Google DeepMind has announced AlphaGenome 2, an AI model that predicts the molecular effects of non-coding DNA variants across hundreds of different cell types. The model replicated aspects of a 2024 finding on inflammatory bowel disease; James Lee said it took 30 minutes to reach the answer, compared with a year and a half of experiments.
The model is an update to AlphaGenome 1, which generated predictions for every possible single-letter change to the human genome but could not specify the precise type of cell in which a variant was likely to act. James Lee, an immunologist and geneticist at the Francis Crick Institute, said this meant the earlier tool had not been hugely useful for his team.
In 2024, Lee led a study that identified ETS2 as a driver of inflammatory bowel disease. The gene is overactive in immune cells due to a DNA variant that predisposes people to the condition. The variant lies in a 'gene desert' far from protein-coding sequences. Lee said the finding took a year and a half of experiments.
To address cell-type specificity, Žiga Avsec, the Google DeepMind researcher leading the AlphaGenome team, and colleagues trained AlphaGenome 2 on gene activity measurements from more than 100 million individual cells across 68 tissues. Adam Gayoso, a DeepMind scientist who co-authored the preprint, said training on individual cells gave a richer, more fine-grained representation of the differences between cells.
Avsec said that understanding variants associated with genetic diseases requires knowing the cell types affected. The new model can generate predictions in the 637 human and 311 mouse cell types it was trained on, and researchers can also feed it their own single-cell gene activity data to generate variant predictions in those cells.
Lee said the model worked out the answer in 30 minutes, compared with the year and a half of experiments his team had conducted. He described the replication as reassuring.
Analysis
Why This Matters
- Dramatically accelerates the identification of disease-causing genetic variants in non-coding DNA, a notoriously difficult area of genomics.
- Cell-type specificity allows researchers to link variants directly to relevant tissues, a critical step for understanding complex diseases.
- Enables researchers to test hypotheses about variants in disease states without lengthy and expensive lab experiments.
Background
The human genome is 98% non-coding DNA, which is difficult to decipher. In 2024, researchers identified an IBD-linked variant in a 'gene desert' using a year and a half of traditional methods. DeepMind's AlphaGenome 1 could predict variant effects but not the affected cell type, limiting its utility. AlphaGenome 2 is the successor trained on data from individual cells.
Key Perspectives
James Lee, Francis Crick Institute: Called the speed of AlphaGenome 2 reassuring, noting it replicated his team's year and a half of experiments in 30 minutes.
Žiga Avsec, Google DeepMind: Says understanding which cell types are affected is essential for connecting genetic variants to disease.
Adam Gayoso, Google DeepMind: Highlights the model's ability to accept new single-cell data from researchers, allowing predictions beyond its original training set.
What to Watch
- Whether AlphaGenome 2's predictions lead to validated drug targets for IBD and other complex genetic diseases.
- The extent of adoption by biomedical researchers and how it changes the experimental genomics workflow.
- Independent validation and benchmarking of the model's predictions by external academic groups.
Checked
11 claims in this story checked against 1 source: 0 supported.
Show each claim
- Not in the sourcesAlphaGenome 2 replicated aspects of a 2024 inflammatory bowel disease finding.Nature: “DeepMind's ultra-powerful AI for genome mapping wows scientists.”
- Not in the sourcesJames Lee said AlphaGenome 2 reached the answer in 30 minutes, compared with a year and a half of experiments.Nature
- Not in the sourcesAlphaGenome 2 predicts the effects of non-coding DNA variants across hundreds of cell types.Nature: “DeepMind's ultra-powerful AI for genome mapping wows scientists.”
- Not in the sourcesAlphaGenome 1 could not specify the precise cell type in which a variant was likely to act.Nature: “DeepMind's ultra-powerful AI for genome mapping wows scientists.”
- Not in the sourcesGoogle DeepMind announced AlphaGenome 2.Nature: “DeepMind's ultra-powerful AI for genome mapping wows scientists.”
- Not in the sourcesAlphaGenome 1 generated predictions for every possible single-letter change to the human genome.Nature
- Not in the sourcesJames Lee said AlphaGenome 1 had not been hugely useful for his team.Nature
- Not in the sourcesIn 2024, Lee led a study identifying ETS2 as a driver of inflammatory bowel disease.Nature
- Not in the sourcesA DNA variant causes ETS2 to be overactive in immune cells and predisposes people to inflammatory bowel disease.Nature
- Not in the sourcesAlphaGenome 2 was trained on gene-activity measurements from more than 100 million individual cells across 68 tissues.Nature
- Not in the sourcesAlphaGenome 2 can generate predictions in 637 human and 311 mouse cell types.Nature
