Deep Learning-Enabled Raman Spectroscopic Identification of Pathogen-Derived Extracellular Vesicles and the Biogenesis Process.

Pathogenic bacterial infections, exacerbated by increasing antimicrobial resistance, pose a major threat to human health worldwide. Extracellular vesicles (EVs), secreted by bacteria and acting as their "long-distance weapons", play an important role in the occurrence and development of infectious diseases. However, no efficient methods to rapidly detect and identify EVs of different bacterial origins are available. Here, label-free Raman spectroscopy in combination with a new deep learning model of the attentional neural network (aNN) was developed to identify pathogen-derived EVs at Gram ± , species, strain, and even down to physiological levels. By training the aNN model with a large Raman data set from six typical pathogen-derived EVs, we achieved the identification of EVs with high accuracies at all levels: exceeding 96% at the Gram and species levels, 93% at the antibiotic-resistant and sensitive strain levels, and still above 87% at the physiological level. aNN enabled Raman spectroscopy to interrogate the bacterial origin of EVs to a much higher level than previous methods. Moreover, spectral markers underpinning EV discrimination were uncovered from subtly different EV spectra via an interpretation algorithm of the integrated gradient. A further comparative analysis of the rich Raman biochemical signatures of EVs and parental pathogens clearly revealed the biogenesis process of EVs, including the selective encapsulation of biocomponents and distinct membrane compositions from the original bacteria. This developed platform provides an accurate and versatile means to identify pathogen-derived EVs, spectral markers, and the biogenesis process. It will promote rapid diagnosis and allow the timely treatment of bacterial infections.