What “prioritized” means¶
prioritized is true for a variant when at least one model's or annotation source's declared rule matched it.
It does not mean:
- pathogenic or benign;
- causal for a phenotype;
- clinically actionable;
- statistically significant;
- higher quality than every unprioritized variant;
- calibrated to the same meaning across model architectures.
How the flag is assembled¶
Each model may declare a rule over its scores and the annotations joined to the variant. Each annotation source may declare a rule over its own columns. Altar evaluates each rule separately and records:
prioritized_model_ids, the models whose rule matched;prioritized_source_names, the sources whose rule matched;prioritized, true when either list is not empty.
A rule counts only when it returns true. Each list follows the order of the models and sources in the run spec. A variant that no rule kept has two empty lists, not missing values.
When you show a prioritized variant, show why it was selected: the two lists, not only the flag.
Each model page and source page states its default rule and where its threshold comes from. A default rule is a screening rule, not a calibrated threshold.
Where the reasons appear¶
Every variant result carries both lists, and the export writes them as its last two columns. A TSV export writes each list as JSON text, and a Parquet export keeps it as a list of strings.
Here is a prioritized-only TSV export with the model's score columns left out. The run has one model,
[models.k562], and the ClinVar source [sources.clinvar] with prioritize_pathogenic = true.
variant_id prioritized clinvar_classification … prioritized_model_ids prioritized_source_names
chr17:43045712:T:C true Pathogenic … [] ["clinvar"]
chr1:1000000:A:G true … ["k562"] []
chr7:117559590:ATCT:A true Pathogenic … ["k562"] ["clinvar"]
ClinVar alone kept the first row, the model alone kept the second, and both kept the third. To keep the rows one
model selected, filter on prioritized_model_ids rather than on prioritized.
Required annotations¶
A model's rule can read annotation columns, such as ChromBPNet's region_type. The model declares each column
it needs. A run stops when no configured source supplies a needed column. It warns when the column comes from a
source that does not declare it follows the same definition, because nothing then states that the values mean the
same thing. See Annotation dependencies.
Missing values¶
A rule has three outcomes: true, false or unknown. If a value the rule reads is missing, the outcome is unknown, not false. An unknown rule does not prioritize the variant. It still differs from a rule that saw the value and returned false. A variant that a configured source has no record of gives such a missing value.
Comparing models¶
A shared flag does not make the underlying scores commensurate. Compare raw values only when their model pages
define the same quantity, units, direction, and context. A ChromBPNet logfc, an AlphaGenome maximum track
effect, and a GPN-Star calibrated likelihood ratio answer different questions.
A weighting across models, or a learned ranking, is a policy of your own. Version it separately from Altar's results. Write a prioritization rule shows how to write your own rules.
Next¶
- Where results live describes the stores and the export a run writes.
- Write a prioritization rule writes your own rules.
- Tutorial 4: write your own prioritization rule writes a rule of your own in a notebook and saves it as a policy.