Skip to content

Altar

Pre-release documentation

These pages describe Altar 0.4.0, which is not yet on PyPI. pip install altar installs 0.3.0, which lacks some commands shown here. To follow these pages now, install the development branch.

Altar scores genomic variants with sequence models and annotates them with published evidence, in one run that writes one table. You describe the run in a TOML file: your variants, the models, the annotation sources and where to keep the scores. Altar runs each model in a container on your machine or a cluster, or calls its hosted API. It looks up the evidence, and records which model, settings and dataset release produced every value.

One altar run: a run spec and variants go in; Altar normalizes the variants, scores each model only on variants not already in the score store, looks each variant up in the annotation sources, and writes one table with provenance. One altar run: a run spec and variants go in; Altar normalizes the variants, scores each model only on variants not already in the score store, looks each variant up in the annotation sources, and writes one table with provenance.

Beta

Altar is in beta. Each interface and feature is labelled Stable, Provisional or Preview. See Maturity labels.

What Altar is

  • A command, altar run, and a Python library for scoring and annotating a variant cohort.
  • Bindings, separate packages that connect Altar to seven models (ChromBPNet, Cherimoya, Enformer, Borzoi, LegNet, Sei and AlphaGenome) and to evidence such as gnomAD, ClinVar, phyloP, AlphaMissense, SpliceAI and enhancer-to-gene links.
  • A score store that keeps every model's scores, so a later run scores only the variants it has not seen.
  • A set of interfaces for adding your own model, dataset, compute backend or store.

What Altar is not

  • A variant-effect model of its own. It runs other people's models.
  • A clinical tool. A prioritized variant is one that a declared rule matched; it is not a pathogenicity call. See Limitations.
  • A consequence annotator such as VEP. It does not produce HGVS or transcript consequences.
  • A hosted service. You install it and run it on your own machine or cluster.

Start here

To understand what the results mean before you use them, read From variants to evidence.