DeepMind’s AlphaGenome 2 predicts how DNA changes play out in specific cell types
Google DeepMind announced AlphaGenome 2 on Nov. 11. The company describes it as a model “that predicts how DNA sequences and their variations impact specific human cell types,” according to its blog post. DeepMind says the model is for research use only.
What it does
The same stretch of DNA can behave differently in a liver cell than in an immune cell. AlphaGenome 2 is built to account for that. According to DeepMind’s technical report, the model takes in a DNA sequence of about 1 million base pairs plus a target tissue, cell type or cell state, and it predicts how that sequence and its variants affect that particular context.
DeepMind says the model can also work out the target cell state from a scientist’s own single-cell RNA sequencing data, a measurement of which genes are active in individual cells, without being retrained. It was trained on human and mouse data, including more than 107 million single-cell profiles, and it adds DNA methylation, a chemical tag that affects whether genes are switched on, as a new type of prediction.
The headline result is DeepMind’s
DeepMind’s main reported result comes from its own evaluation on TenK10K, an immune-cell dataset. The company says AlphaGenome 2 identified “nearly twice as many causal variants (30% vs. 16%)” as a fine-tuned version of the original AlphaGenome and other tools. The report describes this as “doubling recall of causal single-cell eQTLs at 90% precision.” eQTLs are DNA variants linked to changes in how strongly a gene is expressed.
The report is a technical paper released by DeepMind. Assimilated Press found no indication that it has been peer reviewed.
Limits DeepMind states
The report is direct about what the model can’t do. “At high precision, AlphaGenome 2 recovers only a minority of fine-mapped eQTLs and putatively causal GWAS variants,” it says. In other words, when the model is tuned to be confident, it still misses most of the variants researchers have already flagged as likely causes.
DeepMind also writes that the model “can still struggle to distinguish fine-grained states,” and that the model’s learned cell-state embeddings “reflect correlational rather than causal relationships.”
The report’s examples involving inflammatory bowel disease and COVID-19 are computational: the model reproduces or predicts findings. They aren’t new clinical results.
Who can use it
DeepMind says AlphaGenome 2 is available “for non-commercial research via the AlphaGenome API, and in private preview for commercial research.” The company says AlphaGenome 2 Skills in Google Antigravity are coming “soon.”
The blog states that the models are “intended for research use only and have not been validated or approved for clinical or diagnostic purposes.”
The Verge also covered the release.
Written with AI, edited by a human.
