DeepMind launches AlphaGenome 2 DNA model via API
DeepMind's AlphaGenome 2 predicts DNA variant effects by cell type, finding 30% of causal immune variants in its own test vs 16% for AlphaGenome 1.
What it means for founders
- A startup cannot use the AlphaGenome Services for company work. The AlphaGenome terms open them only to individuals, non-commercial organizations and journalism, and bar any use tied to commercial activity, including research done for a company. Your route is the commercial private preview. The forum post says commercial access is coming "soon" and asks companies to express interest through their Google Cloud account rep, or alphagenome@google.com if they have none. No price is public.
- Outputs cannot train a rival model. The terms also forbid using outputs to train models that predict variant effects "similar to AlphaGenome", and they rule out clinical use. One exception, for files the terms call Permissive Use Downloadable Artifacts, applies across all of those restrictions.
- You depend on Google's API for now. The preprint says source code and weights arrive with final publication. Until then query rates vary with demand, and the repository steers live predictions toward small analyses.
- Precomputed scores are another way in. DeepMind posted a filtered set of its precomputed predictions: enhancer to gene links across 640 human conditions, and variant scores across 510 conditions from about 7 million variants it scored. The preprint marks them for research use only. Confirm whether a download falls under the permissive exception before any commercial use.
How AlphaGenome 2 reads DNA one cell type at a time
The first AlphaGenome took a million letters of DNA and predicted thousands of tissue readouts in one pass. AlphaGenome 2 takes the same 1 Mb sequence plus one target condition, whether a whole tissue, one cell type or a particular cell state, and predicts for that condition alone.
Training covered human and mouse data, bulk and single-cell, including more than 107 million single-cell profiles. AG2 adds DNA methylation to the readouts AG1 offered, and it can fill in a missing measurement, such as chromatin accessibility, from one you supply. Upload single-cell RNA data from a few dozen cells and the model conditions on that state with no retraining.
How the DNA model scored against AlphaGenome 1
In DeepMind's preprint, AG2 picked out 30% of fine-mapped causal variants in TenK10K, a blood study covering 28 immune cell types, while keeping 90% precision. AlphaGenome 1, fine-tuned for single-cell work, reached 16%. For calling whether a variant raises or lowers a gene, at 90% accuracy, the figures were 43% and 23%.
DeepMind also says AG2 computationally reproduced a published result linking a distant regulatory variant to inflammatory bowel disease through overactive ETS2 in gut macrophages. Interesting Engineering points out that predictions of this kind still need lab confirmation.
Where AlphaGenome 2 fits in the AI biology race
Enki has also covered OpenAI buying Orchestra Bio and Anthropic's claim that Claude agents found a CRISPR-like enzyme. DeepMind plans AlphaGenome 2 Skills for Google Antigravity, its agentic development platform. DeepMind and the preprint both say the model is for research only and lacks validation or approval for any clinical or diagnostic use.
What we don't know yet
DeepMind has not given commercial pricing or a date for broad commercial access. The paper is a preprint with no date for final publication, the point at which weights ship. Every benchmark figure is DeepMind's own.
The story
Google DeepMind released AlphaGenome 2 on November 11, a DNA model built to forecast what a genetic variant does to gene activity inside specific human cell types, where its predecessor averaged over whole tissues. Non-commercial researchers can use it free through the AlphaGenome API, while companies wait in a private preview.
Sources
Primary sources
Reporting
More in Research
- Physicist says Astra led his Standard Model proofA physicist says OpenAI's GPT-6 Astra took the lead on the detailed proof behind a candidate lattice definition of the Standard Model.OpenAI · 2d ago

- Podqi finds chatbots sending shoppers to risky sitesA Podqi study found Perplexity, ChatGPT and Gemini showed risky or fake retail sites in 4% to 15% of shopping answers.The Verge · 5d ago

- AlphaProof Code verifies Signal's libsignal in Lean
For founders: Specifications are now the expensive part. A team of five finished in four weeks without writing proof code themselves.
Google DeepMind · 8d ago - OpenAI claims AI proofs on Kakeya, ETH and QCD

For founders: Watch the ETH review closely. If it holds, problems long assumed hard under ETH may admit faster exact algorithms, though OpenAI itself says practical gains…
OpenAI · 23d ago - GPT-6.1 Sol and Grok 4.7 pick rival parties in study

For founders: API access may not carry the app's guardrails. If your product calls GPT-6.1 Sol or Grok 4.7 directly, a user may get a political recommendation that ChatGPT…
TechCrunch · 25d ago