Skip to content

Model and source catalog

These are the models and annotation sources Altar can add to a run. A model computes a score for each variant in your run. An annotation source looks each variant up in a released dataset; it runs no model and needs no GPU. Most come in their own binding, a package you install separately, so install only the ones you use. The genomic regions source is part of altar itself.

Each row links to the model page or source page, which has its run spec section, columns, inputs, citation and licence. The validation column says how each one was checked against its upstream implementation or data. A model's label is explained on its page and in the validation status table. An annotation source has one of two labels:

  • Matches upstream on spot checks: when the published copy was built, its row counts were compared with the upstream release and named variants were looked up in both and agreed. The source page says what was checked.
  • Not spot-checked: no recorded comparison with upstream.

Models

Model What it adds Binding Install from Genome builds Variant classes Compute Validation Licence and terms
AlphaGenome Headline scores reduced from many tracks, with every track as details altar-alphagenome PyPI hg38, mm10, set by organism Every class Hosted API; no GPU Not validated Apache 2.0 SDK; the API's terms of use (non-commercial)
Borzoi Top gene and GTEx logSED, with every gene effect as details altar-borzoi PyPI hg38 SNVs NVIDIA GPU per batch Upstream parity (GPU and CPU) Apache 2.0 code; no stated weights licence
Cherimoya Counts effect and profile distance from a BPNet-style model altar-cherimoya PyPI hg38 Every class NVIDIA GPU per fold, then a CPU summary Smoke-tested MIT code; CATv1 weights CC BY 4.0
ChromBPNet Regulatory-effect scores on chromatin accessibility, over five folds altar-chrombpnet PyPI Any; set by the model's files Every class NVIDIA GPU per fold, then a CPU summary Reference comparison MIT code; ENCODE weights, no use restrictions
Enformer Top-track SAD and SADR, with selected-track details altar-enformer PyPI hg38 SNVs NVIDIA GPU per batch Upstream parity (CPU) Apache 2.0 code and weights
LegNet Signed MPRA activity effect altar-legnet PyPI hg38 SNVs CPU only Upstream parity (CPU) MIT code and weights
Sei Top sequence class, with all 40 class effects as details altar-sei PyPI hg19, hg38 SNVs NVIDIA GPU per batch Upstream parity (CPU) Academic and research use only; Zenodo archive lists CC BY 4.0

The container models run in a linux/amd64 image that Altar pulls on first use. AlphaGenome runs in the Altar process and calls a hosted API with your key.

Annotation sources

Source What it adds Binding Install from Genome builds Variant classes Data Validation Licence and terms
AlphaMissense Missense pathogenicity score and class, per transcript altar-alphamissense PyPI hg19, hg38 Missense SNVs, as DeepMind scored them Bring your own, built from DeepMind's release Not spot-checked Predictions CC BY 4.0 (Zenodo version 3)
ClinVar Germline classification, review stars, conditions and genes altar-clinvar PyPI hg38 (published), hg19 SNVs and indels with a ClinVar record, on chr1–22, X, Y and M Published (20260928, hg38), or bring your own, such as a newer weekly release or hg19 Matches upstream on spot checks No NCBI restrictions; attribute ClinVar
ENCODE-rE2G and scE2G atlases Genes a variant's regulatory element is predicted to regulate, from ENCODE-rE2G and scE2G files altar-e2g-atlas Source checkout hg38 Every class, by overlap Bring your own, imported from ENCODE-rE2G or scE2G files Not spot-checked No ENCODE restrictions; cite ENCODE-rE2G, scE2G and ENCODE
Genomic regions Region class, nearest genes and cCRE overlap altar (core) PyPI hg38 Every class Bring your own, region files built from Ensembl, SCREEN and GENCODE Not spot-checked MIT code; open Ensembl, SCREEN and GENCODE data
gnomAD Population allele frequencies, counts and site filters altar-popfreq PyPI hg38 SNVs and indels, on chr1–22, X and Y Published (4.1.1-genomes, hg38), or bring your own Matches upstream on spot checks CC0 data; cite gnomAD
GPN-Star Evolutionary-constraint log-likelihood ratios altar-gpnstar Source checkout ce11, dm6, gg6, hg38, mm39, tair10, one per copy, by score set SNVs between A, C, G and T Bring your own, built from a GPN-Star release Not spot-checked Apache 2.0 scores; cite GPN-Star
Open Targets E2G Genes a variant's regulatory element is predicted to regulate altar-opentargets-e2g PyPI hg38 Every class, by overlap, on chr1–22 and X Published (26.09, hg38), or bring your own Matches upstream on spot checks CC0 1.0 data; cite Open Targets and ENCODE
phyloP 100-way Base-wise conservation score altar-conservation PyPI hg38 SNVs; every class with indels = "max". Bases on chr1–22, X, Y and M Published (2015-05, hg38), or bring your own Matches upstream on spot checks Free for public use; cite UCSC
SpliceAI Precomputed splice delta scores, per gene altar-spliceai Source checkout hg19, hg38 SNVs and indels Illumina scored: substitutions, 1-base insertions and 1–4-base deletions within genes Bring your own, built from Illumina's licensed release Not spot-checked Non-commercial use only; Illumina's terms

"Published" means the Kundaje lab hosts a built copy that the source reads in place, with no download and no credentials. "Bring your own" means you build or import the data first, with the command its page gives.

A source marked "Source checkout" installs from a clone of the repository; see Install from a source checkout. altar sources lists the sources installed in your environment, and altar models the models.

Next