Rapid and accurate identification of pathogenic bacteria phenotypic traits, including virulence, drug resistance, and metabolic activity, is essential for clinical diagnosis and infectious disease control. Traditional methods are time-consuming, highlighting the need for more efficient approaches. This study develops a single-cell Raman spectroscopy approach to detect multiple phenotypic traits of Staphylococcus aureus (S. aureus) as a proof of concept. We constructed a single-cell Raman spectral database encompassing 6240 spectra from 10 strains of S. aureus with diverse phenotypic traits and developed a convolutional neural network (CNN) to predict these phenotypes from the Raman spectra. The CNN model achieved 93.90%, 98.73%, and 98.66% accuracy in identifying enterotoxin-producing strains, methicillin-resistant S. aureus (MRSA), and growth stages, respectively. Characteristic Raman peaks for enterotoxin producers mainly appeared at 781, 939, 1161, 1337, 1451, and 1524 cm⁻¹, whereas MRSA primarily exhibited peaks at 723, 780, 939, 1095, 1162, 1340, 1451, 1523, and 1660 cm⁻¹. During culture, nucleic acid-related peaks weakened, lipid peaks increased, and protein peaks initially increased and subsequently decreased. This integration of Raman spectroscopy and machine learning demonstrates considerable potential for rapid bacterial phenotyping. Future research should expand to a wider range of bacterial species and phenotypes to enhance the diagnosis, prevention, and management of infectious diseases.

The discrepancy between the genotypic and phenotypic expression of enterotoxins in S. aureus had long been a significant challenge in toxin detection. However, the accurate and rapid application of Raman spectroscopy for the genotypic and phenotypic characterisation of S. aureus enterotoxins remains problematic. To address this, the present study utilised a single-cell Raman spectra database from 31 S. aureus isolates, acquired via a Raman laser tweezer system. When combined with convolutional neural network analysis, this approach achieved an average accuracy of 99.71% for identifying single-gene toxin types and 99.44% for multi-gene toxin types, with an average phenotypic identification accuracy of 98.71%. Notably, the phenotypic identification accuracy for the three strains carrying the sea and seb genes reached 100%, and the validation accuracy using unknown genotypes and phenotypes exceeded 85%. Furthermore, the CNN analysis identified characteristic spectral peaks for S. aureus enterotoxin genotypes at 1663–1665 cm⁻¹, 1570 cm⁻¹, and 1117–1119 cm⁻¹, corresponding to protein α-helices, guanine, and nucleic acid backbones respectively. Representative peaks for the phenotype were found at 1302–1314 cm⁻¹ and 912–923 cm⁻¹, corresponding to proteins/lipids and polysaccharides, respectively. Representative peaks for different virulence phenotypes carrying multiple enterotoxin genes were located at 1074–1076 cm⁻¹, 1253–1255 cm⁻¹, 1326 cm⁻¹, and 1327 cm⁻¹, corresponding to proteins, nucleic acids, and lipids, respectively. Furthermore, metabolomic analysis of three S. aureus strains (sea+seb+, sea+seb−, sea−seb+) revealed metabolic differences in fatty acids, purines, phenylalanine, and aspartic acid, consistent with the corresponding distinct Raman spectral peaks (1458, 1179, 1406–1409 cm⁻¹). Thus, this study employed S. aureus as a proof-of-concept, establishing for the first time a method combining Raman laser tweezers with convolutional neural networks for identifying S. aureus enterotoxin genotypes and phenotypes. It clarified the Raman spectral differential peaks and their corresponding biomarkers among five classical enterotoxin genotypes and phenotypic strains, providing a novel approach for accurate toxin typing and virulence characterisation.