Lung adenocarcinoma (LUAD) remains a leading cause of cancer-related mortality worldwide, highlighting the urgent need for non-invasive strategies for early detection. Here, we present a machine learning-assisted metabolomics approach for the early detection of LUAD. Untargeted metabolomic profiling was performed on 199 serum samples from healthy individuals, patients with lung precancerous lesions, and those with stage I LUAD. An ensemble machine learning workflow was developed to identify metabolite panels capable of discriminating clinical status with high accuracy. We observed progressive metabolic alterations in bile acid, lipid, amino acid, and purine metabolism during LUAD initiation and stepwise progression. Notably, ensemble learning identified a six-metabolite panel, including 12-hydroxydodecanoic acid, hypoxanthine, xanthosine, cholic acid, agmatine, and paraxanthine, for accurate detection of early-stage LUAD, and a distinct four-metabolite panel, comprising 7-α,27-dihydroxycholesterol, 11-undecanedicarboxylic acid, biliverdin, and Prolyl-Valine, for precise differentiation between pre-invasive and invasive lesions. Both panels demonstrated promising diagnostic potential, with performance metrices comparing favorably to established methodologies within the current study cohort. This study delineates the evolutionary trajectory of the serum metabolome associated with early LUAD pathogenesis and provides promising biomarkers for non-invasive early detection.
Cai et al. (Wed,) studied this question.