Rapid identification of bacterial pathogens by Raman spectroscopy and transformer

Raman spectroscopy offers numerous advantages in bacterial identification, including rich molecular information, quick processing, and great sensitivity. However, accurately identifying bacterial species remains challenging due to the similarity of Raman spectra among various species. This paper introduces a method that combines Transformer networks and Raman spectra for the swift and precise identification of pathogenic bacteria. Our lightweight transformer model, called RamanFormer, outperforms conventional convolutional neural network (CNN) models in identification accuracy and model complexity on the Bacteria-ID dataset and Custom-built dataset. RamanFormer has only about 1/35 and 1/184 of the network parameters compared with CNNs. On the Bacteria-ID dataset, RamanFormer reached a state-of-the-art (SOTA) isolate-level accuracy of 87.03%. We also evaluated the model using clinical bacterial isolates and discovered that it had a SOTA of 99.98% identification accuracy in the 8-antibiotic empiric group task using just ten bacterial spectra per patient isolate. Additionally, RamanFormer also achieved 97.32% identification accuracy on the Custombuilt dataset. Our approach is thus capable of quickly and correctly classifying different bacterial pathogens based on the Raman spectra and could be used for additional Raman spectra identification tasks. The code for RamanFormer will be accessed at https://github.com/Bo-Zhou-gogogo/Raman-transformer.