Google DeepMind's AlphaGenome 2 Predicts Variant Effects by Cell State

Google DeepMind released AlphaGenome 2 on November 11, 2026, a deep-learning model trained on more than 100 million individual cell profiles that predicts how DNA variants affect gene regulation in specific cell types and states. The model is available immediately for non-commercial research, with commercial research access in private preview.

The release targets a blind spot in tissue-wide genomic models. A variant can matter only in a specialized cell or in a cell activated by disease, and that signal gets diluted when a model averages across an entire tissue. AlphaGenome 2 conditions each prediction on a selected tissue, cell type or cell state — including conditions generated from a researcher's own single-cell experiment.

The stakes are practical. Genetic studies routinely nominate long lists of candidate variants behind disease associations, and pharmaceutical programs need to know which variant acts, through which gene, in which cell. AlphaGenome 2 is built to narrow exactly that search.

What changed from AlphaGenome 1?

AlphaGenome 1 predicted molecular activity across broad tissues and generated fixed outputs for bulk tissues and cell lines simultaneously from a one-million-letter DNA sequence. AlphaGenome Atlas, a related resource, offered predictions for all 9 billion possible single-letter changes in the human genome. DeepMind describes AlphaGenome 2 as a separate model release, not an extension of that precomputed atlas.

The architectural difference is conditioning. AlphaGenome 2 evaluates one condition at a time and supplies that condition throughout the network, which lets the model represent how the same DNA segment is read differently across cell types. Its readouts include gene expression, RNA splicing and chromatin accessibility, with CpG DNA methylation added to the original model's capabilities.

The technical paper describes training on human and mouse bulk and single-cell data. Researchers can supply single-cell RNA-seq data — a snapshot of gene activity — from a new experiment, and a cell encoder turns that data into a condition without further fine-tuning. The model can therefore produce predictions for datasets and cell states it has never seen before.

The same design fills in missing experimental readouts. Given gene-expression data alone, AlphaGenome 2 can predict unmeasured properties such as chromatin accessibility, meaning whether stretches of DNA are open to regulation, which helps researchers prioritize subsequent experiments. The trade-off is computational: every condition requires a separate prediction, making the model more intensive to run than fixed-track multitask models.

How strong are the benchmark results?

DeepMind's paper reports the model's performance on TenK10K, a single-cell immune dataset spanning 28 cell types. The numbers measure the problem the conditional architecture was designed to address: regulatory differences between related cells.

• AlphaGenome 2 recovered 30% of putatively causal expression-linked variants at 90% precision, compared with 16% for AlphaGenome 1 fine-tuned on single-cell data.
• It recovered 43% of variant effect directions at 90% accuracy, versus 23% for the fine-tuned predecessor.
• On genome regions excluded from training, it explained a median 34% of within-tissue, cell-type-specific gene-expression variance, against 23% for the same predecessor.

Which disease mechanisms did the model reproduce?

DeepMind reports two research examples of cell-state-specific analysis.

In an inflammatory bowel disease analysis, AlphaGenome 2 flagged rs2836882 from a seven-variant credible set, linked it to ETS2 approximately 290 kilobases away and localized the predicted effect to gut monocytes and macrophages. That reproduces a mechanism previously established experimentally.

In a second example, the model received single-cell data from a COVID-19 mRNA vaccine study that was new to it. It distinguished vaccine-activated T cells from resting T cells and predicted a stronger effect for a variant associated with severe COVID-19 in the activated state. Together, the examples show how a variant's predicted effect can depend on both cell type and current state.

What precomputed resources ship with the model?

DeepMind precomputed two resources for researchers who do not want to run large-scale comparisons themselves.

• One maps enhancer–gene links across 640 primary single-cell reference conditions and 39,323 protein-coding and long non-coding RNA genes. Enhancers are regulatory DNA segments that can act far from their target genes.
• Another scores more than 6.8 million unique common and trait-associated single-letter variants across 510 human conditions, nominating possible causal variants, target genes, effect directions and relevant cell states.

How can researchers access it?

At launch, AlphaGenome 2 and expression-conditioning Colab tutorials were available through the AlphaGenome API for non-commercial research. The official documentation says non-commercial API use is free and query rates vary with demand. Commercial research remains in private preview.

DeepMind also plans AlphaGenome 2 Skills for Google Antigravity, its agentic development platform.

Genomics PLC CEO Peter Donnelly said he looked forward to incorporating AlphaGenome 2 into the company's Mystra platform and described sequence-prediction models as tools for moving from association signals to the information drug-development programs need.

What are the model's limits?

The paper states the constraints directly. Even at high precision, AlphaGenome 2 recovers only a minority of the expression-linked and disease-associated variants that statistical analyses nominate as likely causal. It can underestimate effects outside accessible chromatin, at very distant regulatory elements or through mechanisms acting after transcription.

The training data carries gaps. Transient, stimulated and disease states remain underrepresented in both training and the precomputed variant panel, and fine-grained neuronal subpopulations remain difficult to distinguish.

A deeper limitation follows from the data itself. The condition representations come from observational, steady-state profiles and therefore capture correlations rather than causal relationships. It remains uncertain whether the model can predict responses to gene knockouts, overexpression or dynamic cell-state transitions.

DeepMind restricts both AlphaGenome versions to research use. Neither has been validated or approved for clinical or diagnostic purposes.

Why does this matter for drug development?

The central consequence is a more targeted route from a disease association to a candidate variant, target gene and cell state. AlphaGenome 2 can narrow that search across hundreds of conditions, but its incomplete variant recovery and its correlational view of cell states mean the hypotheses it produces still depend on experimental follow-up. For now, the model accelerates the shortlist, not the verdict.