Cross-platform transportability of ligand–receptor communication scores for prognosis in lung adenocarcinoma: a systematic benchmark of three scoring constructs across four cohorts

Abstract Ligand–receptor communication scores derived from bulk tumour transcriptomes have been proposed as prognostic markers in lung adenocarcinoma, but whether they survive a platform change has never been benchmarked head-to-head. We quantified communication competence for 2,459 measurable CellChatDB v2 pairs in four cohorts (1,355 tumours, 529 deaths) using three scoring constructs — a z-based construct (Z-CP), a rank co-elevation construct invariant to monotone transformations (RCE) and a cell-type-resolved construct built on marker-anchored deconvolution (CTRC) — with nested cross-validation, formal equivalence testing (two one-sided tests, margin ΔC = 0.05) and a graded platform-shift simulation. In the discovery cohort the constructs performed comparably (cross-validated C-index 0.543–0.585); in the three external cohorts every construct landed in the same modest band (C-index 0.46–0.56). RCE was statistically equivalent to Z-CP in both adequately powered cohorts (TOST p = 0.011 and 0.015) and directionally superior in a third; CTRC lost discovery power and bought no transport benefit. Under simulated platform shift all constructs degraded smoothly and comparably, with no rescue regime. Bulk-derived ligand–receptor scores transport weakly and construct-independently; future benchmarks should report formal equivalence rather than post hoc superiority claims.