Decoding Spectral Signatures of Bacterial Metabolism

Even in the 21^st century, bacterial infections are treated empirically. Antibiotics are prescribed preemptively, because a proper diagnosis relies on culture growth which can delay patient treatment, increasing morbidity. Yet the unnecessary administration of powerful, broad-spectrum antibiotics leads to the proliferation of antibiotic resistance, which currently causes over 35k deaths annually in the US alone. In order to reduce the time required for antibiotic susceptibility tests (AST) needed for informed diagnosis, we use spectral data from plasmonic architectures for metabolomic fingerprinting in order to determine phenotypic susceptibility or resistance to antibiotics. Bacterial responses to stress are critical to their survival in dynamic and challenging environments. Stresses can be general, such as from nutrient deprivation, or specific, such as from targeted antibiotic activity. Regardless of source, bacterial stress response results in rapid and profound shifts in metabolite concentrations within cells to maintain homeostasis. This platform complements existing genomic tests that only provide feedback on known antimicrobial resistance mechanisms.<br/>Yet metabolomics approaches introduce an enormous parameter space, e.g., the <i>E. coli</i> metabolome contains over 2600 different metabolites. We overcome this challenge with plasmonic nanogaps with tunable chemistry providing selective interactions with classes of metabolites to rapidly collect large surface enhanced Raman scattering (SERS) data sets on the scale of minutes for analysis with machine learning (ML) algorithms. Just as one can smell the difference between coffee and chocolate amongst multiple odors, SERS+ML rapidly measures and classifies spectral features of bacterial metabolite signatures in response to antibiotics, which are correlated with antibiotic lethality mechanisms. In order to overcome the longstanding challenge to produce large, high quality and reproducible data sets that are needed for ML analysis, electrohydrodynamic flow is used as a driving force to drive chemical reactions between gold nanospheres in a microfluidic cell. The nanogap distance is controllable on the length scale of angstroms, which is critical as the nanogap distance determines performance in SERS devices. It is low-cost, scalable, and reproducible chemical assembly method able to detect individual molecules over mm² areas. Calibration data evaluated in an implemented convolutional neural network (CNN) regression model trained on SERS data of Rhodamine 800 resulted in limits of detection (LOD) and quantification (LOQ) of 10 fM (~10^-5 ng/mL) with prediction accuracy (r²value) of 0.96 over a dynamic range of 6 orders of magnitude.<br/>Analysis of SERS data using a generative model, the variational autoencoder (VAE), is able to identify metabolic fingerprints associated with antibiotic efficacy in the in the cell lysate data. The high interpretability of the VAE generated spectra allows us to identify useful vibrational information which guides additional targeted data collection. Culture-free and easily acquired datasets of bacterial metabolites augment training data to improve predictive models. Greater than 99% accuracy is achieved with unsupervised Bayesian Gaussian Mixture analysis using just a single spectrum from each class representing different antibiotic exposure conditions when using this data informed transfer learning. Our results show differentiation of bacterial populations of ESKAPE pathogens based on antibiotic susceptibility in 10 min when using SERS + ML. This enormously reduces the amount of time needed to validate phenotypic AST with conventional cell growth assays and outlines a promising approach towards rapid phenotypic AST.