Google DeepMind has introduced AlphaGenome 2, a new version of an artificial intelligence model that predicts how small changes in DNA affect the functioning of specific types of human cells. It is intended to help scientists determine in which cells and through which genes inherited variations contribute to disease. According to the company, the model identifies nearly twice as many variants that alter gene activity as its predecessor.
Every cell in the body carries the same genetic blueprint. But a heart muscle cell activates different genes than a white blood cell currently fighting an infection. International projects such as the Human Cell Atlas map gene activity in millions of cells, but reliably estimating how a tiny variation in DNA will affect it has so far proved difficult. Yet such variations underlie many human traits and diseases.
From whole-tissue averages to individual cells
DeepMind introduced the first AlphaGenome last year, and a detailed article about it was published in the journal Nature this January. From a DNA sequence one million letters long, the model could predict how variants would manifest in different tissues. The company later added the AlphaGenome Atlas, containing predictions for all roughly 9 billion possible single-letter substitutions in the human genome.
However, the older version calculated results for thousands of experiments at once. The approach was efficient, but the model's performance was diluted, causing it to miss subtle differences between related cell types. The team therefore rebuilt the architecture from the ground up. AlphaGenome 2 always focuses on one specific sample—a tissue, cell type, or its current state—and receives this information together with the DNA sequence at every stage of the calculation. This allows it to learn how different cells read the same piece of DNA differently.
“When you develop drugs, you target specific proteins and specific cell types,” Žiga Avsec, who leads the genomics team at DeepMind, explained to The Verge.
The model was trained on more than 100 million individual cell profiles from 68 human and mouse tissues. It can make the same predictions as the original version, including gene activity, histone modifications, and chromatin accessibility, while also adding DNA methylation, which scientists had frequently requested. It can also fill in missing data. If a researcher provides a single measurement, such as gene activity in cells, the model also estimates properties that no one has measured. The laboratory can then make a better-informed decision about which experiment to conduct next.
In half an hour, what took a year and a half to find
On the TenK10K dataset, which covers 28 types of immune cells, AlphaGenome 2 identified 30 percent of the variants that actually cause changes in gene activity. The original model, fine-tuned on single-cell data, found only 16 percent. When predicting whether a variant would increase or decrease gene activity, accuracy improved from 23 to 43 percent.
Scientists also tested the new model on a real discovery. A 2024 study showed that a variant in a distant region of DNA that regulates gene activity contributes to inflammatory bowel disease. It does so by excessively activating the ETS2 gene in immune cells in the gut known as macrophages. AlphaGenome 2 reproduced this conclusion entirely in silico. It correctly identified the culprit and showed that it strengthens ETS2 specifically in macrophages while leaving other cells and genes virtually unchanged.
Immunologist and geneticist James Lee of London's Francis Crick Institute, who led the original research, commented succinctly to Nature. According to him, the model arrived at the answer in 30 minutes, while his team needed 18 months of experiments.
Just a few dozen cells are enough
AlphaGenome 2 can also handle cells it has never seen during training, without needing to be retrained. A scientist uploads a snapshot of gene activity from just a few dozen cells, and the model converts it into a description of the cellular state, which it then uses for its calculations. According to DeepMind, the results are as good as or better than those from models trained specifically on the given data.
“What excited me about AlphaGenome 2 is how it adapts to new data without additional training and predicts the effects of variants in newly described cellular states, even from a small sample,” said Sarah Teichmann, co-founder of the Human Cell Atlas consortium and a professor at the University of Cambridge.
This is well illustrated by data from a study of an mRNA vaccine against COVID-19. The model distinguished T cells activated by vaccination from resting ones and found that a variant associated with severe disease disrupts gene activity more significantly in the activated cells. If scientists had examined only resting cells, this effect would have remained hidden.
A bridge between genetics and the map of the human body
Genome-wide association studies have long identified thousands of locations in DNA linked to diseases and physical traits. However, they often identify only a broader region, leaving scientists with the difficult task of determining which specific variant is responsible, which gene it affects, and in which cell. DeepMind therefore precomputed predictions for these variants across hundreds of cell and tissue types so researchers would not have to run the computationally demanding analyses themselves.
The model is particularly helpful in two areas. First, it links enhancers—distant switches in DNA—to the genes they control. Previous experiments were conducted mainly on cultures grown in laboratories, but AlphaGenome 2 expands this map to cells taken directly from the human body, such as astrocytes, which support and protect neurons in the brain. For traits ranging from height to type 2 diabetes, it also shows which variants and cell types are most likely involved. If the model is sufficiently confident, it also estimates whether a variant will increase or suppress gene activity.
Drug developers are also interested in the tool. Peter Donnelly, CEO of Genomics Limited, praised the way such models help move from statistical associations to what pharmaceutical research needs. His company plans to integrate AlphaGenome 2 into its Mystra platform.
Academic scientists can use the model for non-commercial research through the AlphaGenome API, while companies are currently testing it in closed early access. DeepMind is working with the Wellcome Sanger Institute to characterize additional cell types and plans to make the model available soon on the Antigravity developer platform. It has not yet been validated or approved for clinical or diagnostic use.
Sources: deepmind.google, nature.com and theverge.com



