IT Home, November 12: On November 11 local time, Google DeepMind announced the launch of its next-generation genomic AI model, AlphaGenome 2, which can predict the effects of DNA sequences and their variants on specific human cell types and cellular states, helping researchers identify disease-related genetic variations at the cellular level.

It is understood that most cells in the human body contain the same DNA, but different types and states of cells activate or silence different genes to carry out their specific biological functions. For example, cardiac muscle cells and immune cells responding to an infection exhibit markedly different gene expression patterns. The same DNA variant may also have different effects in different cell types.
Google stated that the previously released AlphaGenome primarily predicted the impact of DNA variations on overall molecular activity across different tissues, making it difficult to fully capture subtle differences within specific cell types. AlphaGenome 2 introduces a novel conditional model architecture that takes the target tissue, cell type, or cell state as input, enabling the model to make predictions tailored to specific biological conditions.

In terms of training data, AlphaGenome 2 utilizes over 107 million single-cell profiles from human and mouse samples, covering 68 tissue types. Compared to the previous generation, the new model can identify nearly twice as many genetic variants associated with gene expression.
In a test on the TenK10K dataset involving 28 primary immune cell types, AlphaGenome 2 identified causal variants responsible for disease at a rate of 30%, compared to 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 accuracy rate of 43%, compared to 23% for the previous-generation model.
In addition to gene expression, AlphaGenome 2 can predict molecular metrics such as histone modifications, chromatin accessibility, and DNA methylation. Among these, DNA methylation is a predictive metric that was not supported by the previous generation of models and had been frequently requested by researchers.
The new model also supports analyzing previously unseen cell types and biological states without retraining or fine-tuning. Researchers can upload small amounts of single-cell RNA sequencing data, allowing the model to generate predictions based on the gene activity profiles of cells and analyze changes in cells caused by factors such as disease progression, drug treatment, or vaccination.
DeepMind researchers cited an example in which, using previously unseen data from a study on a COVID-19 mRNA vaccine, AlphaGenome 2 distinguished between activated and resting immune T cells and predicted that a genetic variant associated with severe COVID-19 has 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 a specific cell type, the model can further predict unmeasured metrics such as chromatin accessibility, offering valuable insights for subsequent experiments.
In a real-world research case, the model successfully reproduced part of the findings from a prior inflammatory bowel disease study by calculating and predicting genetic variants that influence ETS2 gene activity in intestinal macrophages. Such analyses help researchers link disease-associated genetic variants to specific cell types and their target genes.
Currently, AlphaGenome 2 is open to non-commercial research via the AlphaGenome API, while commercial research users may apply through a limited private preview program. Google DeepMind emphasizes that this model is intended solely for scientific research and has not been validated or approved for clinical or diagnostic use.


