Google DeepMind releases AlphaGenome 2, using 107 million single-cell data to predict the impact of DNA variants
Google DeepMind has launched a new-generation DNA sequence model, AlphaGenome 2, which can predict how the same DNA sequence is read in different cell types, as well as the impact of a single base change on specific cells. The model was trained on more than 107 million single-cell data points, covering 68 tissues, and adds DNA methylation prediction. On the TenK10K dataset, the proportion of pathogenic causal variants it identified reached 30%, up from 16% for the previous generation; its accuracy in identifying the direction of gene expression changes was 43%, up from 23% for the previous generation. The model has been opened to non-commercial research through the AlphaGenome API, and commercial users can apply for a private preview version.
Why read this
AI score 69
The original text provides a comparison of the detection rates of the two generations of models on the same dataset, which can be used to assess the improvement in localizing genetic variants achieved by single-cell-level prediction.
Topics
Body · AI translation
Smart Things news on November 12: Yesterday, Google DeepMind launched AlphaGenome 2, a new-generation DNA sequence model. The model can predict how the same DNA sequence is read in different cells, and what effect a single DNA base change will have on a specific cell, helping scientists understand the genetic mechanisms by which diseases occur.
Although cells in the human heart, brain, and immune system possess the same DNA sequence, a change in a single base may trigger disease only in certain cells without affecting others.
Compared with the previous-generation model, AlphaGenome 2 has a redesigned model architecture. While retaining the core DNA function experimental metrics of the previous-generation model, it also adds DNA methylation prediction.
Google says AlphaGenome 2 can combine single-cell biology with human genetics, predicting the effects of genetic variants across hundreds of human cell types by analyzing how different cells read the same DNA sequence in different ways.
On the social platform X, the release of AlphaGenome 2 also sparked discussion among netizens about the prospects of AI medical applications.
One netizen said: "We really have hope of solving health problems in our lifetime."
Another netizen mentioned: "Since March 2025, I have been following a longevity plan, hoping to live past 100. I really hope AI can help achieve this goal. Deep research into individual cells is exactly the research direction I have been looking forward to."
Another netizen joked with a photo of Anthropic co-founder and CEO Dario Amodei: "This is how Dario reacted when he saw the release of AlphaGenome 2 after people mocked him for saying cancer will be cured."
The model was trained on more than 100 million single-cell data points, covering 68 tissues. Compared with the previous-generation model, AlphaGenome 2 identified nearly twice as many genetic variants capable of altering cellular gene expression.
▲Schematic diagram of the AlphaGenome 2 model, source: Google DeepMind
Currently, AlphaGenome 2 is open to non-commercial research through the AlphaGenome API, and commercial research users can apply for a private preview.
1. Second-generation model architecture redesigned, prediction scope refined from tissues to individual cells
Human cells are the basic units of life. Although every cell in the human body carries the same DNA, cells of different types and states selectively turn different genes on or off according to their own functions.
At present, how to predict how subtle DNA variants affect gene expression in different cells remains an unresolved challenge.
In June last year, Google launched the first-generation DNA sequence model AlphaGenome to predict the effects of DNA variants on molecular activity in different tissues. The model can accept DNA sequences of 1 million letters in length and predict results from thousands of experiments across multiple tissues in parallel at once.
Although this approach is relatively efficient, the model's computational power must be spread across a large number of prediction tasks, making it difficult to capture subtle differences between closely related cell types.
The second-generation DNA sequence model AlphaGenome 2 launched this time has redesigned the model architecture, further refining the prediction scope from average data across broad tissues to individual cell types and cell states, thereby significantly improving performance in predicting the effects of genetic variants.
In addition to inputting DNA sequences of 1 million letters, the model also receives target biological conditions as input at various stages of the architecture, thereby learning how the same DNA sequence is read in different cell types.
▲Overall architecture and core capabilities of AlphaGenome 2, source: Google DeepMind
AlphaGenome 2 also supports cross-modal data completion and can predict missing data in experiments. Researchers only need to input a single measurement of single-cell gene activity, and the model can, on the basis of identifying the cell's biological state, further predict properties not yet measured, such as chromatin accessibility, helping researchers determine which experiments to conduct next.
