| Abstract: |
| Keloids are fibroproliferative skin disorders that arise following injury and extend beyond the original wound margins. Despite their clinical burden, their pathogenesis remains poorly understood, and current treatments are associated with high recurrence rates. Identifying biomarkers that uniquely distinguish keloids from other skin/scar types may provide insight into their underlying etiology and guide targeted therapeutic strategies. Previous studies have investigated candidate biomarkers; however, they are often limited by small sample sizes, restricted comparisons (e.g., keloid vs. hypertrophic scars), cross-platform inconsistencies, and reliance on conventional metrics such as fold-change, which may not capture the keloid uniquely distinguishing features. We present a rigorous, multi-layered machine learning framework to identify robust biomarkers that distinguish keloid from non-keloid samples. We curated and harmonized transcriptomic datasets (81 samples, 13 studies) across four tissue types: normal skin, normal scar, hypertrophic scar, and keloids, accounting for batch and technical variability. Using multiple classification algorithms with study-aware cross-validation, feature selection, stability analysis, and bootstrapping, we identified a minimal set of highly consistent genes with strong discriminatory power. Gene enrichment analysis of the five upregulated and three downregulated genes highlights increased lipid metabolism and decreased extracellular matrix organization as the dominant pathways affected, offering potential avenues for targeted treatment. |
|