Machine Learning-Assisted Plasma Metabolomics Identifies a Five-Metabolite Panel for Colorectal Cancer Detection

Abstract

Colorectal cancer (CRC) remains a leading cause of cancer-related mortality worldwide, and improved noninvasive strategies for early detection are urgently needed to complement current screening approaches. We applied an integrated plasma metabolomics workflow that combines untargeted and targeted liquid chromatography-mass spectrometry with machine-learning-assisted feature prioritization. Plasma samples from a discovery cohort of 172 CRC patients and 115 healthy controls were analyzed, followed by validation in an independent cohort of 47 CRC patients and 47 healthy controls. From 146,880 spectral features, random forest analysis was used to robustly rank features, followed by ROC-based prioritization and metabolite annotation. Diagnostic performance was evaluated by using a logistic regression model. Phenotypic assays were conducted to assess the biological activity of selected metabolites in CRC cell models. Five metabolites, N-methylcytisine, 2-piperidone, theophylline, dl-norleucine, and linolenic acid, were consistently reduced in CRC patient plasma across both cohorts. The integrated five-metabolite panel achieved an area under the ROC curve of 0.968 in the validation cohort with a 97.9% sensitivity, an 89.4% specificity, and a 93.7% accuracy. Stratified analyses demonstrated robustness across disease stages and age groups. In vitro assays showed modulation of CRC cell migration and invasion under noncytotoxic conditions. This five-metabolite plasma signature reflects CRC-associated systemic metabolic alterations and demonstrates a strong discriminatory performance. The panel may complement existing screening modalities by contributing to CRC risk stratification and early detection.