Journal Article

Plasma proteomic and machine learning models for differentiating idiopathic pulmonary fibrosis and connective tissue disease–associated interstitial lung disease: findings from a prospective cohort

Yu, Yi-Hsuan, Chen, Yi-Ching, Lai, De-Wei, Wu, Chen-Shiou, Fu, Pin-Kuei, Yang, Lan-Yan

2026

Overview

Background Idiopathic pulmonary fibrosis (IPF) is a progressive fibrotic interstitial lung disease (ILD) with limited treatment options and poor prognosis. Differentiating IPF from connective tissue disease–associated ILD (CTD-ILD) is clinically challenging due to overlapping features, and reliable circulating biomarkers are lacking. Recent studies suggest that multi-marker proteomic models combined with machine learning may enhance diagnostic precision and prognostic assessment in fibrotic ILDs. Methods We prospectively analyzed plasma samples from Taiwanese patients with fibrotic ILDs (IPF, n  = 22; CTD-ILD, n  = 66) using the Olink inflammation panel (92 proteins). Differentially expressed proteins were identified and subjected to integrative network analyses. Predictive classification models were developed using generalized linear modeling (GLM), decision tree, and random forest approaches. Prognostic relevance was evaluated with Kaplan–Meier and Cox regression analyses, and findings were validated in public transcriptomic datasets. Results Among 92 proteins profiled, 23 showed significant differences between IPF and CTD-ILD. Four candidates—MMP-10, FGF-19, ADA, and TWEAK (TNFSF12)—consistently emerged as key discriminatory markers. The GLM model incorporating FGF-19, ADA, and TWEAK achieved the highest diagnostic accuracy (AUC 0.870; sensitivity 0.97; specificity 0.82), outperforming decision tree and random forest models. Transcriptomic validation confirmed TWEAK downregulation in ILD lung tissues and in TGF-β1–stimulated fibroblasts, linking it to canonical profibrotic signaling. Survival analysis showed significantly worse outcomes in IPF versus CTD-ILD (log-rank p  < 0.001), with MMP-10 associated with poor prognosis (HR 2.08, p  = 0.007) and TWEAK with favorable prognosis (HR 0.04, p  < 0.001). Conclusions This study identifies distinct plasma proteomic signatures that differentiate IPF from CTD-ILD and highlights TWEAK as both a diagnostic and prognostic biomarker. A multi-marker GLM model demonstrated excellent diagnostic performance, supporting the clinical utility of plasma proteomics combined with machine learning to improve disease classification and risk stratification in fibrotic ILDs. Clinical implication Integrating plasma proteomics with machine learning enables development of a multi-marker model that enhances diagnostic accuracy and prognostic evaluation in fibrotic interstitial lung diseases, supporting more precise disease classification and risk stratification in clinical practice. Key Point Plasma proteomics combined with machine learning improves IPF vs CTD-ILD classification and identifies prognostic biomarkers, advancing precision diagnosis in fibrotic interstitial lung disease.

• Publisher: BioMed Central,BioMed Central Ltd,Nature Publishing Group,BMC
• Subject: Accuracy / Biomarkers / Care and treatment / Classification / Connective tissue diseases / Connective tissue disease–associated interstitial lung disease / Connective tissues / Datasets / Decision trees / Diagnosis / Fibroblast growth factors / Fibrosis / Generalized linear models / Growth factors / Idiopathic pulmonary fibrosis / Inflammation / Learning algorithms / Lung diseases / Lung diseases, Interstitial / Machine learning / Mann-Whitney U test / Medical prognosis / Medical research / Medicine / Medicine & Public Health / Medicine, Experimental / Multi-marker predictive models / Network analysis / Patients / Physiological aspects / Plasma / Plasma proteomics / Pneumology/Respiratory System / Prognosis / Protein expression / Proteins / Proteomics / Pulmonary fibrosis / Regression analysis / Stromelysin 2 / Surfactants / Transcriptomics / Transforming growth factor-b1 / Tumor necrosis factor-TNF / TWEAK