Combo Contribution Board

Definition: LOO(M) = recall(pool) − recall(pool minus M). Positive = removing M drops recall = M carries irreplaceable signal in this pool; negative = removal improves recall. Overall tab sorts by the 5-disease mean; disease tabs sort by that disease.

⚠ Differences between combos are not corrected for multiple comparisons: in the paper's E8_2, none of the 50 combos survived BH correction; rankings are exploratory, not significance claims.

⚠ Structural gene coupling between the NUT family and DrugBank labels (nutrient targets come from DrugBank): NUT-related contributions on the intervention/mixed boards are systematically inflated — read with this caveat. In addition, disease labels themselves come from DrugBank approved indications (single source), and drugs not listed are implicitly treated as negatives — unapproved does not mean ineffective.

⚠ LOO measures marginal value relative to the reference pool: the same module changes value across pools (autophagy +0.0495 in the A1 pool vs −0.0105 in M1). Exact family-level Shapley removes this reference dependence; module-level Shapley is infeasible (332 modules cannot be enumerated). Negative LOO means removing the module IMPROVES recall (redundancy/interference in high-dimensional strongly-regularized pools), not measurement error.

⚠ Resolution floor: one recall hit = 0.05 per disease (0.053 for Athero, n_pos=19) / 0.010-0.011 on the 5-disease mean. LOO differences are hit-quantized with a minimum nonzero step of 0.010 — single-hit-scale differences may be pure protocol noise: borderline CV-LR ranks are sensitive to floating-point summation order, and a single hit can flip across execution environments (Shapley shares wobble by about ±0.003); family-share ORDER in the intervention pool (I3) was verified reproducible across two execution environments; mixed pools (M2/M3) were not independently recomputed, but their minimum adjacent share gap (0.008) exceeds twice the noise magnitude (2×0.003), so order flips are unlikely (gap-based inference).