Predicting the effects of DNA variants by cell type with AlphaGenome 2

By linking DNA with cell states, Google DeepMind’s model predicts variant effects. Available through an API, it is restricted to research.

In immune cells called macrophages, a variant associated with inflammatory bowel disease affects the activity of ETS2, a gene located approximately 290,000 bases away. This mechanism, experimentally described in a study published in Nature in 2024, serves as a retrospective case study for AlphaGenome 2. According to Google DeepMind, its predictions identify the relevant variant, the target gene, and the cellular context in which that gene’s activity increases.

This analysis illustrates the scope of the new genomic prediction model: estimating how a DNA change affects gene regulation in a specific cell type or state. It does not constitute a new experimental discovery about ETS2. The AlphaGenome 2 results are presented in a preprint, and the tool has neither been validated nor approved for clinical or diagnostic use. Adding cellular context to the DNA sequence

The model takes a sequence of approximately one million bases, the species being studied, and a biological condition: a tissue, cell type, or cell state. This condition enters at different levels of its architecture. The first version produced outputs in parallel for a fixed catalog of experiments; the new version processes one condition at a time to better distinguish closely related cells.

Training brings together human and mouse data, including single-cell profiles aggregated by cell type and 107 million human single-cell gene expression profiles. Outputs cover gene expression, RNA splicing, chromatin accessibility, and other molecular properties. DNA methylation, a chemical modification involved in gene regulation, joins the predicted measurements.

Conditioning also supports estimates of missing data. Given a gene expression profile, the model can predict certain properties that have not been measured in the same condition, such as chromatin accessibility. These estimates can guide the selection of experiments to perform; they do not thereby become experimental observations. Recall of 30% at a specified precision threshold

On TenK10K, which includes 28 blood immune cell types, the team evaluates the identification of variants associated with changes in a gene’s expression. At 90% precision, recall rises from 16% for AlphaGenome 1 fine-tuned on single-cell data to 30% for AlphaGenome 2. In other words, the latter recovers a larger share of the reference associations while maintaining that precision threshold in the test.

The protocol compares 6,346 variant–gene associations with a high probability of being causal, established through statistical fine-mapping, against an equal number of controls. Data from this cohort were not used to train the model, according to the authors. Causality used as the reference is therefore a statistical estimate, without experimental validation of every association.

Even at this level of performance, most reference associations are not recovered at the selected threshold. The manuscript notes that some effects are underestimated, particularly in regions of less accessible chromatin or those far from the affected gene. The 30% figure cannot be extended to all variants across the genome. Describing a new state from single-cell measurements

The API documentation distinguishes conditions selected from a catalog from those constructed using the researcher’s data. In the latter case, an encoder transforms single-cell expression measurements into a representation of the state being studied. That representation then guides predictions without further adjustment of the model’s weights.

The announcement describes inputs derived from a few dozen cells. This capability does not mean that an unfamiliar biological state can be inferred from the DNA sequence alone: measurements describing it must be supplied. The authors also emphasize that transient, stimulated, and disease states remain underrepresented. The learned representations are observational, and responses to gene knockout or dynamic transitions between states remain an open question. Research access and precomputed predictions

Access is available through the AlphaGenome API for noncommercial research. A private preview is offered for commercial research. The manuscript states that the model’s code and weights will be released upon final scientific publication; current API access should not be confused with that announced release.

The team has also calculated scores for nearly seven million common variants or variants associated with human traits, across hundreds of cellular conditions. An atlas linking regulatory regions to genes complements this work. These resources help researchers examine a genetic association signal to identify the most plausible candidate variants, target genes, and cell types.