Metabolomic Subtyping and Machine Learning-Based Diagnosis Reveal Clinical Heterogeneity in Silicosis

Silicosis remains a major occupational health concern worldwide and is characterized by notable clinical heterogeneity in terms of disease progression and complications. However, the underlying metabolic mechanisms contributing to this heterogeneity remain poorly understood.

We conducted a case-control study involving 156 silicosis patients and 132 silica-exposed controls. The plasma samples were analyzed via untargeted metabolomics based on liquid chromatography-mass spectrometry (LC-MS/MS). To explore disease subtypes and potential biomarkers, we applied non-negative matrix factorization (NMF) clustering, weighted gene co-expression network analysis (WGCNA), and machine learning approaches.

A total of 860 differentially abundant metabolites, including elevated pathogen-associated compounds, were identified in silicosis patients. Unsupervised NMF clustering revealed two distinct metabolic subtypes with different clinical features. Patients in the NMF2 subgroup had a 5.3-fold greater risk of pulmonary infections (p = 0.026) than those in the NMF1 subgroup. Metabolomic analysis revealed that NMF2 was enriched in arachidonic acid and unsaturated fatty acid metabolism pathways, with prominent LysoPC accumulation, suggesting inflammation-related lipid peroxidation. In contrast, NMF1 was characterized by increased spermidine biosynthesis and urea cycle activity, along with suppressed saturated fatty acid metabolism and reduced LysoPC processing, potentially affecting membrane integrity and promoting fibrosis. A machine learning-derived dual-metabolite panel, tyrosocholic acid and PI (20:4/0:0), achieved AUC values above 0.85 for both silicosis detection and subtype classification.

These findings highlight metabolic heterogeneity in silicosis and suggest clinically relevant subtypes, providing a foundation for improved stratification, early detection, and targeted interventions.