Google DeepMind's AlphaGenome 2 Doubles Disease Variant Detection Across 28 Cell Types
Google DeepMind's new conditional DNA model predicts variant effects in individual cell types, doubling causal variant recall and adapting to new single-cell data without retraining.
- Google DeepMind released AlphaGenome 2, a conditional DNA model predicting variant effects per cell type.
- Identifies nearly 2x as many causal variants (30% vs 16%) on the TenK10K immune cell benchmark.
- Trained on 107M+ single-cell profiles across 68 tissues from human and mouse samples.
- Adapts to new single-cell experiments from a few dozen cells without any retraining.
- Adds DNA methylation prediction and supports cross-modality imputation of missing readouts.
- Free via AlphaGenome API for non-commercial research; commercial use in private preview.
Google DeepMind has released AlphaGenome 2, a DNA sequence model that predicts molecular activity in specific cell types and states. Researchers can use it today through the AlphaGenome API for non-commercial work. Commercial access is in private preview, and interactive Colab tutorials are available.
On TenK10K, a single-cell dataset covering 28 primary immune cell types, AlphaGenome 2 identified 30% of causal variants, compared with 16% for the original AlphaGenome after single-cell fine-tuning. It predicted the direction of gene effects in 43% of cases, compared with 23%. DeepMind also evaluated the model against specialist tools built for these tasks.
| Task | AlphaGenome 2 | Fine-tuned AlphaGenome |
|---|---|---|
| Causal variants identified | 30% | 16% |
| Direction of gene effect predicted | 43% | 23% |
In this benchmark, causal variant identification means selecting the genetic change responsible for a measured regulatory effect. Directional prediction means determining whether that change raises or lowers a gene’s activity.
Cell state enters every layer
The original AlphaGenome accepted a DNA sequence of up to one million bases and predicted molecular properties across thousands of experiments and tissues in parallel. Sharing model capacity across those outputs made closely related cell types harder to distinguish.
AlphaGenome 2 conditions each prediction on one tissue, cell type, or cell state. The model receives that biological condition alongside the DNA sequence and incorporates it throughout the architecture. A condition can come from a reference catalog or from a researcher’s own single-cell experiment.
- Sequence input: A DNA region containing up to one million bases.
- Condition input: One selected tissue, cell type, or cell state.
- Custom conditions: A cell encoder compresses a single-cell gene-expression profile into a representation the model can use.
- Execution model: Each prediction targets one biological condition at a time.
Conditioning allows the same DNA sequence to produce different predictions for different cells. That distinction matters because regulatory variants often act only in particular cell types or during transient states such as immune activation.
Unseen states without a training run
Researchers can provide a single-cell RNA sequencing profile from a few dozen cells to define a state that was absent from the training set. The cell encoder converts the observed gene activity into a compact representation, which AlphaGenome 2 uses without dataset-specific retraining or fine-tuning.
In a test using previously unseen data from a COVID-19 mRNA vaccine study, the model distinguished activated immune T cells from resting T cells. It also predicted that a variant associated with severe COVID-19 disrupts gene activity more strongly in the activated state. An analysis based on unactivated cells would miss that state-dependent effect.
A training set of 107 million cells
AlphaGenome 2 was trained on bulk and single-cell data from human and mouse samples, including more than 107 million individual cell profiles across 68 tissues. Bulk measurements average signals from many cells, while the single-cell data helps the model separate specialized populations such as heart muscle cells and immune cells.
Its supported outputs include gene expression, histone modifications, chromatin accessibility, and DNA methylation. The first three were available in the original AlphaGenome; DNA methylation is new in this release.
Because the architecture separates a cell’s biological state from the predicted measurements, it can also estimate missing data across experimental modalities. A researcher might supply single-cell gene expression and ask the model to predict chromatin accessibility for the same state. Those estimates can help prioritize later laboratory experiments, although they do not replace direct measurements.
From association to mechanism
Genome-wide association studies, commonly called GWAS, link genetic variants with diseases or traits. Many reported variants lie in non-coding DNA, including regulatory elements called enhancers. These elements can control distant genes and may operate in only one cell type, making it difficult to identify the relevant gene, cell, and direction of effect.
AlphaGenome 2 supports two common GWAS follow-up tasks:
- Enhancer-to-gene mapping: The model can connect regulatory variants with target genes in primary human cell types, including astrocytes that support and protect neurons.
- Variant prioritization: Precomputed scores can rank variants and implicated cell types for complex traits such as height and type 2 diabetes. High-confidence predictions can also indicate whether a variant increases or decreases gene activity.
In an inflammatory bowel disease example, AlphaGenome 2 reproduced a known mechanism across immune cell types in the gut. It selected the causal variant and predicted that the change boosts ETS2 activity in gut macrophages while leaving other tested genes and cell types largely unaffected.
API access with research-only boundaries
Researchers can access the model through the AlphaGenome API at no cost for non-commercial research. Commercial users can request access to a private preview. DeepMind has also published a technical paper and plans to add AlphaGenome 2 Skills to Google Antigravity for agent-based scientific workflows.
- Clinical status: AlphaGenome and AlphaGenome 2 are research tools. Neither has been validated or approved for diagnosis or clinical decision-making.
- Data coverage: Predictions are strongest for biological conditions well represented in the training data.
- Confidence: Directional variant calls should be interpreted alongside the model’s reported confidence.
- Throughput: Each run targets a single condition, giving researchers finer resolution at the cost of the original model’s broad parallel output.
For functional genomics, GWAS follow-up, and target discovery, the main workflow change is the ability to analyze a new single-cell dataset without fine-tuning a separate model. A few dozen profiled cells can provide enough context to query a variant in a specific state, reducing a retraining process that could take weeks to an API-based prediction workflow.
Comments
