DeepMind Launches AlphaGenome 2, Predicting DNA Variant Effects at Single-Cell Level

Google DeepMind released AlphaGenome 2, its next-generation genomics model, on November 11. It predicts how DNA sequences and their variants affect gene expression across hundreds of specific human cell types, nearly doubling the previous generation's ability to identify disease-related mutations, and is seen as a key tool bridging single-cell biology and human genetics.

Confirmed

• The model's predictions are conditioned on cell type and state, covering hundreds of human cell types (e.g., cardiac muscle cells, immune cells). This explains why a single base mutation affects only some cells when the heart, brain, and immune cells share the same DNA sequence
• Training data covers 68 tissues and over 100 million profiles; rather than broad tissue-level analysis (which DeepMind likens to a "blurry image" that loses information), the model focuses on one cell type at a time
• It identifies nearly twice as many mutations that turn genes on or off as the previous-generation model
• It can adapt to lab data without retraining: researchers only need to input data from a few dozen cells to observe how DNA changes play out in concrete scenarios such as vaccine immune responses, intestinal inflammatory diseases, and stem cell differentiation

Why it matters

• Many diseases' genetic clues hide within specific cell types, which broad-spectrum analysis struggles to capture; cell-type-conditioned predictions make pinpointing disease-causing mutations far more precise
• Fine-tuning to new data from small numbers of cells without retraining greatly lowers the barrier for labs to apply the model to their own research (e.g., immunology, inflammation, stem cell work)

sourcerefs corresponds to the original post from Google DeepMind's official account.

2026-11-11 ~ 2026-11-12 · 14 related posts

• [source] DeepMind launches AlphaGenome 2, spotting nearly 2x more gene-regulating mutations — GoogleDeepMind · 2026-11-11
• [source] AlphaGenome 2: trained on 100M+ profiles across 68 tissues, 2x mutation detection — GoogleDeepMind · 2026-11-11
• AlphaGenome 2 adapts to lab data on the fly with just a few dozen cells — GoogleDeepMind · 2026-11-11
• AlphaGenome 2 details: zooming from tissue averages to single cell types and states — Dr_Singularity · 2026-11-12
• Google launches AlphaGenome 2 to predict how DNA variants affect specific cell types — sundarpichai · 2026-11-12
• AlphaGenome 2 now available via API for non-commercial research — sundarpichai · 2026-11-12

8 near-duplicate retellings: GoogleDeepMind · pushmeet · TorturedPoet30 · anshulkundaje · pushmeet · Dr_Singularity · anshulkundaje · demishassabis