Altar AlphaGenome binding

altar-alphagenome connects Google DeepMind's hosted AlphaGenome API to Altar. It registers ALPHAGENOME in Altar's canonical altar.model_plugins group and scores inline: build_scoring_plan returns a JSON-durable InlineScoringPlan, and execute_inline calls the API in process without an execution backend or container.

The binding is not on PyPI. Install it from a clone of the repository, which also installs Altar from the same checkout:

uv pip install -e bindings/alphagenome

The binding depends on the exact alphagenome SDK release recorded in its runtime identity and refuses to execute with a different version. The SDK is imported only when the default scorer runs, so plans can be built, validated, and materialized without loading the client.

An Altar model instance is AlphaGenome plus its configured organism, ontology terms, variant scorers, and sequence length. The binding always calls the production ALL_FOLDS model rather than exposing training folds in user configuration. Human coordinates must be hg38 and mouse coordinates mm10. The API key is a SecretReference in durable configuration; the raw key, an injected scorer, or an SDK client reaches the plugin only through the non-durable AlphaGenomeRuntime.

Each variant yields one scalar headline row (top_output_type, top_track, max_abs_raw_score, top_track_abs_quantile_score, max_abs_quantile_score, n_tracks_scored) and a lossless track_scores detail table with one row per scorer and genomic track. Persist the detail table with any Altar DetailStore.

See the AlphaGenome integration page for the reduction, prioritization rule, and compatibility notes. API quota, availability, privacy policy, and provider terms of use apply independently of this MIT-licensed binding.