Rapid and accurate identification of marine bacteria spores at a single-cell resolution by laser tweezers Raman spectroscopy and deep learning

Marine bacteria have been considered as important participants in revealing various carbon/sulfur/nitrogen cycles of marine ecosystem. Thus, how to accurately identify rare marine bacteria without a culture process is significant and valuable. In this work, we constructed a single-cell Raman spectra dataset from five living bacteria spores and utilized convolutional neural network to rapidly, accurately, nondestructively identify bacteria spores. The optimal CNN architecture can provide a prediction accuracy of five bacteria spore as high as 94.93% ± 1.78%. To evaluate the classification weight of extracted spectra features, we proposed a novel algorithm by occluding fingerprint Raman bands. Based on the relative classification weight arranged from large to small, four Raman bands located at 1518, 1397, 1666, and 1017 cm^-1 mostly contribute to producing such high prediction accuracy. It can be foreseen that, LTRS combined with CNN approach have great potential for identifying marine bacteria, which cannot be cultured under normal condition.