DeepMind Releases AlphaGenome 2 Doubling Disease Variant Detection

IT Home November 11 news, Google DeepMind has released a new-generation DNA sequence model AlphaGenome 2, which can predict the impact of DNA sequences and their variants on specific human cell types, providing researchers with a brand-new tool for uncovering the cellular drivers of disease.

Every cell in the human body carries the same DNA blueprint, but different cell types achieve specific functions by switching different genes on and off. Global human cell atlas projects have mapped gene expression across millions of cells, but predicting how subtle DNA variants affect cellular gene expression remains an unsolved challenge.

Google previously launched the first-generation AlphaGenome, which can predict the impact of DNA variants on molecular activity across broad tissues, and introduced the AlphaGenome Atlas, covering predictions for all 9 billion single-letter variants in the human genome, to support research on rare diseases, complex traits, and basic biology.

AlphaGenome 2 narrows the focus from broad-tissue averages down to individual cell types and states, and reconstructs the model architecture, achieving a step-change improvement in variant prediction performance. It can capture variant effects in disease-relevant cells that are missed by broad-tissue models, and can also make predictions for cell and disease states the model has never seen.

According to IT Home, compared with the previous best model, AlphaGenome 2 can identify nearly twice as many variants that alter gene expression within cells. The model was trained on more than 100 million single-cell atlases across 68 tissue categories, turning static cell atlases into a predictive engine for human genetics. The model is currently available for non-commercial research through the AlphaGenome API, while commercial research is in private preview.

The first-generation AlphaGenome can read DNA sequences up to one million letters in length at once and simultaneously predict relevant properties across multiple tissue types and thousands of experiments. This parallel approach is highly efficient, but it disperses the model's computational capacity, making it difficult to capture subtle differences between closely related cell types.

Google fundamentally restructured the model's logic for handling cell types. AlphaGenome 2 no longer adopts a parallel prediction mode; instead, it is a conditional model that predicts DNA effects for only one condition (tissue, cell type, or cell state) at a time. In addition to DNA sequences up to one million letters in length, the model takes the target biological condition as input at every stage of the architecture, enabling it to learn how the same stretch of DNA is read differently across cell types.

The conditional architecture also supports cross-modal data imputation, allowing prediction of missing experimental data. The model separates a cell's biological state from its output results. If researchers input only single-modality measurements such as single-cell gene activity, AlphaGenome 2 can predict unmeasured properties such as chromatin accessibility, helping researchers plan the direction of follow-up experiments.

AlphaGenome 2 was trained on both human and mouse tissue-block and single-cell data, incorporating more than 107 million single-cell atlases. Moving beyond the tissue-block averages of the first-generation model, it can resolve activity within specialized cells such as cardiomyocytes and blood immune cells, and uncover key DNA variants that act only in specific cell types.

The model retains the core DNA function experimental readouts supported by the first-generation model, such as gene expression, histone modifications, and chromatin accessibility, and adds the highly demanded DNA methylation prediction capability.

AlphaGenome 2 shows significant performance improvements in predicting gene regulation in specific cells, helping researchers precisely pinpoint disease-related variants.

• Improved cell specificity: In genomic regions not covered by training data, AlphaGenome 2 can better explain gene expression differences among different cell types within the same tissue, and can track the continuous changes in gene expression as human hematopoietic stem cells mature into red blood cells and immune cell lineages.
• Improved variant effect prediction: In tests on the large TenK10K single-cell dataset covering 28 types of primary immune cells, AlphaGenome 2 identified causal variants at a rate of 30% and judged the direction of gene effects with 43% accuracy, nearly double the 16% and 23% of the first-generation AlphaGenome, helping researchers discover previously overlooked candidate variants.
• Precise pinpointing of disease-related variants: A previous landmark study found that a variant in a distal regulatory sequence induces inflammatory bowel disease by overactivating the ETS2 gene in intestinal immune macrophages. AlphaGenome 2 computationally reproduced this conclusion, precisely pinpointing the causal variant and demonstrating that it selectively increases ETS2 gene activity only in intestinal macrophages, with almost no effect on other cell types or genes.

AlphaGenome 2 can generalize to entirely new single-cell experimental scenarios without retraining. Researchers only need to upload single-cell RNA sequencing gene activity snapshots from a few dozen cells, and the model can compress them into a cell-state representation and use it as a conditional input to handle never-before-seen cell states, with performance comparable to or even exceeding models trained or fine-tuned directly on the corresponding datasets.

This capability allows AlphaGenome 2 to predict variant effects and gene regulatory patterns in entirely new cell states in scenarios such as disease and vaccination. For example, in tests on unseen single-cell data for COVID-19 mRNA vaccines, AlphaGenome 2 can distinguish vaccine-activated immune T cells from resting T cells, and also revealed variants associated with severe COVID-19 that interfere more strongly with gene activity in the activated cell state — an effect that would be completely missed in unactivated cells.

Building a link between human genetics and single-cell biology is an important path to understanding disease. A key step is to establish correspondences between the thousands of trait- and disease-related variants identified by genome-wide association studies (GWAS) and the increasingly complete human single-cell atlases.

To save researchers from carrying out large-scale computations themselves, Google has precomputed AlphaGenome 2 predictions for trait-related variants across hundreds of human cell and tissue conditions, addressing two core challenges:

• Enhancer-gene association mapping in real human cell types: A large number of disease-associated variants reside in distal regulatory sequences called enhancers, making their effects difficult to trace. Because most CRISPR experiments use stable laboratory-cultured cell lines, researchers often lack regulatory association maps for specific human cells. AlphaGenome 2 extends enhancer-gene mapping capabilities to primary human cells such as human brain astrocytes, helping researchers pinpoint the correspondence between regulatory variants and target genes in the specific cells where disease arises.
• Precisely pinpointing GWAS causal variants and the cells in which they act: Genome-wide association studies typically flag broad DNA segments associated with a disease, and researchers often cannot localize the specific causal variant, target gene, and cell type of action with high confidence. In analyses of complex traits such as height and type 2 diabetes, AlphaGenome 2's precomputed variant scores can help pinpoint the most relevant specific variants and cell types; when confidence is sufficient, the model can also accurately predict whether a variant upregulates or downregulates gene activity, providing key clues for understanding the pathogenic role of genes.

As researchers map more human atlases at the single-cell level, the performance of AI models will continue to improve. Google is carrying out ongoing collaborations with institutions such as the Wellcome Sanger Institute, planning to analyze a broader range of human cell types, developmental lineages, and perturbed cell states. It looks forward to researchers worldwide leveraging AlphaGenome 2's cell-specific predictive capabilities to uncover how genetic variants act across different cells and to reveal the mechanisms underlying basic biology and disease.

Currently, AlphaGenome 2 and the interactive Colab tutorial on expression conditions are open for non-commercial research through the AlphaGenome API, while commercial research is in private preview. Soon, Google will also launch AlphaGenome 2 Skills in its agent development platform Google Antigravity, combining with agentic scientific tools such as Gemini 4 to help researchers accelerate their research and achieve new discoveries.

Official notice: AlphaGenome and AlphaGenome 2 are for research use only and have not been validated or approved for clinical or diagnostic use.