IT Home On November 12, local time on November 11, Google DeepMind announced the launch of a new generation of genomic AI model, AlphaGenome 2, which can predict the impact of DNA sequences and their variations on specific human cell types and cell states, helping researchers identify genetic variations related to diseases at the cellular level.
According to the introduction, most cells in the human body possess the same DNA, but different types and states of cells turn on or off different genes to perform their respective biological functions. For example, there are significant differences in the gene expression patterns between cardiac muscle cells and immune cells that are responding to infections. The same DNA variation may also have different effects in different cells.
Google said that the previously released AlphaGenome mainly predicted the impact of DNA mutations on the overall molecular activity of different tissues, and it was difficult to fully identify the subtle differences among specific cell types. AlphaGenome 2 uses a completely new conditional model architecture, with target tissues, cell types, or cell states as inputs, enabling the model to make predictions under specific biological conditions.
In terms of training data, AlphaGenome 2 used over 107 million single-cell maps from human and mouse samples, covering 68 types of tissues. Compared to the previous generation model, the new model can identify nearly twice as many genetic variations related to gene expression.
In a test using the TenK10K dataset involving 28 primary immune cell types, AlphaGenome 2 identified disease-causing variant proportions of 30%, while the previous generation model achieved 16% after fine-tuning with single-cell data. When predicting the direction of gene expression changes, the new model had an accurate identification rate of 43%, compared to 23% for the previous generation model.
In addition to 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 that the previous generation models did not support, and which researchers have repeatedly requested.
The new model also supports analyzing cell types and biological states that have not been seen before, without re-training 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 analyze cellular changes caused by factors such as disease progression, drug treatment, or vaccination.
According to researchers at DeepMind, in data from a previously unseen study on the COVID-19 mRNA vaccine, AlphaGenome 2 was able to distinguish between the activated and resting states of immune T cells, and it predicted that genetic variations related to severe COVID-19 would have a more pronounced effect on gene activity when immune cells are activated.
IT Home notes that AlphaGenome 2 also supports cross-modal data inference. When researchers provide only single-cell gene activity data for certain cells, the model can further predict indicators such as chromatin accessibility that have not been measured, providing a reference for subsequent experiments.
In actual research cases, this model can predict and reproduce some of the findings from a previous study on inflammatory bowel disease through calculation, and identify genetic variations that affect the activity of the ETS2 gene in intestinal macrophages. Such analyses help researchers link disease-related genetic variations with specific cell types and their associated genes.
Currently, AlphaGenome 2 has been made available to non-commercial research through the AlphaGenome API, while commercial research users can apply for access through a limited private preview program. Google DeepMind emphasizes that this model is used solely for scientific research and has not been verified or approved for clinical or diagnostic purposes.