AlphaGenome 2 predicts the effect of DNA cell by cell
· BehindLines newsroom
On 11 November, Google DeepMind introduced AlphaGenome 2, an AI model that estimates how a variation in DNA changes gene activity in a specific cell type. According to the company's announcement, it succeeds AlphaGenome, which mostly reasoned on averages across whole tissues.
The problem it targets
Almost every cell in the body carries the same genetic material (DNA), but a heart muscle cell, a skin cell and an immune cell do not switch on the same genes. The same genetic variation can therefore have very different effects from one cell to another. According to The Verge, the new model aims to capture these differences.
What changes
AlphaGenome 2 is a so-called conditional model: it receives a DNA sequence of one million letters plus the targeted cell type or state, then predicts the effects for that single condition. According to DeepMind, it was trained on more than 107 million individual cell profiles, from humans and mice, covering 68 tissues. It also predicts DNA methylation, a readout researchers had requested.
According to DeepMind, it spots almost double the number of variants altering how genes are expressed in cells, compared with its previous best model. On data from a Covid-19 vaccine study based on messenger RNA, it told apart the T cells activated by the vaccine from resting ones.
The model can also adapt to new cell types without retraining, The Verge reports, and spot gaps in experimental data, which can help choose the next experiments.
Limits and access
The detailed predictions cost more computing than those of the first model, which is therefore not being replaced, a DeepMind researcher explains to The Verge. AlphaGenome 2 is available for non-commercial research through the AlphaGenome interface, with a restricted early-access preview for commercial research. DeepMind says it wants to generate more data, notably for poorly represented cell states, and eventually considers training its models on several species.
Why it matters
Linking a genetic variation to a specific cell type may help explain the origin of some diseases, since drugs target particular proteins and cells. The announced performance is the company's own, and it states that the model is for research use only and has not been validated or approved for clinical or diagnostic purposes.