Boosting living <i>Bacillus</i> spore identification: Kolmogorov–Arnold network-guided convolutional neural network combined with laser tweezers Raman spectroscopy

As primary carriers of foodborne and zoonotic diseases, <i>Bacillus</i> spores can pose a serious threat to food microbiology and human disease. Thus, the precise identification of <i>Bacillus</i> spores is of great significance for ensuring food safety and human health. Herein, this study proposed a living <i>Bacillus</i> spore identification platform: an adaptive Kolmogorov-Arnold network (KAN)-guided convolutional neural network (CNN) configuration combined with laser tweezers Raman spectroscopy (LTRS). To address the small size of the original single-cell Raman spectral datasets, Gaussian noise-based spectra augmentation was employed to significantly enlarge and enrich it. When the adaptive KAN was introduced into the dense layer of CNN, the prediction accuracy of five <i>Bacillus</i> spore species was as high as 97.80% ± 1.79%. Moreover, the KAN-guided CNN configuration has strong robustness and generalization ability, providing a prediction accuracy of 96% for an independent spectral dataset. To figure out the classification contribution of each Raman band, a blocking individual Raman band method was proposed. The Raman band located at 1655 cm^-1, belonging to the amide I vibration of protein, was determined as the dominant contributor, surpassing two Raman bands belonging to Ca-DPA at 1576 cm^-1 and 1449 cm^-1. It can be foreseen that the KAN-guided CNN configuration combined with LTRS shows great promise for determining microbial identity, especially for unculturable microorganisms.