DeepMind launches AlphaGenome 2: cell-type-specific DNA variant-effect model nearly doubles causal-variant recall (16%→30%) on TenK10K

Why it matters

【Key points】

· DeepMind posted AlphaGenome 2 on 2026-11-11 at 15:00 UTC, with a paper.

· It takes a ~1-million-base DNA sequence plus a cell type or state, and predicts expression, chromatin accessibility and variant effects, including unseen states.

· Trained on bulk and single-cell human and mouse data (>107 million cell profiles, encoder trained on >100 million profiles across 68 tissues).

• · On TenK10K (28 primary immune cell types), causal-variant recall versus single-cell-finetuned AG1 nearly doubles (16%→30%), as does sign prediction (23%→43%).
• These recalls are at 90% precision; causality AUPRC is 0.76 vs 0.72 for AG1-sc and 0.69 for Scooby.

· New versus AG1: DNA methylation, plus a precomputed atlas of enhancer–gene links and scores for ~7 million common variants (640 primary human conditions).

• · Non-commercial API is open today; commercial research is in private preview.
• Weights are promised at final publication, not open today.
• Research use only; not approved for clinical or diagnostic use.

【Our read】

· Background: a DeepMind sequence-to-function model, not a general LLM and not a clinical product. AG1 mostly predicted variant effects on bulk averages; AG2 conditions on specific cell types and states.

· Comparison: “nearly twice” is a high-precision recall gain (16%→30% / 23%→43% at 90% precision), not a doubling of overall accuracy (AUPRC 0.72→0.76). Different class from Carbon-X or Omnii-1.

· Impact: useful for mapping GWAS loci to putative causal variants, genes, cell types and effect direction, but research-only. Watch the commercial preview, promised weights, and replication.

【Verifiable】

• ·
• ·
• ·