Bacterial identification by metabolite-level interpretable surface-enhanced Raman spectroscopy

Surface-enhanced Raman spectroscopy (SERS), as a great potential label-free tool in metabolite detection, offers a strategy for rapid bacterial identification. However, it still lacks experimentally supported spectral interpretation at the metabolite level for complex biosamples. We present a SERS-based method for reliable bacterial intracellular metabolic profiling using plasmonic colloids with high rapidness and cost-efficiency. A convolutional neural network model was constructed to accurately classify eight types of bacteria with an overall accuracy as high as 90.44% and identify the key spectral features for classification by Shapley Additive Explanations. Molecule-level interpretation of the SERS metabolic profiles has been further realized in combination with laser desorption/ionization mass spectrometry, evidencing the primary metabolite contribution to the bacterial spectral signatures and molecule-level distinctions among different bacterial types. We provide insights into the mechanism of bacterial identification by label-free SERS and pave the way for interpretable SERS diagnostic tools for various diseases.