Boosting identification of microsporidian spores originating from different hosts: single-cell Raman spectroscopy combined with self-attention mechanism-driven convolutional neural network.

As a class of special intracellular parasites, the microsporidian pathogens parasitized in various hosts are shown to be a serious threat to agriculture production. Therefore, precise identification of microsporidian pathogens is crucial for controlling microsporidian-related agriculture diseases. However, conventional identification methods have shown limitations including low sensitivity, destructive operation, and complicated preprocessing. We proposed an advanced identification platform that integrates single-cell Raman spectroscopy with a self-attention mechanism (SAM)-driven convolutional neural network (CNN) configuration, which can realize convenient, non-destructive, high-precision identification of microsporidian spores from 11 various host sources at a single-cell resolution level. Considering that yielded microsporidian spores are difficult to cultivate, an interpolation algorithm-based spectra shifting approach was proposed to significantly enlarge the size of single-cell Raman spectra datasets, overcoming possible overfitting caused by training small samples of original Raman spectra datasets of microsporidian spores. Owing to the collaboration of both SAM and spectra augmentation, the averaged prediction accuracy of microsporidian spores from 11 various hosts can be significantly enhanced from 88.17% ± 1.05% provided by a single optimal CNN model to be as high as 95.16 ± 1.61% provided by the SAM-driven CNN configuration. To figure out which spectral features contributed to such high prediction accuracy, the global spectral features were systematically extracted by the SAM curve. These four highlighted Raman bands located at 541, 718, 915, and 1081 cm