Fig. 1. SERSome acquisition of bacteria lysate. (a) Schematic workflow including thermal lysis of bacteria, removal of debris, addition of Ag NPs and salt ions, injection of the mixture into a capillary tube, and detection using Raman spectroscopy. (b) Extinction spectrum and TEM image of Ag NPs. (c) Heatmaps of SERSomes and (d) the typical single SERS spectra from S. aureus at various concentrations. All spectra were normalized. (e) Spectral SNRs of S. aureus at different concentrations (n=20) calculated using the Raman band indicated by yellow in panel (d). P values were calculated using the Wilcoxon rank-sum test.

Fig. 2. Analysis of SERS spectra of eight bacteria. (a) Average SERS spectra of eight bacteria. The 200 to 300cm−1 region was scaled across bacterial species. The fingerprint region was scaled to [0,1]. The shaded area represents the standard deviation (n=1000). PCA plot of the spectra of the eight bacteria: (b) points are colored blue for Gram− and red for Gram+ bacteria with the confidence intervals indicating their distributions; (c) points are colored by eight types of bacteria instead with the confidence intervals indicating the distinctive types. (d) Average spectra in the range of 200 to 400cm−1. (e) Box plots for the intensities at 236cm−1 of the eight bacteria. For the data in panels (b), (c), and (e), each point indicates the mean value from an independent test (n=100). The color used for each bacterium in panels (c) and (d) can be referred to in panel (e). (f) Schematic illustration of the ligand replacement on the surface of Ag NPs after the addition of NaCl. Cl− ions occupy the surface of Ag NPs (left) and metabolites with strong competitiveness displace the Cl− ions (right). (g) Spectra (200 to 400cm−1) for two bacteria (ECO and SAU) and (h) two metabolites (CoA and adenine) at different concentrations.

Fig. 3. Bacteria identification using interpretable CNN. (a) Architecture of CNN for bacterial classification. SHAP is used for model interpretation. (b) Confusion matrix on the testing set. (c) SHAP values of eight bacteria species. For clarity, an average spectrum overlaid with SHAP values is also shown at the top.

Fig. 4. Metabolite identification by Ag NP-assisted LDI-MS. (a) Workflow of Ag NP-assisted LDI-MS. The Ag NP-adsorbed metabolites were separated and dropped on the LDI plate and measured by LDI-MS. (b) Spearman’s correlations among samples (n=3). (c) Different m/z features in Gram+ and Gram− bacteria. (d) PCA plot based on the LDI-MS features (n=12 for Gram+ and Gram−). (e) Distribution of metabolite superclasses and the total number of metabolites identified from HMDB (top) in each bacterium. (f) Boxplot of MS intensities for Gram+ and Gram− bacteria in different superclasses. All intensities are standardized according to the median. The correlation of the intensity at 236cm−1 with the (g) total MS intensity of each bacterium and (h) MS intensities of seven metabolite superclasses. (i) Detailed correlation with (I) organoheterocyclic compounds, (II) benzenoids, and (III) nucleosides, nucleotides, and analogs. The point color for each bacterium can be referred to panel (g). (j) Original mass spectra (174.010 to 174.022m/z) (left) and corresponding replicate intensities at 174.016m/z (right) in each bacterium.

Fig. 5. Metabolite-level interpretation of bacterial SERS spectra. (a) Schematic workflow for spectral decomposition. The metabolite panel was established based on the results of LDI-MS. Coefficients were derived by spectral decomposition based on NNLS. (b) Representative original bacterial spectra (black) and fitting spectra (blue) of the eight bacteria. (c) Box plots based on cosine similarity between the original spectrum and the fitting spectrum. Confusion matrices of (d) random forest and (e) support vector machine for bacterial identification based on coefficients from spectral decomposition.