Estimand-aware and donor-aware triangulation of predefined gene-set signals across human tendon transcriptomic datasets

We analysed five predefined programmes across three human tendon transcriptomic contexts: acute exercise-associated fibroblast pseudobulks (E-MTAB-15400; 4 control and 4 exercise samples), mechanically stretched donor-derived cells (GSE150482; 3 normal and 3 tendinopathy donors), and paired lesional versus grossly normal-appearing tendon (GSE26051; 23 donor pairs). Inference was conducted at the sample, donor or donor-pair level. Non-stimulated GSE150482 single-cell RNA sequencing comprised only N1 and D1, whereas stretched single-cell RNA sequencing comprised N1-N3 and D1-D3; the conditions also used Drop-seq and 10X Chromium, respectively. A donor-replicated disease-by-stimulation interaction is therefore not estimable. Within-context Benjamini-Hochberg adjustment supported acute mechanical-response, extracellular-matrix (ECM) and protein-folding differences and paired-lesion collagen, integrin and ECM differences, but no stretched donor-background difference. Exact label permutations, a conservative global 15-test adjustment, leave-one-unit analyses, a Fib1-Fib4 pooled sensitivity and a paired exact Wilcoxon sensitivity qualified these findings. Integrin was fully nested in ECM, while collagen overlapped 52/79 genes with ECM. Collagen fibril organization showed the most directionally consistent matrix-associated pattern across the three non-equivalent contrasts, without establishing a disease-specific loading response or causal mechanism. The analysis demonstrates why public tendon datasets must be integrated by estimand and biological inference unit rather than pooled as interchangeable replications.