Skip to content

From variants to evidence

An Altar run attaches three kinds of information to each variant: scores a model computed in this analysis, annotations looked up in a dataset that already exists, and links from the variant to genes. Each kind keeps its own meaning, units and source.

Altar does not combine these into one biological score. If you rank, filter or train on them, that is your own policy. Write it down and version it.

Model scores

A model score is a value a model computed for one variant in this analysis. It belongs to one model instance: a trained model with a fixed biological scope, such as a cell type, tissue, set of folds, checkpoint or selection of tracks. The model's page defines each score: its meaning, units and direction.

Two models can report a quantity with the same name, such as a log2 fold change. Each value stays attributed to the model that produced it, so the columns do not collide. Related models keep different names when their quantities differ. ChromBPNet's logfc and Cherimoya's counts_log2fc, for example, are defined separately.

Altar records which computation produced each stored score. Score reuse and provenance describes what that record holds and when a later run reuses the score.

Annotations

An annotation is a value looked up in a dataset that existed before the analysis: an observed fact such as a population allele frequency, a published prediction such as a precomputed SpliceAI or AlphaMissense score, or a summary of several source records. A lookup runs no model.

A dataset of model predictions is still an annotation. Precomputed SpliceAI scores are looked up, so they carry the dataset's release, not a model run's identity. An annotation source can have its own prioritization rule. Models, sources and bindings describes annotation sources.

A variant–gene link connects a variant, or a regulatory element it falls in, to a gene the element is predicted to regulate. One variant can have many links. A link can keep:

  • the variant and the linked regulatory element;
  • the target gene's identifiers;
  • the biosample or tissue;
  • the link's score and its direction;
  • distances along the genome;
  • the method, dataset release, genome build and source record.

A link is a prediction or an observation from its source. Altar reports it and does not assert that the variant acts through that gene.

A link source adds its links to the export as columns: the strongest link, the number of links and genes, and a list of every link. Each link source's page lists them.

Variant results

A variant result is the assembled record for one variant. It holds:

  • each model's scores, with the identity of the computation that produced them;
  • each annotation and link column;
  • prioritized, and the models and sources whose rules matched.

What prioritized means explains the flag. A model's output with many values per variant, such as per-track effects or a matrix, stays out of the variant result, in a detail table or an artifact file; see Where results live.

Missing is not zero

An empty value means there is no value. It can mean the dataset does not cover the position, the variant type is not supported, the model could not score the variant, or a needed file was absent. A zero is an observed value. Altar keeps missing values empty and never fills them with zero.

Next