Abstract

Background: Breast cancer remains one of the most common causes of cancer mortality in women globally. Although therapeutic options have improved, resistance to treatment continues to limit clinical success, high lighting the need for innovative approaches. One defining feature of malignancy is metabolic reprogramming. In breast cancer, tumor cells exhibit metabolic patterns that differ markedly from those of normal tissue, and characterizing these differences may help uncover new therapeutic targets aimed at disrupting tumor growth and improving outcomes.

Objectives: This study evaluated whether multiple machine learning methods could differentiate the metabolic signatures of breast cancer patients from those of healthy individuals, with the goal of identifying metabo lite-based targets relevant to precision oncology.

Methods: Plasma samples from 102 women with breast cancer and 99 control participants were profiled using targeted liquid chromatography-tandem mass spectrometry (LC-MS/MS). Six classification algorithms were trained and compared, and metabolite contributions to model performance were assessed using mean squared error-based importance measures.

Results: Breast cancer samples showed reduced levels of several amino acids, including alanine, histidine, tryptophan, tyrosine, methionine, and proline. Among the evaluated models, Random Forest demonstrated the strongest performance (accuracy = 0.90, specificity = 0.85, sensitivity = 0.95). K- Nearest Neighbors produced similar sensitivity but lower specificity, while Logistic Regression provided balanced performance (specificity = 0.90, sensitivity = 0.86; accuracy = 0.88). Naive Bayes and Support Vector Machine yielded in termediate accuracy (0.83). The Decision Tree model had the lowest sensitivity (0.76) but the highest positive predictive value (0.89). Feature-importance analysis consistently identified glutamic acid, ketocholesterol, cystine, ornithine, succinate, acetylcarnitine, asparagine, tryptophan, and palmitic acid as influential metab olites.

Conclusion: Machine learning-based metabolic profiling revealed several metabolites that may represent ac tionable metabolic constraints in breast cancer. These findings support the potential of metabolomics-driven modeling to inform targeted interventions and individualized therapeutic strategies.

DOI: doi.org/10.63721/26JPMHC0115

To Read or Download the Article  PDF