect gene regulation in a particular human cell type or cellular state. The intended use is more specific than flagging a variant as potentially consequential somewhere in the body: researchers can ask which cells may be affected.
That distinction addresses an awkward fact of human biology. Most cells carry the same DNA, but heart, skin and immune cells activate different sets of genes. Their patterns of gene activity can also change with cellular state. A genetic variant may therefore have different consequences depending on the cell in which it appears, as The Verge reported.
How does Google AlphaGenome 2 differ from the original model?
The first AlphaGenome was a sequence-to-function model: it could take as much as 1 million DNA base pairs and predict a range of molecular features, including gene expression, chromatin accessibility, histone modifications, transcription-factor binding, chromatin contacts and RNA-splicing outputs. Research on that earlier model was published in Nature in January 2026.
AlphaGenome 2 adds biological condition as part of the prediction problem. DeepMind says it can model the expected effect of variants in selected cell types and states, rather than offering only the broader predictions of its predecessor. The company trained the model on more than 100 million individual cell profiles from 68 tissues.
DeepMind also says scientists can supply their own single-cell RNA-seq data to adapt the model to biological states it has not encountered, without fine-tuning. The company says AlphaGenome 2 identifies nearly twice as many gene-expression-linked variants in particular human cell types as its prior model. Those are DeepMind performance claims, not independent validation.
What AlphaGenome 2 can and cannot tell researchers
The model is meant to help prioritize hypotheses about how a variant could alter molecular processes tied to gene regulation in a particular cellular context. That could make a large list of genetic candidates more manageable for research into disease mechanisms. It does not establish that a variant causes a disease, prove a treatment will work, or diagnose a patient.
There is no supplied evidence of independent validation for AlphaGenome 2's specific results, peer-reviewed publication of the new model, clinical validation, or demonstrated patient benefit. Secondary reporting says it is not validated for clinical or diagnostic use. The gap matters: a promising model output remains a prediction, even when it points toward a plausible cell type.
Researchers can access AlphaGenome 2 for non-commercial work through the AlphaGenome API, according to DeepMind's announcement summary. Commercial research access is in private preview.
Do not confuse AlphaGenome 2 with AlphaGenome Atlas
- Original AlphaGenome: the earlier DNA sequence model, built to predict several kinds of functional genomic measurements.
- AlphaGenome 2: the newer model that DeepMind says adds cell-type and cell-state context to variant-effect predictions.
- AlphaGenome Atlas: a separate catalog of AlphaGenome predictions for possible single-letter DNA changes, including the company's Variant Impact score.
That naming scheme is doing no favors. Still, the substantive change in AlphaGenome 2 is clear enough: DeepMind is trying to move variant interpretation closer to the cells in which a predicted effect may occur, while leaving the work of proving biological and clinical significance unresolved.
This story draws on original reporting from The Verge.
