Skip to content

Glossary

The terms these docs use, in plain words.

Term Meaning
annotation A value looked up for a variant in an existing dataset, such as an allele frequency or a precomputed score.
annotation source A dataset a run looks annotations up in, and the installed code that reads it, such as gnomAD or ClinVar. It runs no model. A run spec names one in a [sources.<name>] section, and altar sources lists the installed ones.
artifact A file a model writes for a variant set, such as a matrix or a track, stored as a file rather than as table rows.
artifact store Where a run keeps artifact files, named in [artifact_store].
backend Short for execution backend.
binding A package that connects one model, dataset or service to Altar, such as altar-legnet or altar-popfreq.
connection A named way of reaching a bucket or a warehouse, declared in a [connections.<name>] section. It uses the credentials of the environment, so a spec never holds one. Provisional.
contract What an annotation source declares about its columns: a source ID and a schema version. Sources with the same contract give the same columns with the same meaning. It is part of a source's identity.
detail store Where a run keeps detail tables, named in [detail_store]. Without that section, the details go beside the scores.
detail table A table of a model's rich output, with many rows per variant, such as one row per track.
execution backend What runs a model's container tasks: local Docker, Kubernetes or Modal. A run spec names it in [backend].
export The table a run writes: one row per variant, with its scores and annotations.
genome build The reference assembly a variant's coordinates are on, such as hg38.
identity What a stored value records about how it was produced, as a hash. A score's identity is the ID of its lineage. A source's identity is its contract, release and genome build. Altar reuses a stored value only when its identity matches the one the run asks for.
inline model A model Altar runs in its own Python process instead of in a container, such as AlphaGenome, which calls a hosted API. It needs no execution backend.
lineage The scores stored from one computation: one model with one set of weights and score-changing settings, one plugin version and one runtime image. See Score reuse and provenance.
A binding that fills a variant–gene link store from a published atlas, such as the ENCODE-rE2G and scE2G atlases. A run reads the stored links through the variant_gene_links annotation source.
model architecture A modeling approach, such as ChromBPNet, Cherimoya or AlphaGenome.
model instance One trained model in a fixed biological scope, such as a K562 ChromBPNet model. A run spec declares it in a [models.<name>] section.
model key The <name> of a run spec's [models.<name>] section. The store keeps the model's scores under it, and the run summary and the export's provenance name the model by it. The spec key model_name is only a display name for the model's export columns. Provenance's model_name field records it, and holds the model key when the spec sets none.
normalization Writing a variant in one standard way: trimming shared bases, then left-aligning indels against the reference.
plan What a model binding builds for one scoring request. A container plan lists the container tasks to run and which of their output files hold the scores, detail tables and artifacts; any execution backend can run it. An inline model's plan is the request it sends from Altar's own process.
plugin The class a binding registers with Altar through an entry point, such as a model plugin or an annotation source class.
preparation Work some models do once before scoring, deriving files from their weights and reference data, such as ChromBPNet's per-fold calibration. altar prepare runs it and publishes the files altar run reads.
prioritized A declared rule matched the variant. It does not mean pathogenic, causal or clinically significant. See What prioritized means.
provenance The record, in every export, of what produced its values: for each model, the lineage the run asked for and the lineages its rows came from; for each source, its release and genome build. See Score reuse and provenance.
published copy The copy of a source's dataset that its publisher hosts. A source reads it by default, with no download. altar pull copies it into storage you own.
release The version of a source's dataset, such as gnomAD 4.1.1-genomes or ClinVar 20260928. It is part of the source's identity, and the export records the release each source served.
resource manifest A JSON file a binding's resources command writes, naming a model's weights and other files with their digests. A [models.<name>] section points to it with resource_manifest.
run spec A TOML or JSON file that describes one run: its variants, models, sources, store and export. See the run spec reference.
runtime The container image that holds a model's framework and its command-line program.
score A value a model computed for one variant.
score store Where scores are kept between runs, such as a SQLite file or a BigQuery table, named in [store].
shard One part of work or output that is split into several: one of a plan's parallel container tasks, such as one fold, or one file of a split output. An interpretation product names its shard with labels, such as allele=ref for the reference allele's file.
store Where a run keeps results between runs: the score store, the detail store or the artifact store.
task One unit of work that runs on its own in one container, such as one fold of a ChromBPNet model, the task that combines its folds, or one batch of LegNet.
type The type key of a run-spec section: the installed model, source, backend, store or adapter the section uses, such as "LEGNET" or "gnomad". altar models and altar sources list the model and source types.
variant ID A variant written as chr:pos:ref:alt, such as chr1:100:A:G.
variant result The assembled record of one variant: its scores, its annotations and the reasons it was prioritized.
A relation from a variant, or a regulatory element it falls in, to a gene.