Machine learning-enhanced direct mass spectrometry analysis of non-volatile breath metabolites for rapid and accurate lung cancer screening

Abstract

Breath analysis by direct mass spectrometry faces significant challenges due to the inherent complexities in sample collection, low analyte concentrations, and accurate compound identification. While current breath analysis primarily focuses on volatile organic compounds (VOCs) for disease research, non-volatile organic compounds (nVOCs) remain largely unexplored despite their diagnostic potential. Here, we present a novel breath analysis method for lung cancer diagnosis based on nVOCs, integrating non-invasive breath analysis with machine learning algorithms for comprehensive characterization of 98 clinical breath samples. This study leverages a machine learning-driven database docking methodology to overcome the bottleneck of metabolite direct mass spectrometry conventional identification. This approach enables rapid and precise screening of non-volatile differential metabolites while effectively excluding exogenous confounders (e.g., pharmacological or environmental interference), enhancing nVOC detection in breath. The approach identified 29 statistically significant nVOC biomarkers, including fatty acids and amino acids, achieving a 0.9878 prediction accuracy for lung cancer detection. For distinguishing between NSCLC and SCLC, the area under the curve (AUC) value can reach 0.9, and the out-of-bag error of random forest is 0.00402. Notably, specific nVOCs including fatty acids and amino acids have high diagnostic potential, with an AUC of up to 0.67 of individual metabolites for the differentiation of SCLC from NSCLC. Finally, significantly altered metabolic pathways were explored by metabolite pathway and transcriptome analysis, showing that the fatty acid metabolism is a potentially regulatable pathway. Our approach facilitates rapid, non-invasive discrimination of NSCLC and SCLC in metabolic analysis, showing promise as an efficient, low-cost clinical test.