Skip to content

Start with Altar

Altar is a Python framework for analyses that combine genomic variants, executable model predictions, and precomputed biological evidence. It defines the interfaces between those parts and ships reference implementations for running and storing them.

Altar is useful when

  • several model architectures need to produce results through one application;
  • the same model work must run on different compute systems;
  • precomputed scores and annotations need to sit beside newly computed scores;
  • model and evidence integrations should be independently installable;
  • every result must retain model, source, schema, and genome-build identity.

Altar is not

  • a single variant-effect model;
  • a hosted analysis service or user interface;
  • a clinical interpretation system;
  • a knowledge graph;
  • a package that bundles every model framework or set of weights.

Applications decide which integrations to install, how to schedule work, where credentials live, and how results are presented to users. Altar supplies the portable scientific and execution interfaces underneath.

Current scope

  • GRCh38 first. Most integrations and the variant region annotations target GRCh38 (hg38). Cherimoya, Enformer, Borzoi, and LegNet require hg38, and Sei accepts hg19 or hg38. ChromBPNet uses the genome of its configured artifacts. AlphaGenome scores mouse variants on mm10. AlphaMissense and SpliceAI sources are built from an hg19 or hg38 release, the gnomAD source reads gnomAD v4 on hg38, the ClinVar source reads ClinVar's GRCh38 or GRCh37 files, and GPN-Star releases include non-human assemblies.
  • SNVs first. Enformer, Borzoi, LegNet, and Sei score SNVs only. Their manifests declare that, so batch preparation leaves out indels and multi-nucleotide variants and lists them in a report; a mixed cohort still scores its SNVs. ChromBPNet, Cherimoya, and AlphaGenome declare no variant-class limit. GPN-Star covers SNVs only. See variant eligibility.
  • No structural variants. VCF ingestion drops symbolic and structural-variant alleles such as <DEL>.
  • No consequence annotation. Altar does not run VEP or produce HGVS or transcript-consequence annotations. The variant layer assigns coarse region labels, such as coding, intronic, splice, or promoter, by interval overlap.
  • Population and clinical evidence as lookups. The gnomAD binding supplies gnomAD v4 population frequencies, and the ClinVar binding supplies ClinVar's germline classifications. Both report released data; Altar makes no clinical classification of its own. Variant annotation in the variants runtime image also reports the gnomAD frequencies that Open Targets publishes. See population allele frequencies.
  1. Check installation and optional dependencies.
  2. Run the real model-scoring tutorial.
  3. Review the component matrix for the integration you need.
  4. Use the terminology consistently in application code and documentation.