Google DeepMind in the UK has released AlphaGenome 2, a research model that predicts how DNA changes affect particular cell types. It combines DNA sequences with a description of the cell being studied, moving beyond broad tissue averages. Scientists can also supply new single-cell measurements to investigate cell states absent from its training data.
Artificial Intelligence··Night
DNA predictions gain a cell address
Google DeepMind, the UK-based AI research laboratory, announced AlphaGenome 2 on 11 November to predict how genetic changes affect particular human cell types. The model examines the molecular effects of DNA sequences, including regions that regulate other genes. Its new cell-specific approach addresses a difficulty in interpreting disease-linked variants: the same stretch of DNA can act differently in different cells.[1], [2]
The previous AlphaGenome generated predictions across many tissues in parallel from a million-letter DNA sequence. Its successor takes a condition specifying a tissue, cell type or cell state alongside that sequence. The condition enters different stages of the model, allowing it to describe how the sequence is read in a selected biological setting rather than only across a broad tissue average.[1]
New cell measurements can guide the model
Training combined bulk measurements with single-cell data from humans and mice. A researcher can also supply a new snapshot of gene activity from individual cells. A cell encoder compresses that measurement into a representation that conditions the model. DeepMind describes this as a way to generate predictions for newly profiled cell states without retraining the entire network for each experiment.[1], [2]
The predicted readouts include gene activity, the accessibility of DNA’s packaging and changes to proteins associated with DNA. DNA methylation is an added output in this version. DeepMind also describes estimating missing measurements: when a researcher provides gene-activity data, the model can predict another property, such as how accessible the surrounding genetic material is to cellular machinery.[1]
A bowel-disease finding is revisited computationally
One example revisits a variant already linked experimentally to inflammatory bowel disease. It lies far from protein-coding sequences and affects ETS2, a gene that becomes overactive in gut immune cells called macrophages. AlphaGenome 2 reproduced aspects of that result computationally, identifying the relevant variant and cell setting. The underlying biological discovery came from earlier experimental research.[1], [2]
DeepMind has also prepared predictions connecting trait-associated variants with candidate genes and cell types. Non-commercial researchers can use the model through its programming interface, with interactive examples accompanying access. Commercial research is in private preview. DeepMind specifies that AlphaGenome 2 is intended for research and has not been validated or approved for clinical or diagnostic use.[1]
Related columns
For more information on this topic, you can read the related columns.
