A DNA model that asks "which cell?"
On 11 November 2026, Google DeepMind released AlphaGenome 2, the follow-up to last year's AlphaGenome DNA sequence model. The authors are Adam Gayoso, Natasha Latysheva, Pushmeet Kohli and Žiga Avsec. The model's job hasn't changed. You give it a stretch of DNA and it predicts how that sequence, and any small change to it, affects the molecular machinery that reads genes. What's changed is the question it answers. The first model gave a broad answer across many tissues at once. AlphaGenome 2 answers for one specific cell type or cell state.
That matters because a disease variant can be harmless in most of the body and only cause trouble in one population of cells. DeepMind's showcase example is a variant linked to inflammatory bowel disease. The new model reproduced a published finding that it works by overactivating the ETS2 gene in gut macrophages.
How the model works
AlphaGenome 2 reads up to 1 Mb of human or mouse DNA, the same window as its predecessor. Each request now also carries a "condition", which can be a tissue, a cell type or a cell state. The original model filled a fixed catalogue of thousands of output tracks in parallel, and DeepMind says that spread its capacity thin. The new design feeds the condition into every stage of the network instead.
According to MarkTechPost's write-up of the technical paper, the architecture is U-Net style: a convolutional encoder, an 11-block transformer and a decoder. The condition becomes a 256-dimensional embedding that adjusts activations through FiLM layers throughout the network. The paper found that conditioning only the output end barely helped, and that most of the gain came from conditioning the encoder and transformer. DeepMind hasn't disclosed the parameter count.
Researchers can pick a condition in two ways:
- Reference conditions: a named entry from DeepMind's catalogue, identified by ontology codes.
- Expression conditioning: you upload your own single-cell RNA-seq data as raw UMI counts with Ensembl gene IDs, and a cell encoder turns it into a condition embedding. The API docs cap this at 256 cells per call. DeepMind says a few dozen cells can be enough to describe a cell state the model has never seen, with no retraining.
The model predicts gene expression, chromatin accessibility, histone marks, transcription factor binding, splicing and 3D contact maps. DNA methylation is new and was a common request. It can also impute one measurement from another, for example estimating chromatin accessibility from single-cell gene activity. DeepMind trained it on bulk and single-cell data from human and mouse samples. The blog puts the single-cell total at more than 100 million cell profiles across 68 tissues.
What DeepMind says it achieves
The headline result uses TenK10K, a single-cell dataset covering 28 primary immune cell types. At 90% precision, AlphaGenome 2 found 30% of causal variants that change gene expression. The original AlphaGenome, fine-tuned on single-cell data, found 16%. For predicting whether a variant pushes gene activity up or down, recall rose from 23% to 43%.
The paper, as summarised by MarkTechPost, reports more gains:
- On GTEx tissues, median correlation for cross-tissue expression changes rose from 0.61 to 0.71.
- On held-out ENCODE-rE2G CRISPRi data, enhancer-gene linking reached an AUPRC of 0.78. AlphaGenome 1 scored 0.73 and a Borzoi ensemble scored 0.61.
- For imputing missing assays, it beat the Avocado baseline on 166 of 181 held-out tracks.
- On an unseen Parkinson's cohort, it reached 41% recall against 17% for a cohort-tuned Scooby model.
There's a weak spot too. Outside accessible chromatin, its AUROC for predicting effect direction is only 0.63. DeepMind ran every one of these evaluations itself, and we haven't found any independent replication yet. Treat the numbers as the developer's own claims until other labs test them.
Access, cost and terms
AlphaGenome 2 runs as a hosted service through the AlphaGenome API, and DeepMind has published Colab tutorials for expression conditioning. Here's what the official pages say:
- Price: API keys are free for non-commercial use, according to the GitHub README. DeepMind hasn't published pricing for commercial use. The blog says commercial research access is in private preview, and the README points commercial users to Google Cloud.
- Weights: you can't download them yet. MarkTechPost reports they'll come "upon final publication" of the paper.
- Quotas: the README gives no fixed numbers and says query rates vary with demand. Looking up precomputed results is faster than running live predictions.
- Usage limits: outputs are for non-commercial use, and you can't use them to train other machine learning models. Predictions are for research and modelling only. DeepMind says the model hasn't been validated for clinical or diagnostic use.
- Software licence: the Python client (
pip install alphagenome) is Apache-2.0. The docs and examples are CC-BY 4.0.
What changes for people already using AlphaGenome 1
The comparison guide in DeepMind's docs lists a few breaking changes. Every AlphaGenome 2 call needs exactly one condition, so if you want to compare cell types you'll loop over conditions or use the batch methods. Variant scoring now returns only raw scores. Version 1 also returned quantile scores ranked against about 300,000 common variants, so any pipeline with thresholds built on those percentiles will need recalibrating. PRO-cap output has been removed, and the docs suggest CAGE as a stand-in. Splice site outputs are now specific to each condition. Contact maps use different targets, so their values can't be compared directly with version 1. Each track also comes labelled as observed, imputed with support, or imputed and exploratory, which tells you how much to trust it. DeepMind still recommends version 1 for reproducing the original paper, scoring every tissue at once, or getting PRO-cap tracks.
What builders and researchers can do with it
There are a few practical uses here:
- Fine-mapping GWAS hits: DeepMind has precomputed scores for trait-linked variants across hundreds of cell and tissue conditions. MarkTechPost reports one atlas covering about 7 million common and GWAS variants across 510 conditions, and another linking enhancers to genes across 640 primary cell conditions. DeepMind gives height and type 2 diabetes as examples where the model helped find causal variants and the cells they act in.
- Your own single-cell data: labs with scRNA-seq from a tissue or state that isn't in the catalogue can condition directly on their own cells. In a COVID-19 mRNA vaccine study, the model told activated T cells apart from resting ones. It also found that a variant tied to severe COVID-19 disrupts gene activity more in the activated state.
- Drug discovery screening: Genomics Limited chief executive Peter Donnelly says the company plans to build the model into its Mystra platform. We'd expect other commercial teams to wait for the preview to open up and for pricing to appear.
The view from Aotearoa
University and Crown research institute teams doing non-commercial work can start using the API now. Startups and companies should treat it as a research tool until commercial terms are public. Anyone uploading single-cell data should check the API terms against their own consent and ethics approvals first. For any project involving Māori genomic data, the HRC-funded Te Mata Ira guidelines for genomic research with Māori, led by Associate Professor Māui Hudson at the University of Waikato, are the place to start. They cover collaboration and the cultural and ethical expectations involved, so read them before sending samples' data to an overseas service. DeepMind is taking use-case proposals until 30 November, and chosen projects may get extra quota and credits.
Sources
- AlphaGenome 2: deciphering how cells read DNA to better understand disease (Google DeepMind blog)
- AlphaGenome 1 vs AlphaGenome 2 (AlphaGenome documentation)
- google-deepmind/alphagenome (GitHub)
- Driving indigenous data sovereignty (Health Research Council of New Zealand)
- Google DeepMind Releases AlphaGenome 2 for Cell-Type-Specific Variant Effects (MarkTechPost)
Build with people working on the same things
Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.