Machine learning–enhanced label-free Raman spectroscopy for accurate serum-based detection of early-stage breast cancer via tumor-associated antigens

Abstract Early-stage detection of breast cancer remains a critical challenge for improving clinical outcomes. In this study, an experimentally label-free Raman spectroscopy approach combined with principal component analysis and supervised machine learning is reported for the detection of tumor-associated antigen signatures in serum samples. Raman spectra from early-stage and advanced breast cancer patients, as well as healthy controls, were analyzed alongside recombinant tumor-associated antigen proteins, including $$\alpha$$ -1 antitrypsin, peptidyl-prolyl cis-trans isomerase A, peroxiredoxin-2, and triosephosphate isomerase-1. Distinct spectral variations were observed in lipid-, nucleic acid-, and protein-related bands. In particular, a characteristic peak at $$\sim$$ 519 $$\mathrm {~cm}^{-1}$$ , associated with oxidized glutathione and oxidative stress-related structural alterations, was prominently observed in early-stage breast cancer and tumor-associated antigen spectra. Machine learning models trained on principal component analysis-reduced features, including linear discriminant analysis, support vector machine, decision tree, and Naïve Bayes, demonstrated high diagnostic performance, achieving ROC-AUC values above 0.94 for early-stage breast cancer versus controls and above 0.85 for early- versus advanced-stage discrimination. Furthermore, spectral and principal component analysis comparisons revealed partial clustering overlap between early-stage breast cancer samples and tumor-associated antigen proteins, supporting the biochemical relevance of the detected signatures. This integrated spectroscopic and computational approach provides a non-invasive and accurate strategy for early breast cancer detection and stage differentiation, highlighting its potential for clinical diagnostics.