Abstract
Surface-enhanced Raman scattering (SERS) spectroscopy represents a powerful analytical platform that combines non-destructive, label-free molecular identification with exceptional sensitivity for trace-level detection. Its capacity to generate information-rich spectral fingerprints makes SERS particularly advantageous for simultaneous multi-analyte analysis across diverse sample matrices, including complex biological systems. This study addresses the analytical challenges associated with identifying and quantifying multiple molecular species in complex environments by integrating SERS with advanced machine learning methodologies. We developed a hierarchical analytical framework that leverages the complementary strengths of deep learning and regression techniques: A multi-label convolutional neural network (CNN) for discriminating structurally similar analytes from SERS spectral data, coupled with a support vector regression (SVR) model for semi-quantitative determination of relative concentration ratios among identified species. The methodology was systematically validated using binary mixtures of short-chain fatty acids (SCFAs) as representative biomolecular targets, with performance rigorously benchmarked against established multivariate statistical methods and conventional machine learning approaches. Experimental validation demonstrated robust classification accuracy for both analytes at physiologically relevant concentrations, maintaining consistent performance across simple aqueous media and complex cell culture environments. These results establish the viability of the integrated SERS-CNN-SVR approach for advanced mixture analysis applications where precise identification and quantification of multiple biomarkers is essential.