Skip to content

specificity-rules.md

Latest commit

 

History

History
47 lines (37 loc) · 2.77 KB

specificity-rules.md

File metadata and controls

47 lines (37 loc) · 2.77 KB
Raw

Designing the specificity criterion

Core principle: the per-cell-type peak is the only screening criterion

Whether a gene is "specific to a certain cell type" is judged solely by whether its peak cell type in the HPA single-cell-type matrix falls inside the target cell-type group. This is the conclusion reached after two rounds of iteration in a real project:

  • The first version also required "highest expression in bone marrow tissue" for megakaryocytic specificity, which wrongly killed a set of genes that are highly specific at the cell level but not at the tissue level (a tissue is a mixture of cell populations, so cell-type specificity is diluted at the tissue level).
  • The correct approach: the cell-type peak decides inclusion; tissue information is additional annotation only. Most genes that satisfy cell-type specificity are indeed also highest in the corresponding tissue (such as bone marrow), but that is a result, not a condition.

Decision logic (implemented in screen_specificity.py)

For each gene:

  1. top_cell = the cell type with the highest expression among all cell types; the peak must be > 0.
  2. driver = the first group in specificity_groups that contains top_cell (the order of the groups in the config is the priority, e.g. put the narrowest lineage group first).
  3. Selected: driver is non-empty and the highest value inside the group ≥ the highest value outside the group (i.e. the peak lies inside the target group).
  4. Borderline: the highest value inside the group ≥ the highest value outside the group × borderline_fraction (default 0.85) but the selection criterion is not met; listed separately for manual review — typical examples are genes such as RUNX1 whose peak sits at the edge of the target group.
  5. specificity_ratio = highest inside the group / highest outside the group, used for ranking and candidate grading; housekeeping genes (such as ACTB, TMSB4X) are expressed at very high levels but have a ratio close to 1, so when commenting on candidates judge mainly by ratio and secondarily by absolute level.

Additional tissue-level metrics

With reference_tissue configured (e.g. "bone marrow"), the output includes that tissue's expression value and its rank among all tissues; used for display and cross-validation only, never for screening.

Cell-type names

Cell-type names must match the column names returned by HPA exactly (case, hyphens, plurals), e.g. Megakaryocyte-Erythroid progenitors, Platelets, cDC, pDCs. After running step 1 you can check the names in the config against the header of the output Excel; the effect of a misspelled name is that the type silently takes part in no group, so it is advisable to check the specificity_groups echo in screening_summary.json.