AlphaGenome 2 predicts genetic effects at single-cell resolution
Google DeepMind has introduced AlphaGenome 2, a genomics model designed to predict how genetic variants affect particular human cell types. Unlike its predecessor, which assessed sequences across many tissues simultaneously, the new system receives a specified cellular state and calculates the likely effects for that state. It was trained on more than 107 million individual cell profiles from 68 human and mouse tissues.
The model predicts gene activity, DNA packaging in the nucleus and DNA methylation, which can influence whether genes are active or silenced. It can also infer additional properties from limited measurements and adapt to previously unseen experiments without retraining. On the TenK10K benchmark covering 28 primary immune cell types, it identified about 30 percent of causal variants and correctly determined whether a variant increased or reduced gene activity in 43 percent of cases. A fine-tuned AlphaGenome 1 reached 16 and 23 percent, respectively.
AlphaGenome 2 is available through an API for noncommercial research, while commercial access is in closed preview. It is not approved for clinical diagnostics. Google DeepMind plans to add dedicated features to Google Antigravity, alongside tools including Gemini 4.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.