Limitations¶
Altar runs models and looks up evidence. The limits of each model and dataset still apply to the results.
Not for clinical use¶
A variant-effect model predicts a consequence within its training data, architecture, assay, organism and biological context. A prediction does not establish that a variant causes disease or is clinically significant.
Altar has no clinical validation and no regulatory approval, and it makes no claim to support clinical decisions.
Context mismatch matters¶
A model trained on one cell type predicts effects in that cell type. Do not generalize it to unrelated tissues without a reason. A broad hosted model can score many tracks, but taking the maximum across tracks changes the statistic and can saturate. Record the context you configured, and interpret the scores in that context.
Reference datasets have coverage boundaries¶
A missing precomputed value can mean the variant type is not covered, the assembly differs, the chromosome is absent, the alignment has a gap, the licensed data were not supplied, or the source does not cover the position. A missing value is not a neutral score. See Missing is not zero.
Prioritization rules are policies¶
A threshold may come from the literature, from the provider's recommendation or from exploration. Each model page and source page states where its threshold comes from. A default rule is a screening rule, not a calibrated threshold and not a validated classifier. See What prioritized means.
Next¶
- Use Altar from Python runs a spec and reads the results from your own code.
- Cite Altar says how to cite Altar and each model and dataset you use.