Skip to content

Pathograph overlap as a qualitative mirror of shared heritability

Prototype motivated by Zhao et al., Nature Genetics (2026), "Pleiotropic shared heritability quantifies the shared genetic variance of common diseases" (doi:10.1038/s41588-026-02607-w).

Idea

That paper's method (PHBC) quantifies how much genetic variance a target disease shares with a set of auxiliary diseases (a scalar, h2_pleio/h2), but has no representation of what is shared. dismech is the complementary layer: every disorder is a pathograph (pathophysiology nodes annotated with cell types, biological processes, genes, phenotypes, and causal edges). The primitive operation is therefore a pairwise overlap(A, B) — intersect two pathographs into a pair-specific subgraph of shared mechanism. Pre-curated mechanism modules (kb/modules/) are just the overlaps recurrent enough to be worth naming; the overlap function needs no module to run on an arbitrary pair.

Method (scripts/pathograph_overlap.py)

  • Extract mechanistic terms (CL, GO, gene symbols, HP) from every disorder pathograph (1,214 disorders, ~7,300 unique terms).
  • Specificity inference: IDF-weight terms over the corpus so ubiquitous mechanisms (inflammation, apoptosis, seizure) self-decay and specific shared mechanisms dominate.
  • Near-parent inference: optionally expand each term to its 2-hop is_a ancestors (OAK), then recompute IDF on the expanded corpus, so over-generic ancestors self-cancel while a shared moderately specific ancestor connects diseases that used different child terms.
  • Score = symmetric weighted Jaccard (>= 2 shared terms required).

Run: 

python3 scripts/pathograph_overlap.py            # corpus stats + top pairs
python3 scripts/pathograph_overlap.py validate   # vs genetic rg (exact match)
python3 scripts/pathograph_overlap.py validate-expand  # + near-parent inference

Result

Against 22 disease pairs with published LDSC genetic correlations (illustrative literature consensus; see RG_PAIRS in the script), spanning strong to near-zero:

metric exact match + near-parent
Spearman rho(|rg|, overlap) 0.40 0.50
mean overlap, strong pairs (|rg| >= 0.30) 0.109 0.134
mean overlap, near-zero controls (|rg| < 0.15) 0.020 0.044
Schizophrenia <-> Bipolar (rg 0.68) 0.000 0.046 (recovered)

Overlap tracks genetic correlation; near-parent inference improves the rank correlation and recovers pairs exact matching missed. Near-zero-rg negative controls correctly sit near-zero overlap.

Most actionable output: high-rg / zero-overlap pairs = curation gaps

Type 2 Diabetes <-> Coronary Artery Disease: rg ~ 0.35, overlap = 0.000 even after ancestor expansion. The two pathographs share no annotated mechanism because CAD is curated around atherosclerotic-plaque/foam-cell biology and T2D around insulin/beta-cell biology; the connecting insulin-resistance -> endothelial-inflammation layer is not yet curated. Established genetics asserts shared mechanism the KB has not captured — a ranked, evidence-backed curation to-do list.

Caveats / next steps

  • The rg table is hand-curated literature consensus, not a harmonized pull; swap for a real genetic-correlation source (e.g., LDSC atlas / LD Hub) for a systematic test.
  • Gene matching is by symbol from the genetic section; could be tightened to HGNC ids.
  • Ancestor expansion is is_a only and 2 hops; a Resnik-style information-content measure over most-specific common ancestors is the principled successor.
  • Generate the full ranked high-rg / low-overlap gap list across all pairs to drive curation priorities.