Google DeepMind Releases AlphaGenome 2: Cell-Type-Specific DNA Variant Effect Prediction Trained on 100M+ Single-Cell Profiles, Nearly Doubling Causal Variant Recall, Now on API for Non-Commercial Research

Background and Industry Pain Points

Nearly every cell in the body carries the same DNA, yet heart muscle cells and immune cells switch different genes on and off, so a single-letter variant may cause disease in only one kind of cell. Genome-wide association studies (GWAS) flag DNA regions linked to disease but rarely pin down the causal variant, target gene and cell type. The first AlphaGenome, released last year, predicted averages across broad tissues, so effects confined to a few cell types could be averaged away.

Architecture Highlights and Mechanism

AlphaGenome 2 is a conditional model. Alongside a million-letter DNA sequence it takes one target condition (a tissue, cell type or cell state) and predicts results for that condition only. The condition can be picked from a reference catalog or generated on the fly by a cell encoder from a researcher's own single-cell experiment. Because cell state is modeled separately from outputs, the model can impute missing modalities, for example predicting unmeasured chromatin accessibility from single-cell gene activity. Beyond the first model's gene expression, histone marks and chromatin accessibility, it adds DNA methylation. Training used human and mouse bulk and single-cell data, about 107 million single-cell profiles.

Benchmarks and Evaluation

All figures are vendor-reported. On TenK10K, a single-cell map of 28 primary immune cell types, AlphaGenome 2 recovered 30% of causal variants versus 16% for the original model fine-tuned on single-cell data and other specialized tools, and 43% versus 23% of directional gene effects. DeepMind also says it computationally reproduced a known finding that a distant regulatory variant drives inflammatory bowel disease by overactivating ETS2 in gut macrophages. There is no independent third-party evaluation yet. Project lead Žiga Avsec told The Verge the new model is more expensive to run and will not replace the first one; the AlphaGenome Atlas of 9 billion single-letter variants still uses the original model.

Developer Quick Start

Non-commercial researchers can call AlphaGenome 2 through the AlphaGenome API today, with interactive Colab tutorials for expression conditioning; commercial groups must apply for the private preview. DeepMind has also precomputed scores for GWAS trait variants across hundreds of cell and tissue conditions, so you can look them up without running large computations. AlphaGenome 2 Skills for Google Antigravity are coming soon, so Gemini 4 agents can call the model. It is for research use only and has not been validated for clinical or diagnostic purposes.