Donor-aware cross-cohort transcriptomics prioritizes P2RX4 in human diabetic retinopathy

Small public retinal cohorts can generate unstable candidates unless donor dependence, clinical covariates, and external checks are separated. We analysed GSE160306 macular RNA sequencing using 26 donors with diabetes and a predefined 158-gene inflammatory space. Total-association and DME-conditioned severity models used HC3 finite-sample t inference, 4,999 wild-bootstrap-t replicates, 2,000 donor bootstraps, deletion and influence diagnostics, and alternative severity encodings. P2RX4 ranked first in both primary estimands (total-association β = 0.688, P = 0.00149; wild-bootstrap P = 0.0008), although no gene passed 158-gene FDR in either primary model and its bootstrap top-five frequency was 61.7%. A post-protocol reconstruction of GSE276892 raw reads (8 PDR, 9 surgical controls) yielded a positive but imprecise disease-only negative-binomial estimate (log2 fold change 0.295, 95% CI -0.854 to 1.443; two-sided Wald P = 0.615); source, age/sex, reference, and low-expression-sample sensitivities remained inconclusive. In post-protocol GSE179568 membranes, the unadjusted rank result favoured PDR (P = 0.00967), but PDR and controls had no age overlap and the age/sex-adjusted HC3 estimate was null (β = 0.102, P = 0.773). Thus P2RX4 is a reproducibly prioritized hypothesis with heterogeneous ocular-compartment observations, not a validated disease effect, cell source, biomarker, or mechanism.