In terms of training, AlphaGenome 2 used bulk tissue data and single-cell data from human and mouse samples, including more than 107 million single-cell data points.
This means the model is no longer limited to predicting tissue-wide averages like the previous-generation model, but can identify activity inside specific cells, thereby discovering important DNA variants that act only in specific cell types.
In terms of prediction content, AlphaGenome 2 retains the core DNA function experimental metrics of the previous-generation model, including gene expression, histone modification, and chromatin accessibility, while adding DNA methylation prediction, that is, using an AI model to predict which positions in a DNA sequence may undergo methylation and how different DNA variants or cell states affect methylation levels.
Sarah Teichmann, Professor of Stem Cell Medicine at the University of Cambridge and co-founder of the Human Cell Atlas consortium, said that AlphaGenome 2 can directly adapt to new single-cell data and, without retraining, predict the effects of DNA variants in newly discovered cell states based on a small number of cell samples.
2. Genetic variant identification capability nearly doubled, locating disease-related genes
Google says that AlphaGenome 2 has made significant progress in predicting gene regulation in specific cells, which can help researchers more accurately identify genetic variants associated with disease.
First, in terms of cell type identification, even when faced with genomic regions outside the training data, AlphaGenome 2 can more accurately explain gene expression differences between different cell types within the same tissue.
For example, the model can track the continuous changes in gene expression as human hematopoietic stem cells gradually develop into red blood cells and different immune cell lineages.
Second, in terms of predicting the effects of genetic variants, AlphaGenome 2 has also achieved better performance.
On the TenK10K dataset, which contains a large-scale single-cell atlas of 28 types of human primary immune cells, compared with the first-generation AlphaGenome fine-tuned using single-cell data and other specialized tools, the number of pathogenic causal variants identified by AlphaGenome 2 nearly doubled, reaching 30%, compared with 16% for the original model.
In addition to overall performance testing, the research team also evaluated AlphaGenome 2's ability to predict the effects of genetic variants in different cell types through multiple datasets and specific gene loci. The tests covered changes in gene expression, determination of variant causality, and cell type-specific analysis, and demonstrated through specific gene loci how the model explains the effects of different variants.
In terms of identifying the direction of gene expression changes, AlphaGenome 2 achieved an accuracy of 43%, compared with 23% for the original model. The model can not only identify which genetic variants may have an effect, but also more accurately predict whether these variants will enhance or weaken the activity of the relevant genes.
▲AlphaGenome 2 prediction results for gene expression and genetic variant effects in different tissues and cell types, Source: Google DeepMind
In addition, AlphaGenome 2 can also locate genetic variants associated with disease.
Google mentioned that a recent important study published in Nature found that a genetic variant in a long-distance regulatory switch triggers inflammatory bowel disease (IBD) by overactivating the ETS2 gene in intestinal immune cells.
AlphaGenome 2 made predictions for different immune cell types in the intestine and reproduced this research finding at the computational level. The model not only identified the genetic variant that causes this phenomenon, but also found that the variant selectively enhances ETS2 gene activity in intestinal macrophages, while its effects on other cell types and genes are relatively limited.
3. The second generation does not require retraining and can predict previously unseen cell states
In addition to improving the prediction capability for existing cell types, AlphaGenome 2 can also transfer prediction capability to entirely new single-cell experimental data without retraining the model for the new data.
Researchers can upload single-cell RNA sequencing data, even if the data come from only a few dozen cells. The model first compresses the gene activity information of these cells into a cell state representation, and then uses this representation as a condition to predict previously unseen cell states.
Google says that AlphaGenome 2's performance on these new data can match or exceed models trained or fine-tuned directly on the corresponding data.
Researchers can then use AlphaGenome 2 to predict new cell states that appear after disease or vaccination, as well as the effects of genetic variants on gene regulation in these states.
Google cited a study on COVID-19 mRNA vaccines as an example. Faced with single-cell data that the model had not encountered before, AlphaGenome 2 successfully distinguished immune T cells in a vaccine-activated state from immune T cells in a resting state.
To further verify the model's generalization ability for unseen cell states, the research team conducted tests in peripheral blood mononuclear cells (PBMCs) after vaccination and in a cohort of Parkinson's disease patients. The results showed that AlphaGenome 2 can use a small amount of single-cell data to establish new cell state representations and further predict molecular activities such as chromatin accessibility and gene expression.
