Skip to content
AI.info

The Pulse

Google details AlphaGenome research uses after cell-specific model release

Google DeepMind says researchers have used AlphaGenome Atlas to narrow genetic variants for studies of rare disease and body mass index. Its newer AlphaGenome 2 model predicts how variants affect specific cell types and is available for non-commercial research.

Google details AlphaGenome research uses after cell-specific model release

Researchers have used Google DeepMind’s AlphaGenome Atlas to narrow lists of genetic variants for studies of rare disease and body mass index, according to the company. In a Google interview published Nov. 17, research scientist Žiga Avsec described those uses alongside a newer model designed to predict how DNA changes affect specific cell types.

The original AlphaGenome model takes a partial DNA sequence of 1 million base pairs and makes thousands of predictions about it. Google released that model in June 2025 and introduced the Atlas in September 2026.

Ranking genetic changes

The human genome contains 3 billion DNA base pairs, with letters representing the bases that make up the sequence. Because each letter can be replaced by three others, there are 9 billion possible single-letter changes.

Google used AlphaGenome to predict the effects of those changes and assemble the Atlas, a 1-petabyte dataset. The Atlas provides an AlphaGenome Variant Impact, or AVI, score that combines predictions for coding and non-coding regions and helps researchers prioritize variants for further study.

Avsec described a hypothetical patient with a rare disease whose DNA differs from the standard reference human genome at 4 million variants. Even after researchers narrow the list to rare variants, hundreds or thousands of candidates could remain, he said.

Early research cases

Gareth Hawkes, a Medical Research Council fellow at the University of Exeter, used the Atlas to narrow a list of 1 billion genetic variants, according to Google. He focused on the most impactful 1% and identified 19 variant groups linked to body mass index for research.

At the Broad Institute, research fellow Laura Covill used the AVI score to prioritize variants in research on unsolved rare diseases. The Atlas identified a variant in the DNM1 gene and predicted that it created an incorrect splice site, providing supporting evidence that helped solve the case, Google said.

Moving to individual cells

AlphaGenome 2, released on Nov. 11, is designed to predict how variants affect individual cell types rather than relying on broad tissue-level outputs. The earlier model was trained mostly on data from whole tissues or cell lines, Avsec said.

The newer model takes a DNA sequence and a biological condition—such as a tissue, cell type or state—as inputs. Researchers can also supply their own single-cell gene-activity data to adapt it to previously unseen cell states without retraining the model.

Google DeepMind says the model was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles. On the TenK10K dataset of 28 primary immune cell types, the company reported that it identified 30% of causal variants versus 16% for the original model fine-tuned on single-cell data and other specialized tools; the corresponding figures for directional gene effects were 43% versus 23%.

AlphaGenome 2 is available for non-commercial research through the AlphaGenome API and in private preview for commercial research through Google Cloud. Avsec said the team is researching how to measure confidence in predictions and advised users to interpret them alongside other sources of information.

Sources

  1. Google
  2. Google DeepMind
  3. Google

Companies in this story

Related stories

Explore

More articles