Mahir Dığış, Kısmet Kaya, Betul Demir
Validation and robustness checks converted an apparently supported mechanistic network into a result with no surviving gene-level candidate. The most consequential was a library default: within-cohort ROC direction selection turned a contradicted gene into the apparently strongest validation. Such checks should be standard, not optional. No causal or exposure-related association was demonstrated.
BACKGROUND: Network-toxicology and molecular-docking pipelines are widely used to link environmental chemicals to disease targets, but their outputs are rarely subjected to external validation, background-aware significance testing, or promiscuity controls.
METHODS: This fully computational (in silico) study applied a conventional network-toxicology and molecular-docking pipeline to vitiligo and eleven compounds-caffeine, four antibiotics, two angiotensin-receptor blockers, two statins, a proton-pump inhibitor and an antidepressant-adopted verbatim from a published study, holding chemical input constant. The output was stress-tested by core-tier threshold sensitivity, external cross-platform validation in two independent cohorts with the receiver-operating-characteristic (ROC) direction fixed in advance, background-restricted overlap-significance testing, and caffeine-exclusion sensitivity analysis.
RESULTS: The 102-gene target set did not intersect the high-confidence vitiligo core tier (overlap = 0, stable above 0.170) and gave a 22-gene extended-tier intersection, in which HIF1A was the highest-degree node and five key differentially expressed genes discriminated lesional from non-lesional skin at areas under the ROC curve (AUC) of 0.73-0.87. Every check degraded this. The overlap was enriched against a protein-coding background (fold 2.49, p = 5.1×10-5) but not against the 5601-gene mineable universe (fold 1.29, p = 0.12). External discrimination did not replicate: with the direction fixed, seven of ten gene-cohort estimates fell below 0.50, none with a lower confidence bound above 0.50; HIF1A fell to 0.32 and 0.45, and EDNRA, apparently validating at 0.96 under automatic direction selection, reversed reproducibly to 0.04 (95% CI 0.00-0.24). Caffeine exclusion left two genes. No gene cleared all four checks.
CONCLUSION: Validation and robustness checks converted an apparently supported mechanistic network into a result with no surviving gene-level candidate. The most consequential was a library default: within-cohort ROC direction selection turned a contradicted gene into the apparently strongest validation. Such checks should be standard, not optional. No causal or exposure-related association was demonstrated.