AI Potluck
Back to Gap Map Model components / Scientific AI models

AlphaGenome

Google
restricted / Overall score: 2.0

AlphaGenome is Google DeepMind's genomic sequence model for predicting gene expression, splicing, chromatin features, contact maps, and variant effects. It accepts sequences up to one million base pairs and produces most outputs at single-base-pair resolution through an API, a precomputed Atlas, and a gated downloadable checkpoint.

The repository documents both API and local-checkpoint surfaces; the downloadable model remains subject to separate non-commercial terms. Verified 2026-09-10 via the Hugging Face API, model terms, and Google DeepMind repository.

Openness

2 high confidence
2.0
weights
open(gated downloadable checkpoint files)
data
closed(training corpus not released)
code
partial(Apache-2.0 API client, examples, and notebook
license
AlphaGenome-Model-Terms(non-commercial use only)

The checkpoint is now downloadable after accepting terms, correcting the original API-only roster claim, but the terms prohibit commercial use and therefore cap the score at restricted.

Adoption

2 medium confidence
2.0

The gated Hub endpoint reports zero downloads, so 2,080 GitHub stars are used as the last-resort signal.

Capability

2 high confidence
2.0

One attribute. The multi-instrument credit was withdrawn on re-verification: the attribute asks what the model ACCEPTS, and AlphaGenome accepts DNA sequence - the diversity is in its assay outputs. That is the same distinction that excludes GEDI from granite-geospatial, where lidar is a label source rather than an input. The resolution span stands, and a broad prediction surface is not model adaptation.

Verified 2026-09-10