Altar

Altar is an extensible Python framework for analyses that combine genomic variants, executable model scores, and precomputed biological evidence.

The package provides:

• canonical VCF/TSV variant ingestion and validation;
• model-binding interfaces for container and hosted-API scoring;
• portable execution-task and storage interfaces;
• score, annotation, and variant–gene relation sources;
• attributable per-variant result assembly;
• conformance suites for third-party extensions.

Install the framework and the bindings you need from PyPI:

pip install altar
pip install altar-chrombpnet altar-alphamissense

altar-chrombpnet, altar-cherimoya, altar-alphamissense, and altar-opentargets-e2g are on PyPI. The other bindings install from a source checkout of the repository.

Cloud providers, hosted-model SDKs, and the experimental investigation agent are optional extras, such as pip install "altar[bigquery]". Model frameworks and weights stay in separately locked runtime images. See installation for every binding and extra, and for installing from a source checkout.

Documentation

The full site is published at https://kundajelab.github.io/altar/.

• Start with Altar
• Real scoring tutorial
• Scientific guide
• Integration catalog
• Extension guide
• Public API policy

External packages should import only the compatibility-controlled façade modules: altar.models, altar.execution, altar.results, altar.sources, altar.predicates, altar.registry, altar.access, altar.scoring, altar.testing, and altar.variants. The authentication (altar.access) and job-queue names are provisional: importable, but outside the compatibility commitment until they have a public adapter and consumer.

Plugin discovery and Python imports both use the canonical altar.* namespace. See the plugin packaging guide for the nine entry-point groups, two of them provisional, and validation rules.

License

Altar is copyright 2026 Riya Sinha and released under the MIT License. See the repository's LICENSE and CITATION.cff files for license and citation metadata.