Google DeepMind's AlphaGenome 2 Doubles Causal Variant Detection With 107 Million Single Cells

Google DeepMind released AlphaGenome 2 on Tuesday, the second generation of its genomics foundation model, and moved its resolution down from whole tissues to individual cell types and cell states. The company positions it as a research tool for connecting the thousands of variants surfaced by genome-wide association studies to the specific cells where they actually do damage.

The model was trained on more than 107 million single-cell profiles spanning 68 human tissues, combining bulk and single-cell data from human and mouse samples. Where the original AlphaGenome predicted thousands of tissue-level readouts in parallel across a one-million-letter DNA sequence, version 2 is a conditional model: it takes one condition at a time — a tissue, a cell type, or a cell state — and threads that condition through every stage of the architecture.

The measurable claim comes from TenK10K, a dataset covering 28 primary immune cell types. On it, AlphaGenome 2 reliably identifies causal variants at roughly 30 percent versus 16 percent for its predecessor, and directional gene effects at 43 percent versus 23 percent — close to a doubling on both counts. DeepMind is explicit that these are its own evaluation results and that the model is not validated or approved for clinical or diagnostic use.

The second headline capability is generalization without retraining. Researchers can upload single-cell RNA-seq data from as few as a few dozen cells, and the model compresses it into a cell-state representation it can condition on. In a COVID-19 mRNA vaccine example, the model distinguished vaccine-activated T cells from resting ones and found that a variant linked to severe COVID interferes with gene activity far more strongly in the activated state — an effect invisible in bulk analysis.

To spare labs the compute bill, DeepMind has precomputed AlphaGenome 2 scores for trait-associated variants across hundreds of human cell and tissue conditions, supporting enhancer-to-gene mapping in primary cells such as brain astrocytes and causal-variant pinning for traits from height to type 2 diabetes. The model also adds DNA methylation as a new prediction output alongside gene expression, histone marks and chromatin accessibility, and can impute unmeasured modalities from a single input assay.

Availability is split: the model is open today through the AlphaGenome API for non-commercial research, with interactive Colab tutorials, while commercial use sits in private preview. AlphaGenome 2 skills for Google's Antigravity agent environment, combining the model with Gemini 4, are coming soon. Genomics Limited says it plans to integrate the model into its Mystra platform, and Human Cell Atlas co-founder Sarah Teichmann endorsed the cell-level resolution as the right lens for interpreting variant effects.

For a company whose public narrative is dominated by frontier models and data centers, AlphaGenome 2 is a reminder that DeepMind's science track is still shipping. The competition in AI-for-genomics is real — EvolutionaryScale, OpenAI's biology partnerships and academic groups are all in the race — but a free non-commercial API with precomputed GWAS scores is a strong default position for the research community that will actually validate or refute these claims.