▲AlphaGenome 2 extends to unseen cell states through expression conditioning, Source: Google DeepMind
The model also found that a genetic variant associated with severe COVID-19 infection interferes more strongly with gene activity when immune cells are in an activated state. If only inactive cells were studied, this effect might be missed.
4. Connecting human genetics with single-cell research to help screen disease targets
To better understand disease, researchers need to connect human genetics research with cell biology.
One important task is to match the large number of trait- and disease-associated genetic variants discovered by genome-wide association studies (GWAS) with the expanding human single-cell atlas.
GWAS can identify DNA regions associated with disease, but researchers often find it difficult to further determine which specific genetic variant is responsible, which target gene it affects, and in which cell type this effect occurs.
To help researchers establish these connections while reducing the burden of large-scale computational tasks, Google precomputed AlphaGenome 2's predictions for trait-associated genetic variants across hundreds of human cell and tissue conditions, focusing on solving two problems.
First, mapping enhancer–gene relationships in real human cells.
Many disease-related genetic variants are located in regulatory regions far from their target genes; these regions are called enhancers. Enhancers can regulate the activity of target genes, but because the two may be far apart, it is difficult for researchers to trace the regulatory relationships between them.
Currently, most experimental CRISPR screens use stable cell lines cultured in the laboratory, so scientists often lack maps of gene regulatory relationships in specific human cell types.
AlphaGenome 2 extends the mapping of enhancer–gene relationships to human primary cell types, linking genetic variants within regulatory regions to the target genes they act on.
In the analysis of enhancer–target gene associations, the research team evaluated the consistency between model predictions and experimental data, and expanded the scope of analysis to 640 biological conditions to further compare gene regulatory relationships across different cell types. Through specific gene loci such as BCL11A, the model demonstrated the correspondence between predictions and experimental validation.
▲AlphaGenome 2 predicts regulatory relationships between enhancers and target genes. Source: Google DeepMind
Second, pinpointing causal variants in GWAS studies and the cells in which they act.
GWAS can flag large DNA regions associated with disease, but researchers may not be able to determine with high confidence the specific causal variants, target genes, and cells of action within them.
For a variety of complex traits ranging from height to type 2 diabetes, the genetic variant scores precomputed by AlphaGenome 2 can help researchers screen for the specific variants and cell types most likely to be involved in the formation of the relevant traits and diseases.
When the model has high confidence, it can also accurately predict whether a given genetic variant will enhance or reduce the activity of a gene. This provides important clues for researchers to understand how the relevant genes are involved in disease occurrence.
In the analysis of genetic variants associated with complex traits and diseases, the research team further scored non-coding variants for multiple traits and, combined with predictions across different cell types, screened for genetic variants that may influence disease occurrence. The model also demonstrated, through specific gene loci, the analytical process from genetic variant to target gene, cell type, and potential mechanism of action.
English text is being prepared.
Peter Donnelly, CEO of genomics company Genomics Limited, said that human genetics is one of the most powerful tools in drug discovery, with GWAS playing a central role, but going from discovering genetic association signals to identifying the specific targets needed for drug discovery remains difficult. AlphaGenome 2 is an important advance in this field, and the company looks forward to integrating it into its Mystra platform.
Currently, AlphaGenome 2 and an interactive Colab tutorial for setting expression conditions are available to non-commercial research through the AlphaGenome API, and commercial research users can apply for a private preview.
Google also plans to launch AlphaGenome 2 Skills in its agent development platform Google Antigravity. The company hopes to combine specialized models such as AlphaGenome 2 with agentic scientific tools such as Gemini 4 in Antigravity to help scientists accelerate research and discover new scientific results.
Conclusion: From genomic prediction to cell-level analysis
From AlphaGenome to AlphaGenome 2, Google DeepMind is advancing DNA sequence analysis from tissue-level overall prediction to fine-grained analysis at the single-cell level.
From an industry application perspective, the value of AlphaGenome 2 lies not only in improving the accuracy of DNA variant prediction, but also in helping AI-driven drug discovery further translate genetics research findings into verifiable drug discovery leads.
Source: Google DeepMind