Google DeepMind releases AlphaGenome 2 model for predicting genetic variant effects at the cellular level
Google DeepMind has introduced the AlphaGenome 2 model, the successor to last year's AlphaGenome. It predicts how genetic variants affect specific cell types and states, and identifies nearly twice as many variants that change gene expression as the previous model.
Google DeepMind has announced the AlphaGenome 2 model, the successor to the AlphaGenome genomic model released last year. While nearly all cells in the body carry the same DNA, different cell types switch different genes on and off depending on what they are doing at the time — so the same genetic variant can have a different effect in different cells. AlphaGenome 2 is designed to capture these differences and predict the effect of a specific variant at the level of individual cell types and states, which according to the company provides a more detailed view than broader models such as the original AlphaGenome.
The model was trained on more than 100 million single-cell profiles across 68 tissues. According to Google DeepMind, this turns static cell atlases (large databases mapping individual cells) into a "predictive tool for human genetics." The company states that AlphaGenome 2 identifies nearly twice as many variants affecting gene expression as the previous best model. According to the company, researchers found that the model can make predictions even for new cell types and biological states without the need for retraining, and it can also flag gaps in experimental data, which can help scientists decide what to investigate next in the lab.
However, more detailed predictions are computationally more expensive than with the original AlphaGenome, according to the company's genomics lead Žiga Avsec, which is why the new model does not replace the older version — that one continues to be used, for example, for the genome atlas covering every possible DNA letter change in the human genome, released in September. Since its release, AlphaGenome 2 has been available via the AlphaGenome API for non-commercial research; for commercial use, only a limited "private preview" is running for now.
The company notes that the emergence of such a model was made possible only by new technologies capturing individual cells and their states at high resolution, which enabled the creation of the necessary large-scale datasets. Going forward, the company wants to train models across more species, and in the long term sees follow-up models as part of an effort to create "virtual cells" — digital simulations of biological systems.
Why it matters
For scientists and researchers in genomics and biomedicine, the model offers a more specific tool for determining how genetic variants identified in large populations affect specific cells and potentially relate to diseases. For pharmaceutical research, the granularity at the cell-type level is particularly relevant — drug development targets specific proteins and specific cell types, so more accurate predictions can narrow the range of experiments that need to be carried out in the lab.
Release card
AlphaGenome 2
Google DeepMind
- Inputs
- The input is a DNA sequence and genetic variants, the output is predictions of variant effects on gene expression in individual cell types and states.
- Availability
- Available via the AlphaGenome API for non-commercial research; for commercial use, only a limited private preview is available for now.
- Identifikace variant měnících expresi genů téměř dvakrát více než dosavadní nejlepší model The company says the model captures more genetic variants affecting gene expression than the previous best model.
- Prediction of the effect of a specific genetic variant on individual cell types and states
- Predictions for new cell types and biological states without the need for retraining
- Identification of gaps in experimental data as guidance for further laboratory research
- More detailed predictions are computationally more expensive, so the model does not replace the original AlphaGenome
- Not used for creating a large-scale genome atlas covering every possible DNA letter change, where the older model continues to be used
- For commercial use, it is available only in a limited private preview, not fully
According to the company, it provides a more detailed view at the level of specific cell types and states than broader models such as the original AlphaGenome.
The card summarizes information from the article and any dated corrections, with a link to the original source. It is not our assessment of the model. It does not yet have a dedicated editorial profile. Model selection and other announcements →
What was added since the original report
Verified updates
-
Trained on more than 100 million single-cell profiles; Covers 68 tissues; Identifies approximately 2x more genetic variants affecting gene expression; Can make predictions on new cell types without the need for retraining; Identifies gaps in experimental data
- Trained on more than 100 million single-cell profiles
- Covers 68 tissues
- Identifies approximately 2x more genetic variants affecting gene expression
- Can make predictions on new cell types without the need for retraining
- Identifies gaps in experimental data
Relevant practical impact
What this means
For a business
Companies in genomics, biotechnology, and pharmaceutical research can use the model to identify variants affecting gene expression in specific cell types, which may narrow and reduce the cost of the experimental phase of disease research and drug development. Non-commercial access via the API is available immediately; commercial use is for now limited to a closed private preview.
DevelopmentCheck the original
Event sources
independently confirmed · 2 publishers, 1 independent. We count feeds from the same owner only once.