Predicting the effects of variants for each cell type
Google DeepMind, Google's research division, on Nov. 11 U.S. time (Nov. 12 Japan time) announced AlphaGenome 2, an AI model that predicts how DNA sequences and their variants affect each type of human cell. From the same day, it is available through the "AlphaGenome API" for noncommercial research, while access for commercial research is in a private preview. The company positions it as a tool to help scientists pinpoint the cellular mechanisms that cause disease.
Every cell in the body carries the same DNA, but which genes are active differs by cell type and state, such as heart muscle cells or immune cells that fight infection. How small differences in DNA (variants) change the activity of these genes (expression) is considered a major factor in human traits and disease, and the company said predicting it remains an unsolved problem.
Last year, the company released the first "AlphaGenome," which predicts how variants affect molecular activity as an average across whole tissues. It later also released "AlphaGenome Atlas," which contains predictions for all 9 billion possible single-letter changes in the human genome. The new model narrows the unit of prediction from tissue averages to individual cell types and states, and its architecture has been rebuilt. The company said it finds nearly twice as many variants that alter gene activity in cells as the previous state-of-the-art model, and that it was trained on data from more than 100 million individual cells across 68 tissues.
Predictions made by specifying conditions one at a time
The first model took a DNA sequence of 1 million letters and predicted the results of thousands of experiments across many tissues all at once. The company said this approach spread performance thin and made it hard to capture fine differences between closely related cell types. AlphaGenome 2 instead takes tissues, cell types and states as "conditions" specified one at a time and makes predictions together with a 1-million-letter DNA sequence. The company said this lets the model learn how the same DNA is read differently depending on cell type.
Training used whole-tissue and single-cell data from human and mouse, including data from more than 107 million individual cells, according to the company. The predicted outputs are the same as the first model's — gene expression, histone modifications and chromatin accessibility (how readable the DNA is) — plus DNA methylation, which had been widely requested. It can also take the results of just one type of measurement as input and estimate other outputs that were not measured.
When researchers input single-cell data measured in their own experiments, the model can make predictions about new cell states without being retrained. According to the company, gene activity data from a few dozen cells is enough. On data from a study of COVID-19 mRNA vaccines, the company said the model distinguished immune cells (T cells) activated by the vaccine from resting cells and showed that variants linked to severe illness strongly disrupt gene activity in the activated state.
Causal variant identification rate of 30% on immune cell data, vs. 16% for comparison
According to the company, AlphaGenome 2 identified causal variants 30% of the time on "TenK10K," a large single-cell dataset covering 28 types of immune cells. It correctly predicted whether gene activity would go up or down 43% of the time. The comparisons, the original AlphaGenome fine-tuned on single-cell data and other specialized tools, scored 16% and 23%, respectively, which the company described as nearly twice. The blog does not state the denominators of the rates or the detailed evaluation conditions.
In inflammatory bowel disease (IBD), recent research had shown that a variant in a distant gene regulatory region causes a gene called ETS2 to be overactive in intestinal immune cells (macrophages). AlphaGenome 2 reproduced this result purely computationally by making predictions across a range of intestinal immune cells, according to the company. The company said the model identified the causal variant and showed that it selectively boosts ETS2 activity in intestinal macrophages while having little effect on other cell types or genes.
The company has precomputed predictions across hundreds of cell and tissue conditions for variants linked to diseases and traits. It said these can narrow down likely causal variants, and the cell types in which they act, for multiple traits and diseases ranging from height to type 2 diabetes.
- AlphaGenome 2
- Original AlphaGenome fine-tuned on single-cell data and others
All performance figures were presented by Google DeepMind on its official blog, and ORIENS has not confirmed any third-party evaluation.
Free for noncommercial research, private preview for commercial research
AlphaGenome 2 is available through the AlphaGenome API for noncommercial research. According to the API page, noncommercial use is free, and the same API can also call the first model and the precomputed results of the Atlas in addition to AlphaGenome 2. The company also released an interactive Colab tutorial for trying predictions with single-cell data.
Under the terms of use, the API is limited to individuals, universities and nonprofit research institutions, educational and government bodies, and news organizations. For-profit companies cannot use it, even for noncommercial work, and it cannot be used for research on behalf of companies. Google does not currently charge for it but said it may introduce fees with one month's notice. According to the GitHub documentation, the rate of queries accepted varies with demand, and more queries are accepted for lookups in the precomputed Atlas. It says on-the-fly predictions with the model are intended for small-scale analyses and are not suited to large-scale, genome-wide analyses.
Access for commercial research is in a private preview, and the company has not disclosed who is eligible or the pricing. The GitHub documentation states that AlphaGenome is also available for commercial use through Google Cloud. It does not specify which models are covered.
The company said it will soon release "Skills" for using AlphaGenome 2 in its development environment "Google Antigravity," but it did not give a timeline.
The company states clearly that AlphaGenome and AlphaGenome 2 are for research only and have not been validated or approved for clinical or diagnostic purposes. The terms of use also prohibit using them for clinical purposes or relying on them as medical advice.
Reactions from researchers and industry
Sarah Teichmann, co-founder of the Human Cell Atlas and a professor at the University of Cambridge in Britain, praised the model's ability to handle new single-cell data without retraining and to predict how variants act in new states even from a small number of cells. Peter Donnelly, CEO of Britain's Genomics Limited, said it is difficult to go from associations found in genome-wide association studies (GWAS) to the information needed for drug discovery programs, that DNA sequence prediction models are an important tool for bridging that gap, and that he looks forward to incorporating AlphaGenome 2 into the company's "Mystra."
