Provenance-Preserving Cross-Tissue Transcriptomic Prioritisation of Predicted Metabolite Targets Across The Diabetic Gut-Kidney Axis

Abstract

Background: Computational target-prioritisation studies can become misleading when prediction, species-specific transcriptomics, network connectivity and pathway enrichment are merged without preserving the evidential boundary of each layer. We developed a provenance-preserving workflow to prioritise predicted metabolite targets across gut and kidney transcriptomes in diabetes while retaining discordant and null findings.

Methods: An audited human-target prediction workflow was identifier-standardised before independent analysis of three public transcriptomic datasets: human jejunal enteroendocrine-enriched tissue (DS001; GSE132831), supportive mouse ileum (DS002; GSE210876) and human diabetic-nephropathy glomeruli (DS003; GSE96804). Dataset-specific differential-expression evidence was retained without cross-dataset effect-size pooling or combined P values. Direction status was assigned only when at least two datasets robustly supported a target. A predefined 31-target human gut-kidney set was examined using high-confidence STRING functional and physical network contexts and controlled GO, KEGG and Reactome over-representation analysis.

Results: Eight eligible prediction runs yielded 800 preserved records. Deduplication and component-level identifier verification resolved 449 unique approved human genes, all retained as computationally predicted targets. Robust FDR support was observed for 113 targets in DS001, 19 in supportive DS002 and 209 in DS003. Across the 449-target universe, 31 targets were direction-concordant, 22 were direction-discordant and 396 had insufficient multi-dataset support for direction assessment. The 31-target functional STRING network contained five associations connecting nine submitted seeds while 22 remained isolates; the matched physical-network sensitivity query returned no edges. No GO, KEGG or Reactome term survived the prespecified Benjamini-Hochberg FDR threshold.

Conclusions: Cross-tissue evidence can narrow a broad predicted-target universe without converting computational prediction, overlap, network connectivity or nominal enrichment into validation. The resulting target set is hypothesis-generating; sparse network structure and null FDR-controlled enrichment argue for target-level, rather than pathway-level, interpretation.