Robust Single-Cell Raman Identification Enabled by Target-Interference Library

Foodborne pathogens, especially spore-forming Bacillus species, present a continuous challenge to food safety owing to their resilience against conventional sterilization methods and their ability to rapidly produce toxins. Single-cell Raman spectroscopy enables rapid and label-free detection of bacterial cells. However, its practical application is often limited by a low signal-to-noise ratio (SNR). This issue becomes particularly pronounced in food testing, where complex sample matrices and diverse interfering components can significantly elevate spectral background noise, thereby compromising analytical accuracy and reliability. To overcome these challenges, we propose a Target-Interference Library (TIL) Decomposition framework designed to actively learn the characteristic features from spectral libraries of the target bacteria and known interferences. The trained TIL framework is capable of decomposing noisy single-cell spectra into constituent target bacterial signals and interference components. By leveraging the isolated target bacterial signals, it significantly enhances the precision of identification. We applied this method to differentiate four closely related Bacillus species namely Bacillus licheniformis, Bacillus subtilis, Bacillus cereus, and Bacillus thermoamylovorans, achieving a classification accuracy of 97.1%, significantly surpassing conventional methods. Crucially, the method achieved a notably high identification accuracy of 96.12% for single spores within a real-world, complex milk matrix. This demonstration provides a practical pathway for rapid, on-site surveillance, addressing the pressing needs of modern food safety surveillance.