Google DeepMind Unveils AlphaGenome 2 with Near-Doubled Single-Cell Variant Detection
Google DeepMind has released AlphaGenome 2, a next-generation genomic AI model built on a conditional architecture that takes tissue, cell type and cell state as inputs, pushing prediction from tissue-level to single-cell resolution. Trained on over 107 million single-cell profiles, it nearly doubles the ability to detect gene-expression-related variants, identifying 30% of causal disease variants in an immune-cell benchmark versus 16% for its predecessor. It adds DNA methylation prediction and zero-shot generalization to unseen cell types, and is available via API for non-commercial research.
Related Entities
Full Report
IT之家 November 12 news — On November 11 local time, Google DeepMind announced the launch of its next-generation genomic AI model AlphaGenome 2, which can predict the impact of DNA sequences and their variants on specific human cell types and cell states, helping researchers identify disease-related genetic variants at the cellular level.
According to the introduction, most cells in the human body possess the same DNA, but different cell types and states turn different genes on or off to carry out their respective biological functions. For example, the gene expression patterns of cardiomyocytes and immune cells responding to an infection differ markedly. The same DNA variant may also exert different effects in different cells.
Google stated that the previously released AlphaGenome mainly predicted the impact of DNA variants on the overall molecular activity of different tissues, making it difficult to adequately identify subtle differences in specific cell types. AlphaGenome 2, by contrast, adopts a new conditional model architecture that takes the target tissue, cell type, or cell state as input, enabling the model to make predictions for specific biological conditions.
In terms of training data, AlphaGenome 2 used over 107 million single-cell atlases from human and mouse samples, covering 68 tissues. Compared with the previous-generation model, the new model can identify nearly twice as many genetic variants related to gene expression.
In a test on the TenK10K dataset covering 28 primary immune cell types, AlphaGenome 2 identified 30% of causal disease variants, compared with 16% for the previous-generation model after fine-tuning on single-cell data. In predicting the direction of gene expression changes, the new model achieved an accurate identification rate of 43%, compared with 23% for the previous-generation model.
Beyond gene expression, AlphaGenome 2 can also predict molecular indicators such as histone modifications, chromatin accessibility, and DNA methylation. Among these, DNA methylation is a prediction indicator not yet supported by the previous-generation model and one that researchers had repeatedly requested.
The new model also supports analysis of previously unseen cell types and biological states without retraining or fine-tuning. Researchers can upload a small amount of single-cell RNA sequencing data, allowing the model to generate corresponding predictions based on the gene activity characteristics of the cells, and to analyze cellular changes caused by factors such as disease progression, drug treatment, or vaccination.
DeepMind researchers gave an example: in a previously unseen dataset from a study of an mRNA COVID-19 vaccine, AlphaGenome 2 was able to distinguish vaccine-activated immune T cells from resting ones, and predicted that a genetic variant associated with severe COVID-19 would have a more pronounced effect on gene activity when immune cells are activated.
IT之家 noted that AlphaGenome 2 also supports cross-modal data inference. When researchers provide only single-cell gene activity data for a certain cell type, the model can further predict indicators that have not yet been measured, such as chromatin accessibility, providing a reference for subsequent experiments.
In a real-world research case, the model was able to computationally reproduce part of the findings of a previous inflammatory bowel disease study, identifying genetic variants that affect the activity of the ETS2 gene in intestinal macrophages. Such analyses help researchers link disease-related genetic variants to specific cell types and their target genes.
Currently, AlphaGenome 2 is open to non-commercial research through the AlphaGenome API, while commercial research users can apply for access through a small-scale private preview program. Google DeepMind emphasized that the model is intended for scientific research only and has not been validated or approved for clinical or diagnostic use.