C23-20 Integrated Proteomic and MicroRNA Profiling of Tracheal Lavage Fluid-derived Extracellular Vesicles Using Machine Learning to Identify Molecular Signatures of Tracheal Repair

Background Tracheal injury caused by intubation, tracheostomy, or tracheobronchial tuberculosis often leads to epithelial disruption and aberrant repair, resulting in fibrotic stenosis that severely impairs airway function. Bronchoscopic observations revealed excessive tracheal secretions in injured patients. Our previous study identified abundant extracellular vesicles (EVs) in these secretions, suggesting their potential role in tracheal repair. This study profiled EVs isolated from tracheal lavage fluid (TLF) using integrated proteomics and miRNA sequencing combined with machine learning to systematically delineate molecular features and regulatory mechanisms underlying distinct types of tracheal injury. Methods TLF samples were collected from patients with mechanical or infectious tracheal injury and healthy controls across multiple centers. EVs were isolated by ultracentrifugation and validated by TEM, nanoparticle tracking, and Western blotting. LC–MS/MS and small RNA sequencing were performed to generate proteomic and miRNA datasets. Principal component analysis (PCA) and t-SNE were used to identify injury-specific patterns. LASSO regression and random forest models were applied to screen key molecular features. Integrated analyses constructed a “protein–miRNA–pathway” interaction network to elucidate mechanisms of tracheal repair. Results In the discovery cohort (n = 69), 2,143 EV proteins were identified, with 187 upregulated and 146 downregulated in injured groups. Differential proteins were enriched in epithelial repair, immune response, and extracellular matrix remodeling pathways. The random forest model based on proteomic features achieved an AUC of 0.91 for classifying mechanical versus infectious injury, while LASSO identified 15 signature proteins, with DAD1, P2RY13, and LILRA3 showing the highest weights. miRNA sequencing detected 326 differentially expressed miRNAs, and PCA/t-SNE revealed distinct clustering among groups. The miRNA-based random forest model reached an accuracy of 0.88 and an AUC of 0.89, highlighting miR-223-3p and miR-29a-3p as major contributors. Integrated analysis identified 127 significantly negative protein–miRNA correlations, mainly involved in epithelial regeneration, inflammation regulation, and matrix remodeling. Notably, the miR-223-3p–FOXO1 and miR-29a-3p–COL12A1 axes exhibited injury type–specific patterns (FDR<0.01). Conclusion This study provides a comprehensive proteomic and miRNA atlas of TLF-derived EVs from patients with distinct tracheal injuries. Machine learning identified DAD1, P2RY13, LILRA3, miR-223-3p, and miR-29a-3p as key molecular signatures. Integrated multi-omics analysis reveals that EV-associated molecules are tightly linked to epithelial repair, immune modulation, and extracellular matrix remodeling, offering new insights into tracheal injury repair mechanisms and potential biomarkers.