Label-Free and Strain-Resolved Characterization of AI-2 Associated Quorum Sensing Functional Phenotypes in Bacterial Isolates Using Single-Cell Raman Spectroscopy and Deep Learning

Abstract We evaluated the relationship between Raman spectral profiles obtained by single-cell Raman spectroscopy (SCRS) and autoinducer-2 (AI-2)-associated quorum sensing (QS) functional phenotypes using experimentally measured relative BB170 reporter responses as functional reference values. Based on this relationship, this study provides a rapid, label-free SCRS-based model for characterizing AI-2-associated QS functional phenotypes in bacterial isolates. The optimized residual neural network (ResNet) achieved an R2 of 0.8351 under spectrum-level 10-fold cross-validation, while strain-level aggregated evaluation showed good agreement with experimentally measured responses within the represented dataset (R2 = 0.8840). In an independent-isolates set of 21 strains from multiple sources spanning the No-, Low-, Medium-, and High-response categories, the model achieved a mean absolute error (MAE) of 0.092, a root mean squared error (RMSE) of 0.148, an R2 of 0.805, and an overall category accuracy of 80.95%. Reverse transcription quantitative polymerase chain reaction (RT-qPCR) detected luxS expression in all 21 independent isolates, providing complementary evidence for their potential AI-2-associated signaling capacity. In addition, furanone C-30 treatment of W37 resulted in concordant decreases in experimentally measured and model-estimated relative BB170 reporter responses, further supporting the responsiveness of the captured spectral phenotype to within-strain functional perturbation. Integrated Gradients (IG) analysis highlighted coordinated Raman spectral regions associated with proteins, lipids, polysaccharides, nucleic acids, and sulfur-containing groups, suggesting that model estimates were based on multiband spectral patterns rather than a single characteristic feature.