Combining metabolomics and machine learning to discover biomarkers for early-stage breast cancer diagnosis

There is an urgent need for better biomarkers for the detection of early-stage breast cancer. Utilizing untargeted metabolomics and lipidomics in conjunction with advanced data mining approaches for metabolism-centric biomarker discovery and validation may enhance the identification and validation of novel biomarkers for breast cancer screening. In this study, we employed a multimodal omics approach to identify and validate potential biomarkers capable of differentiating between patients with breast cancer and those with benign tumors. Our findings indicated that ether-linked phosphatidylcholine exhibited a significant difference between invasive ductal carcinoma and benign tumors, including cases with inconsistent mammography results. We observed alterations in numerous lipid species, including sphingomyelin, triacylglycerol, and free fatty acids, in the breast cancer group. Furthermore, we identified several dysregulated hydrophilic metabolites in breast cancer, such as glutamate, glycochenodeoxycholate, and dimethyluric acid. Through robust multivariate receiver operating characteristic analysis utilizing machine learning models, either linear support vector machines or random forest models, we successfully distinguished between cancerous and benign cases with promising outcomes. These results emphasize the potential of metabolic biomarkers to complement other criteria in breast cancer screening. Future studies are essential to further validate the metabolic biomarkers identified in our study and to develop assays for clinical applications.