--- title: "AlphaGenome 2 launches with cell-specific DNA predictions: benchmarks and research access" description: "AlphaGenome 2 adds cell-specific DNA variant predictions. See GTEx and TenK10K comparisons, research API availability and what the results mean." url: "https://promptszone.com/alphagenome-2-benchmarks-research-access" published: "2026-11-12T03:00:05Z" categories: ["New Models"] --- # AlphaGenome 2 launches with cell-specific DNA predictions: benchmarks and research access Google DeepMind’s AlphaGenome 2 adds cell-specific variant predictions. Compare its GTEx and TenK10K benchmarks, API access and research limits. ![AlphaGenome 2 TenK10K comparison: 30% versus 16% recall at 90% precision; 43% versus 23% recall at 90% sign accuracy, compared with AlphaGenome 1-sc.](https://promptszone.com/assets/uploads/fcc8a33d869b5e9b27ec31d1f60e0b9a.png) **Summary:** AlphaGenome 2 brings cell-context conditioning to DNA variant predictions, with stronger developer-reported recall at fixed quality thresholds and non-commercial research API access. Google DeepMind released **AlphaGenome 2 on November 11, 2026**, extending its DNA model from broad tissue predictions to the effects of genetic variants in specific cell types and biological states. The model is available through the AlphaGenome API for non-commercial research; commercial research access is in private preview. The practical advance is context: the same change in DNA can behave differently in different cells. AlphaGenome 2 conditions its predictions on a biological setting, helping researchers prioritize which variants to investigate and where their effects may appear. ## What changed from AlphaGenome DeepMind says the new model combines bulk genomic measurements with more than 107 million single-cell profiles spanning 68 tissues in human and mouse data. It takes a DNA sequence of up to one million letters together with biological context, rather than relying only on a fixed set of tissue outputs. Researchers can condition the model on new single-cell RNA-sequencing data without retraining it. AlphaGenome 2 also adds DNA methylation predictions, alongside readouts such as gene expression and chromatin activity. These are predictions for research workflows, not measurements of an individual patient. ## How AlphaGenome 2 compares on variant-effect benchmarks The launch paper compares the full-genome distilled model with earlier AlphaGenome variants and specialist models on held-out test chromosomes. The following scorecard separates GTEx tissue-level evaluations from TenK10K single-cell evaluations. All scores are developer-reported; higher is better within each row. ![AlphaGenome 2 benchmark scorecard: GTEx causality AUPRC 0.74 versus AlphaGenome 1 0.72 and Borzoi 0.69; GTEx sign accuracy–recall AUC 0.85 versus 0.83 and 0.77. TenK10K causality AUPRC 0.76 versus AlphaGenome 1-sc 0.72 and Scooby 0.69; sign accuracy–recall AUC 0.85 versus 0.79 and 0.77.](https://editor.blogmaker.app/web/assets/uploads/cfdba4f9f69c1fb780a67dab3ad7a637.png) **AG1-sc is a separate baseline:** the original AlphaGenome fine-tuned on single-cell data. It should not be conflated with the original AlphaGenome baseline in the GTEx rows. Missing cells are not ranked, and the rows are not combined into an overall score. The causality task distinguishes fine-mapped, putatively causal variants from matched negative controls. AUPRC is the area under the precision–recall curve. The sign metric instead measures the area under a sign accuracy–recall curve: how well the model predicts whether a variant increases or decreases expression. It is not a conventional ROC AUC. ## Why the fixed-threshold results matter On TenK10K, covering 6,346 causal single-cell eQTLs across 28 primary immune cell types, AlphaGenome 2 recovered **30% of variants at 90% precision**, compared with 16% for AG1-sc. For effect direction, it reached **43% recall at 90% sign accuracy**, versus 23%. Those figures describe recall at specific quality thresholds, not 30% or 43% overall accuracy. They suggest that a researcher maintaining a high-confidence cutoff could recover more candidates, while still leaving many variants unresolved. The causality evaluation used equal numbers of positive variants and matched negatives; performance in a different screening population may differ. For GTEx across 49 tissues, the paper reports 26% versus 17% recall at 90% precision, and 40% versus 32% recall at 90% sign accuracy, compared with the original AlphaGenome. Those tissue-level results answer a different question from the single-cell comparison and should be read separately. ## Availability and research limits Non-commercial researchers can access AlphaGenome 2 through the AlphaGenome API, subject to its access conditions. Commercial research access is a private preview, not general commercial availability. DeepMind says AlphaGenome 2 Skills for Antigravity are coming soon; they are not part of the currently available release. The model is not validated or approved for clinical or diagnostic use. Its predictions can help prioritize experiments, but do not establish biological causality or replace experimental validation. The most relevant next step for a research team is to test whether its cell types, biological conditions and intended variant-analysis task fit the model’s documented coverage. ## Sources - [Google DeepMind: AlphaGenome 2 launch announcement, November 11, 2026](https://deepmind.google/blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/). - [AlphaGenome 2 paper](https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphagenome-2-deciphering-how-cells-read-dna-to-better-understand-disease/alphagenome-2.pdf), especially Figure 2c/e and the variant-effect evaluation methods. - [AlphaGenome documentation and FAQs](https://google-deepmind-alphagenome.readthedocs-hosted.com/faqs.html). - Charts visualize the developer-reported results, not an independent evaluation. Google provider logo: [Source: Google](https://blog.google/image-library/), used under the image library’s publication permission. The provider logo is a fallback, not a distinct AlphaGenome 2 logo.