Cerebrospinal Fluid Metabolomics and Machine Learning Identify Novel Biomarkers for Lung Cancer Leptomeningeal Metastasis
Menée à partir d'un algorithme d'apprentissage automatique ainsi que d'une analyse, par spectrométrie de masse haute résolution, d'échantillons de fluide cérébrospinal prélevés sur 218 patients atteints d'un cancer du poumon puis validée à partir d'échantillons sériques prélevés sur 233 patients complémentaires, cette étude identifie une signature métabolomique pour détecter précocement des métastases leptoméningées
BACKGROUND : Lung Cancer Leptomeningeal Metastasis (LC-LM) severely impacts patient survival and quality of life, yet current diagnostic methods lack sufficient sensitivity and specificity, particularly for early detection. Cerebrospinal fluid (CSF) metabolomics may reveal specific biomarkers reflecting brain metastasis.
METHODS : We performed untargeted metabolomic profiling of CSF samples by high-resolution mass spectrometry (HRMS) in a cohort of 218 participants, including 99 samples from LC-LM (with cancer cells detected in the CSF), 12 samples from the lung cancer parenchymal brain metastases (with no cancer cells detected in the CSF), 27 samples from the control group, 21 samples from the breast cancer LM, 15 samples from patients with LM from other tumors such as melanoma and gastric cancer, and 36 samples from other diseases. Significant metabolites were identified and validated. Subsequently, targeted metabolomics was conducted on serum samples from an independent cohort (n = 233), including 50 LC-LM patients, 150 patients with primary lung cancer (stages I-III), and 33 benign pulmonary nodules.
RESULTS : Untargeted CSF metabolomics revealed a distinct metabolic signature in LC-LM patients. Differential analysis identified metabolites significantly altered in LC-LM, notably elevated lactic acid, N1, N12-diacetylspermine, and altered amino acid metabolites (e.g., L-proline, L-glutamic acid), each demonstrating strong diagnostic accuracy individually, with area under the receiver operating characteristic (ROC) curve (AUC) > 0.90. Machine learning classification models based on CSF metabolite panels achieved perfect diagnostic performance (AUC = 1.00) in distinguishing LC-LM from controls and other groups. Targeted validation of five top metabolites in serum samples confirmed their diagnostic utility, with N1, N12-diacetylspermine achieving an AUC of 0.882, superior to traditional protein biomarkers.
CONCLUSION : CSF-based metabolomic profiling combined with machine learning offers a highly accurate and minimally invasive diagnostic tool for LC-LM. Serum validation further supports its translational potential, emphasizing its significance in clinical practice for improving early detection and potentially enhancing patient management and outcomes.
Neuro-Oncology , résumé, 2025