Betanin Promotes Wound Closure and Drives a Context-Specific Transcriptional Repair Program in HaCaT Keratinocytes: A Multi-Evidence AI-Guided Prioritization Study

Abstract

Keratinocytes coordinate re-epithelialization, barrier restoration, inflammatory control, and matrix remodeling during cutaneous repair. Although betanin has reported cytoprotective and antioxidant properties, its wound-relevant transcriptional program in keratinocytes remains incompletely defined. This study therefore asked whether betanin engages barrier, inflammatory, and remodeling-related keratinocyte programs under inflammatory conditions, and whether existing wound-healing databases can account for the resulting transcriptional pattern. Scratch closure was quantified by ImageJ measurement and by an independent rule-based automated image-analysis pipeline with first-order kinetic modeling; transcriptional responses in IFN-γ/TNF-α-stimulated HaCaT cells were interpreted through database-weighted evidence scoring and machine learning-based stress testing. Betanin (20 μg/mL) significantly reduced residual wound area at 24 h (p = 0.020) and 48 h (p = 0.035) by ImageJ quantification, and automated image analysis with kinetic modeling yielded an approximately 1.5-fold higher first-order closure rate constant relative to the vehicle-treated wounded control. Betanin partially restored FLG, reduced KRT14, and selectively attenuated inflammatory mediators including IL1B, ICAM1, CCL22, and CXCL8. Among remodeling-associated genes, TGFB1, COL1A1, and VEGFA were suppressed while COL3A1 was partially restored. NFE2L2 and HMOX1 were only modestly affected, indicating a dominant anti-inflammatory and barrier-modulating response rather than canonical antioxidant axis activation. Database-weighted evidence scoring stably prioritized FLG, COL3A1, COL1A1, and ICAM1, corroborated by betanin-specific experimental responses. Three architecturally distinct machine learning models showed no evidence that database-derived wound-healing features predict betanin-induced transcriptional responses, indicating that the selected prior-knowledge annotations were insufficient to account for the observed pattern in this 15-gene, single-context dataset.