Non-targeted metabolomics and machine learning reveal metabolic dysregulation in lymph node tuberculosis

Lymph node tuberculosis (LNTB) is the most prevalent form of extrapulmonary tuberculosis; however, differentiating it from non-LNTB remains challenging due to overlapping clinical features and suboptimal diagnostic methods. Current diagnostic methods for LNTB lack both sensitivity and specificity. This study aimed to characterize the metabolic differences between LNTB and non-LNTB patients, elucidate the pathological mechanisms underlying LNTB, and identify diagnostic biomarkers using machine learning models.

Serum samples from 40 LNTB patients and 30 non-LNTB patients were analyzed using ultra-high-performance liquid chromatography-mass spectrometry. Differential metabolites were identified based on a variable importance in projection >1, false discovery rate-adjusted p-value <0.05. Pathway enrichment analysis was performed using the Kyoto Encyclopedia of Genes and Genomes (KEGG). Machine learning, including support vector machines and random forest, were employed to screen for diagnostic biomarkers, which were validated by receiver operating characteristic curves.

Among the 1294 detected metabolites, 89 exhibited significant differences between the two groups. By integrating KEGG enrichment with topological analysis, phenylalanine, tyrosine, and tryptophan biosynthesis possessed the highest impact, followed by phenylalanine metabolism, and aminoacyl-tRNA biosynthesis. Machine learning identified four biomarkers: Leu-Ala [area under the curve (AUC) = 0.8292], evodiamine (AUC = 0.7558), fenazaquin (AUC = 0.7175), and acetol (AUC = 0.7117). Leu-Ala demonstrated the highest diagnostic accuracy, with a sensitivity of 73.5 % and specificity 86.7 % at a cutoff value of 0.62.

Untargeted metabolomics revealed dysregulation in the biosynthesis of phenylalanine, tyrosine, and tryptophan, phenylalanine metabolism, as well as in aminoacyl-tRNA biosynthesis in LNTB. Additional, Leu-Ala was identified as a novel diagnostic biomarker. The integrating of metabolomics with machine learning presents a promising approach for LNTB detection, though larger validation studies are necessary.