Overview
| Name | AlphaGenome · aliases: AlphaGenome |
|---|---|
| Full name | The developer uses the model name only |
| Ambiguous names (not auto-linked) | AlphaGenome 2 |
| Family | Not a family entry. The successor AlphaGenome 2 (2026-11-11) is recorded as the latest event here; no family entry |
| Organization | Google DeepMind |
| Organization type | Big-tech lab |
| Country | United Kingdom |
| Model class | Genome sequence-to-function model / variant effect prediction |
| Domain | Genomics & RNA |
| First release | † partly verified Basis: Developer blog announcement with API preview (preprint 2025-06-27) Note: The announcement date was not directly re-verified and comes from aggregators source |
| Other milestones |
|
| Latest update | · successor Basis: Successor model AlphaGenome 2 launch and API v2.0.0 AlphaGenome 2 launched: a conditional model predicting one tissue or cell type at a time, reported by the developers as trained on over a hundred million single-cell profiles; available by API for non-commercial use with a commercial private preview, weights not mentioned Note: The published record did not include this event source The most recent sourced event, not a claim that nothing changed since. |
| Status | Superseded by a successor · The successor AlphaGenome 2 has been released |
| Openness | Code: Open · documentation and example code CC-BY-4.0 Weights: Non-commercial · non-commercial terms, acceptance required Weights and inference code public from 2026-01-28; API open for non-commercial use; the successor's weights are not mentioned |
| Benchmarks reported | 3 counted (self 1, modified protocol 2) |
| Wet-lab validation | None recorded |
| Unverified fields | † first_release, release_milestones.official_announcement, training_data.a_public |
| As of | 2026-11-16 |
What it does
Reads 1 Mb of DNA and predicts expression, transcription initiation, chromatin accessibility, histone marks, transcription-factor binding, contact maps and splice-site usage at base resolution, scoring non-coding variants by the difference between reference and alternate predictions. Compared with Borzoi it lengthens the input and merges modalities into one model; unlike self-supervised DNA language models such as Evo 2 and NTv3 it is supervised on functional-genomics tracks.
Benchmarks reported
Every row is one pairing from the survey's relation file; a score belongs to the benchmark page's ranking, where it is shown with metric, eval set, source type and date. 3 of 3 pairings are counted.
| Benchmark | Relation | Eval set / protocol | Source type | Note and evidence |
|---|---|---|---|---|
| 08 Basenji/Enformer test set | Developer report, official protocol | — | Self-reported | Self-reported on this data's track test set · evidence |
| 11 ClinVar hold-out protocol | Developer report, modified protocol | — | Self-reported | Its own variant-effect suite, including splicing classes from this database †partly verified · evidence |
| 37 TraitGym | Developer report, modified protocol | — | Self-reported | Uses the variant set and matched controls but reports enrichment rather than AUPRC, so it is not board-comparable. †partly verified · evidence |
Other evaluations (outside the 44)
- QTL and causality tasks (Own task set) Self-reported: Direction and coefficient correlation for expression and accessibility QTLs, plus a supervised classifier trained on multimodal scores, all developer-defined
- Public reporter-assay datasets (Public benchmark not among the 44) Self-reported: Splicing and regulatory evaluations on two public MPRA datasets, outside the 44
- Third-party probing comparison (2026-04) (Comparison against baselines only) Third-party: A third-party study claims probes trained on another genome model's embeddings beat this model; those are probes rather than zero-shot runs and are not directly comparable
- BEND (Public benchmark not among the 44) Self-reported: The original paper does not report this benchmark
Developers report: one model matches or beats the strongest external model per task on most track and variant-effect evaluations, with multimodal prediction in a single 1 Mb context.
Independent evidence: No independent evidence found. No third-party evaluation under a site benchmark protocol was found even after the weights were released; existing third-party work uses a probing setup
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | Public functional-genomics and variant resources; the paper states no new raw experimental data was generated source |
| b. Private or licensed data | Not disclosed | |
| c. Own wet-lab data used for training | No | source |
| d. Wet lab used only for validation | No | The reporter-assay and variant validations in the paper use public data only source |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: A London AI lab inside a US-listed technology group, founded in 2010, acquired in 2014 and merged with another group research unit in 2023-04. Headcount framings differ: a 2025-09 media report cites about 5,600 (a wider remit) while a data firm counts 3,678 for its UK entity (2025).
- Funding: Funded from the group's R&D budget; a sister AI drug-design company in the same group announced a $600M first external round on 2025-03-31.
- Products and partnerships: A sibling co-folding model AlphaFold 3; two earlier sequence-to-function models, one of which (Borzoi) is both predecessor and comparator; at API launch the developers said thousands of scientists were testing it; the group also has the single-cell model C2S-Scale with a US university.
- Related models: Borzoi, AlphaFold 3, C2S-Scale, Evo 2, NTv3
Reading notes
- All win counts are self-reported and measured against the strongest external model per task, not a rank on a common third-party board.
- The 2025-06 preview and the journal version count differently; always state which version is cited.
- The ClinVar side covers splice-related classes only and does not show general pathogenicity classification is usable.
- Long-range regulation (enhancer-gene links beyond 100 kb) is an acknowledged weakness, and only TraitGym's variant sets were used.
Caveats
- The preview and journal versions count differently and must not be mixed.
- Weights carry non-commercial terms, and the model card and paper conflict on whether one tissue dataset is in the released version.
- All site benchmark entries are self-reported; no third-party figures are citable.
- The successor is API-only, with weights and evaluation detail undisclosed.
Sources
- Journal paper (journal)
- Journal full text (public archive) (journal)
- Developer blog (preview) (company blog)
- Developer blog (successor model) (company blog)
- Code and weights repository (repository)
Corrections
2 corrections
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.