Many MindedAlphaGenome 2 predicts how DNA variants act in individual cell types
Google Deepmind has released AlphaGenome 2, a genomics model that predicts how variants in DNA sequence affect specific human cell types. It was trained on more than 107 million individual cell profiles from 68 tissues in humans and mice. Its predecessor read a DNA sequence one million letters long and predicted properties across thousands of experiments in many tissues at once, a bulk view.
The new version is a conditional model: it takes a specific cell state as an extra input alongside the sequence. Besides gene expression, it predicts DNA accessibility, meaning how the DNA is packaged in the nucleus, and now also DNA methylation. It can infer extra properties from a single measurement, such as packaging when only gene activity was measured. It can also adapt to new experiments without retraining, using gene activity data from a few dozen cells.
On the TenK10K benchmark, covering 28 primary immune cell types, Deepmind reports it identified about 30 percent of causal variants and got the up or down direction right in 43 percent of cases. A fine-tuned AlphaGenome 1 scored 16 and 23 percent. In a computational test it reproduced a known inflammatory bowel disease mechanism, isolating a distant variant that overactivates ETS2 in gut macrophages. On COVID-19 mRNA vaccine data, it showed a severe-COVID-linked variant disrupts activity mainly in activated T cells.
Why it matters: most disease variants sit in non-coding DNA, and knowing which cell type they act in is a bottleneck for turning genetic associations into drug targets. Precomputed predictions for hundreds of cell and tissue conditions could cut that analysis cost for researchers anywhere.
Caveats: the performance figures come from Deepmind and are not independently verified, and the IBD result reproduced an already-known mechanism. The model is not approved for clinical diagnostics. Access is via API for noncommercial research now, with a closed preview for commercial users. Integration into Google Antigravity alongside tools such as Gemini 4 is planned, not released.
Source: The Decoder