Article · AI translation
IT Home reported on November 12 that Google DeepMind announced the launch of its next-generation genomic AI model, AlphaGenome 2, on November 11 local time. The model can predict how DNA sequences and their variants affect 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 have the same DNA, but cells of different types and states turn different genes on or off to perform their respective biological functions. For example, heart muscle cells and immune cells responding to an infection have markedly different gene expression patterns. The same DNA variant may also have different effects in different cells.
Google said the previously released AlphaGenome primarily predicted the effects of DNA variants on molecular activity across entire tissues, making it difficult to fully identify subtle differences in specific cell types. AlphaGenome 2 uses a new conditional model architecture that takes a target tissue, cell type or cell state as input, enabling predictions for specific biological conditions.
For training data, AlphaGenome 2 used more than 107 million single-cell profiles 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 associated with gene expression.
In a test using 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. For predicting the direction of changes in gene expression, the new model correctly identified 43%, compared with 23% for the previous-generation model.
In addition to gene expression, AlphaGenome 2 can predict molecular measures such as histone modifications, chromatin accessibility and DNA methylation. DNA methylation is a measure that the previous-generation model did not support and that researchers had repeatedly asked to be able to predict.
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 cells' gene activity patterns and analyze cellular changes caused by factors such as disease progression, drug treatment or vaccination.
DeepMind researchers gave the example of data from a previously unseen study of a COVID-19 mRNA vaccine. AlphaGenome 2 could distinguish immune T cells in vaccine-activated and resting states, and predicted that a genetic variant associated with severe COVID-19 would have a more pronounced effect on gene activity when the immune cells were activated.
IT Home noted that AlphaGenome 2 also supports inference across data modalities. When researchers provide only single-cell gene activity data for a particular cell type, the model can further predict unmeasured measures such as chromatin accessibility, providing a reference for subsequent experiments.
In a real-world research case, the model was able to computationally reproduce some findings from an earlier inflammatory bowel disease study, identifying genetic variants that affect the activity of the ETS2 gene in intestinal macrophages. This type of analysis helps researchers link disease-related genetic variants to specific cell types and the genes they affect.
AlphaGenome 2 is currently available for noncommercial research through the AlphaGenome API, while commercial research users can apply for access through a limited private preview program. Google DeepMind emphasized that the model is intended only for scientific research and has not yet been validated or approved for clinical or diagnostic use.