AlphaGenome 2 Finds Twice as Many Gene-Altering DNA Variants
Google DeepMind's AlphaGenome 2 predicts how small DNA changes alter genes in single cell types and found twice as many variants, per its creators.

Google DeepMind’s AlphaGenome 2, an AI model that predicts how small changes in DNA affect gene activity in individual cell types, found twice as many expression-altering variants as its predecessor, according to its creators. The result was reported by Chemistry World and comes from a preprint that has not been peer reviewed.
The model is meant to help researchers study the cell types that matter in disease, rather than averaging over whole tissues.
How the new model was trained
The first AlphaGenome read DNA sequences up to one million base pairs long and predicted how variations changed gene activity. Most of its training data came from bulk tissue, so the measurements described a tissue as a whole. Differences between closely related cells were blurred.
For AlphaGenome 2, the training data included more than 100 million single-cell profiles across 68 tissues. The architecture was also changed. The new model still reads the same one million base pairs, covering protein-coding and non-coding regions, but it now predicts for one cell type or cell state at a time.
“The new model predicts one cell type at a time and in that way it can use the full bandwidth to focus on that,” said Ziga Avsec, a genomics researcher at DeepMind. In a blog post, the developers called the result a “step-change improvement” in predicting the effects of variants.
A test on inflammatory bowel disease
To check the model, the team had it revisit an earlier study of inflammatory bowel disease (IBD). That study had found, through laboratory experiments, a non-coding DNA sequence that turns up a gene driving stronger inflammation in specific immune cells. The AI model correctly identified the responsible DNA variant, the gene it affects and the cell type where it matters most, which is the macrophage, a kind of immune cell.
Large genetic studies can point to a stretch of DNA linked to a disease, but experiments are then needed to find the exact variants responsible. The work follows the recently released AlphaGenome Atlas, which charted the molecular effects of mutating every DNA base pair in the human genome one at a time.
What scientists say about the limits
James Lee of the Francis Crick Institute in the UK led the original IBD discovery. He said researchers now test variants experimentally to find those likely to change gene expression, and that the model could speed that up if its predictions are accurate. “This is a big if though,” he said, adding that some experimental validation will probably still be needed.
The developers agree. Avsec said the model returns a prediction of zero for many variants, which usually means it does not know how the variant acts. People will still need to do a lot of experimental follow-up to confirm the hypotheses it generates, he said.
Peter Koo, a quantitative biologist at Cold Spring Harbor Laboratory in the US, noted that models linking DNA sequence to function have existed for a decade. In his view, the new version is essentially tuned to be more aware of cell types, and much of the advance lies in the engineering needed to make that work. He is nonetheless impressed by the model’s scale. Access to the full model, he said, will let researchers explore how genetic variation acts across cell types, though it will take time to learn where it works well and where it does not.
Why it matters
Most of the DNA that influences disease does not code for proteins, and its effects depend on which cell is reading it. A model that predicts per cell type could let biologists narrow a long list of candidate variants before running experiments. Adam Gayoso at DeepMind said any biologist with single-cell RNA sequencing data can plug it into the model and get predictions about the effect of variants in their context, although the team cannot test every possible cell state.
For readers new to the topic, how DNA works covers the genes and proteins behind these predictions. The twofold gain in variants found is the developers’ own figure from a preprint, and independent testing of the full model has yet to be reported.
Source: Chemistry World



