We propose an innovative technology based on the combination of Raman microspectroscopy and deep learning to classify the Mechanism of Action (MoA) of antimicrobials and predict their novelty. Raman microspectroscopy provides chemical and physical signatures of molecular structures in bacteria to reveal phenotypic responses to antimicrobials. Deep learning techniques are powerful tools to extract discriminative features from complex Raman spectra and classify them. We developed the RaMoA technology and assessed its performance exclusively on the wild-type Escherichia coli (E. coli) ATCC 25922, after 1 h of treatment with 27 antibiotics representing 5 conventional functional classes (i.e., 12 cell wall synthesis inhibitors, 9 protein synthesis inhibitors, 3 DNA replication inhibitors, 2 RNA synthesis inhibitors, and 1 cell membrane function inhibitor). First, using preprocessed Raman spectra as input to a 1D Convolutional Neural Network, we classified Raman spectra of the treated bacteria into the correct MoA class with 96% accuracy. Aggregation of spectra predictions led to the correct MoA assignment for 100% of the antibiotics. Second, we showed how such a reference spectral dataset and autoencoder architecture could address a more difficult task: assessing the novelty of the MoA of a candidate antibiotic. After moving the single cell membrane inhibitor (i.e., colistin), one of the cell wall (i.e., cefazolin), and one of the protein (i.e., chloramphenicol) synthesis inhibitors to the test set, our tool successfully assigned colistin as a novel MoA, while cefazolin and chloramphenicol were rightly identified as cell wall and protein synthesis inhibitors, respectively.
Courbon et al. (Fri,) studied this question.