Google DeepMind releases a new generation of DNA sequence models AlphaGenome 2, which can predict the impact of DNA sequences and their variations on specific human cell types, providing researchers with a new tool to uncover the cellular drivers of diseases.

Each cell in the human body carries the same DNA blueprint, but different cell types perform specific functions by turning on or off different genes. Global human cell atlas projects have mapped the gene expression of millions of cells, but predicting how subtle DNA variations affect cell gene expression remains an unsolved challenge.

Previously, Google had launched the first-generation AlphaGenome, which could predict the impact of DNA mutations on the molecular activity of broad organizations. It also introduced AlphaGenome Atlas, which covers the prediction of all 9 billion single-letter mutations in the human genome, facilitating research in rare diseases, complex traits, and basic biology.

AlphaGenome 2: By narrowing down from the average level of broad tissues to a single cell type and state, and by reconstructing the model architecture, the performance of mutation prediction has achieved a leapfrog improvement. It is now capable of capturing the mutation effects in disease-related cells that were overlooked by the broad tissue models, and can also make predictions about cells and disease states that the model has not previously seen.

Compared to the previous best model, AlphaGenome can identify nearly twice as many mutations that can alter gene expression within cells. This model is trained based on over 100 million single-cell profiles from 68 types of tissues, transforming static cell profiles into a predictive engine for human genetics. Currently, this model is available for non-commercial research through AlphaGenome API, while commercial research is still in the private preview phase.

The first-generation AlphaGenome is capable of reading a sequence of DNA length of up to one million letters at once, and can simultaneously predict relevant attributes for multiple types of organizations and thousands of experiments. This parallel approach is quite efficient, but it disperses the model's computing power, making it difficult to capture subtle differences between closely related cell types.

Google has made a fundamental restructure to the logic by which the model processes cell types. AlphaGenome It no longer uses a parallel prediction mode but instead adopts a conditional model that predicts the DNA effect for only one condition at a time (tissue, cell type, or cell state). In addition to DNA sequences that are millions of letters long, the model takes the target biological condition as input at each stage of its architecture, enabling it to learn the differences in reading the same sequence DNA among different cell types.

Conditional architectures also support cross-modal data filling, allowing for the prediction of missing experimental data. The model separates the biological state of cells from the output results. If researchers provide only single-class measurement data such as single-cell gene activity, AlphaGenome, it is possible to predict unmeasured attributes such as chromatin accessibility, helping researchers to plan the direction of subsequent experiments.

AlphaGenome Trained simultaneously using tissue blocks and single-cell data from both humans and mice, incorporating over 107 million single-cell profiles. Beyond the average level of the initial generation of models, this model is capable of analyzing the activity within specialized cells such as cardiac muscle and blood immune cells, and can identify key DNA mutations that are only active in specific cell types.

This model retains the core functional experimental indicators supported by the initial model, such as gene expression, histone modification, and chromatin accessibility, and has also added the much-needed DNA methylation prediction capability.

AlphaGenome 2 has significantly improved performance in gene regulation prediction of specific cells, which can help researchers accurately locate disease-related mutations.

Cell-specific enhancement: In genomic regions not covered by training data, AlphaGenome 2 can better explain the differences in gene expression between different cell types within the same tissue, and is capable of tracking the continuous changes in gene expression during the maturation of human hematopoietic stem cells into red blood cells and immune cell lineages.

Variant effect prediction improvement: In tests on a large TenK10K single-cell dataset that covers 28 types of primary immune cells, AlphaGenome identified causal variants accounting for 30% of the cases, with an accuracy rate of 43% in determining the direction of gene effects. This is nearly double the initial rates of 16% and 23% achieved by AlphaGenome, and can help researchers discover candidate variants that were previously overlooked.

Precise identification of disease-related mutations: A previous milestone study found that a mutation in a remote regulatory sequence could induce inflammatory bowel disease by overactivating the ETS2 gene in intestinal immune macrophages. AlphaGenome replicated this conclusion at the computational level, precisely identifying this causal mutation and proving that it only selectively enhances the activity of the ETS2 gene in intestinal macrophages, with virtually no effect on other cell types or genes.

AlphaGenome It can be generalized to brand-new single-cell experiment scenarios without the need for retraining. Researchers only need to upload single-cell RNA sequencing gene activity snapshots of dozens of cells, and the model can compress them into cell state representations. When used as conditional inputs for never-seen cell states, its performance can rival or even surpass that of models trained or fine-tuned directly on the corresponding datasets.

This capability enables AlphaGenome 2 to predict the variation effects and gene regulation patterns of new cell states in scenarios such as disease prediction and vaccination. For example, in tests with single-cell data for the mRNA vaccine against COVID-19 that have not been seen before, AlphaGenome 2 was able to distinguish between vaccine-activated immune T cells and resting T cells. It also revealed mutations associated with severe cases of COVID-19, which have a stronger interference with gene activity in activated cell states—a effect that would be completely overlooked in unactivated cells.

Establishing a connection between human genetics and single-cell biology is an important path to understanding diseases. A key aspect of this process is to correlate the thousands of traits and disease-related variations identified through whole-genome association studies ( GWAS ) with the increasingly sophisticated human single-cell atlases.

To prevent researchers from conducting large-scale computations on their own, Google has pre-calculated AlphaGenome predictions for hundreds of variations related to traits under various conditions of human cells and tissues, solving two core challenges:

• Enhancer-gene association mapping in real human cell types: A large number of disease-related mutations are located in remote regulatory sequences known as enhancers, making it difficult to trace their effects. Since most CRISPR experiments use stable laboratory-cultured cell lines, researchers often lack regulatory association maps for specific human cells. AlphaGenome extends the enhancer-gene mapping capability to primary human cells such as human brain astrocytes, helping researchers to pinpoint the correspondence between regulatory mutations and target genes in the specific cells where diseases occur.
• Precise identification of causal mutations and their target cells: Whole-genome association studies typically identify broad segments associated with diseases, but researchers often cannot pinpoint specific causal mutations, target genes, and cell types with high confidence. In the analysis of complex traits such as height and type 2 diabetes, pre-computed mutation scores can help identify the most relevant specific mutations and cell types. When there is sufficient confidence, the model can also accurately predict whether mutations upregulate or downregulate gene activity, providing key clues for understanding the pathogenic effects of genes.

As researchers map more human atlases at the single-cell level, the performance of the AI model will continue to improve. Google is collaborating with institutions such as the Wellcome Sanger Institute, planning to analyze a wider range of human cell types, developmental lineages, and perturbed cell states. It is hoped that researchers around the world can leverage the cell-specific prediction capabilities of AlphaGenome 2 to uncover the patterns of gene variations in different cells and reveal the mechanisms linking basic biology to diseases.

Currently, AlphaGenome 2 and the interactive expression condition Colab tutorial are available to non-commercial research through AlphaGenome API. Commercial research is in a private preview phase. Soon after, Google will also launch AlphaGenome 2 within the Agent Development Platform Google Antigravity, in conjunction with intelligent agent science tools such as Gemini 4, to help researchers accelerate their research process and make new discoveries.

Official Note: AlphaGenome and AlphaGenome are intended for research purposes only and have not yet been verified or approved for clinical or diagnostic use.