THE ESSENTIALS IN 30 SECONDS
- AlphaGenome 2 takes a million-letter DNA sequence plus a chosen tissue, cell type or cell state, and predicts regulatory activity for that condition.
- On 28 immune cell types, DeepMind reports it recovers 30% of causal expression variants versus 16% for a fine-tuned AlphaGenome 1.
- Access is through the AlphaGenome API for non-commercial research; DeepMind says the model is not validated for clinical or diagnostic use.
Google DeepMind published AlphaGenome 2 on November 11, 2026, a follow-up to its DNA sequence model that aims to say not only whether a genetic variant changes gene regulation, but in which kind of cell it does so. The release comes with a technical paper and access through the existing AlphaGenome API.
From tissues to cell states
The first AlphaGenome predicted many tissues and cell lines in parallel from a single stretch of DNA. DeepMind's blog argues that this spread the model's capacity thin and made closely related cell types hard to tell apart. AlphaGenome 2 is instead conditional: it receives a one-million-letter sequence together with a single target condition, such as a tissue, a cell type or a cell state, and that condition shapes the computation from the earliest layers onward. Users can choose a reference condition or build one from their own single-cell RNA data, which DeepMind says can work with a few dozen cells and no retraining.
The model was trained on bulk and single-cell data from humans and mice. According to the paper, its single-cell encoder learned from more than 100 million human cell profiles across 68 tissues. Beyond gene expression, it predicts chromatin accessibility, histone marks and, as a new addition, DNA methylation, and it can infer one type of measurement from another.
Reported results
On the TenK10K dataset covering 28 primary immune cell types, DeepMind reports that AlphaGenome 2 identifies 30% of causal expression variants at high precision, against 16% for the original model fine-tuned on single-cell data. It gets the direction of a variant's effect right in 43% of cases, against 23%. The paper also reports better correlation with measured gene expression in GTEx tissues. DeepMind adds that the model reproduced, in silico, a published finding linking an inflammatory bowel disease variant to higher ETS2 activity in gut macrophages.
Access and limits
The model is available free for non-commercial research via the AlphaGenome API and in private preview for commercial users, along with precomputed predictions for trait-linked variants across hundreds of conditions. DeepMind states it is for research only and not approved for clinical or diagnostic purposes. The paper itself notes that the model still misses most fine-mapped variants at high precision, underrepresents transient and disease states, and needs a separate run per condition, which raises compute costs. Interesting Engineering, which covered the launch, points out that predictions still require laboratory confirmation; the benchmark figures are DeepMind's own.
