Explain the 98% of the genome that does not code for proteins, so researchers can find which DNA changes behind rare and common diseases actually matter.
Open weights and Nature paper (Jan 2026); Atlas of 9B DNA variant predictions released Sep 2026.
Updated 8 Sep 2026Checked 11 Oct0 updates this week
Convolutional layers scan the sequence for short letter patterns, such as the spots where control proteins bind.
Transformer layers let every part of the million-letter window inform every other, since switches can sit far from the gene they control.
Output heads predict gene activity, splicing, open DNA and protein binding across hundreds of cell types, trained on public lab data such as ENCODE and GTEx.
Run the normal and mutated sequence and compare the predictions. The Atlas precomputes this for every possible single-letter change.
| Spec | AlphaGenome |
|---|---|
| DNA read per prediction | Up to 1 million base pairsR (reported) |
| Output resolution | Single DNA letterR (reported) |
| Variant-effect benchmarks led or matched | 24 of 26R (reported) |
| Training time | 4 hours on TPUsR (reported) |
| Scientists using the API (Jan 2026) | 3,000+R (reported) |
| API requests per day (Jan 2026) | About 1 millionR (reported) |
| Variants scored in the Atlas | 9 billionR (reported) |
| Atlas dataset size | About 1 petabyteR (reported) |
R reported by the company
Google DeepMind
Gemini, AlphaFold, Genie, Veo, TPU chips
Google DeepMind is Alphabet's AI lab, formed in 2023 from DeepMind and Google Brain. It builds the Gemini models and leads in AI for science. AlphaFold won a share of the 2024 Nobel Prize in Chemistry.
