Combining Acoustic Bioprinting with AI-Assisted Raman Spectroscopy for High-Throughput Identification of Bacteria in Blood

Abstract

Identifying pathogens in complex samples such as blood, urine, and wastewater is critical to detect infection and inform optimal treatment. Surface-enhanced Raman spectroscopy (SERS) and machine learning (ML) can distinguish among multiple pathogen species, but processing complex fluid samples to sensitively and specifically detect pathogens remains an outstanding challenge. Here, we develop an acoustic bioprinter to digitize samples into millions of droplets, each containing just a few cells, which are identified with SERS and ML. We demonstrate rapid printing of 2 pL droplets from solutions containing S. epidermidis, E. coli, and blood; when they are mixed with gold nanorods (GNRs), SERS enhancements of up to 1500× are achieved.We then train a ML model and achieve ≥99% classification accuracy from cellularly pure samples and ≥87% accuracy from cellularly mixed samples. We also obtain ≥90% accuracy from droplets with pathogen:blood cell ratios <1. Our combined bioprinting and SERS platform could accelerate rapid, sensitive pathogen detection in clinical, environmental, and industrial settings.

Keywords: acoustic bioprinting; bacteria; gold nanorods; infectious disease; machine learning; surface-enhanced Raman spectroscopy.

Figure 1

(a) Schematic of acoustic printing platform and confocal Raman setup. Droplets containing bacteria (purple) and nanorods (gold) suspended in EDTA solution are acoustically printed onto a glass slide coated in 200 nm of gold (see also Supplementary Figures 2–4). (b) Stroboscopic images of the time evolution of upward droplet ejection at ∼3.5 m/s from an open pool at an acoustic frequency of 44.75 MHz and a droplet ejection repetition rate of 1 kHz. Images were captured with an exposure time of 40 ms, and as such, each frame is composed of 40 droplet ejections, highlighting ejection stability. The scale bar is 100 μm (see also Supplementary Figure 2). (c) Graph of droplet diameter versus ultrasound transducer resonant frequency. Droplets were printed with 4.8, 17, 44.75, and 147 MHz and had droplet diameters of 300, 84, 44, and 15 μm, respectively, highlighting the tunability of acoustic droplet ejection. (see also Supplementary Figure 1). (d) Raman spectra of dried cellular samples, including S. epi, E. coli, and red blood cells (RBCs) on a gold-coated slide.

Figure 2

Patterned droplet ejection from cellular stock solution. All droplets were ejected at 147 MHz. (a) Pattern printout of the Stanford University logo printed from droplets containing a 1:1 mixture of S. epi bacteria and mouse RBCs onto a gold-coated slide. The image on the left shows a photograph of print (top) with a scale bar of 4 mm. The brightfield image (bottom) was collected using a 5× objective lens and has a scale bar of 500 μm. (middle) SEM of the top portion of the tree region of the print with a scale bar of 100 μm. (right) A single row of 4 droplets from the large-area print, and then a magnified image of a single droplet with false coloring showing RBCs in red and S. epi bacteria in blue. The scale bar is 5 μm. (b) Droplets containing E. coli bacteria were printed onto an agar-coated slide and incubated at 37 °C for up to 36 h to demonstrate the cellular viability of printed samples. 100 droplets were placed at each location to ensure each droplet contained cells. The scale bar is 2 mm.

Figure 3

Spectral identification of cells printed with GNRs. (a) SEMs showing single droplets printed from varying cellular samples suspended in our EDTA solution at a concentration of 1e9 cells/mL. The left column shows samples without GNRs, and the right column shows cells printed with GNRs. From top to bottom, droplets contain S. epi, E. coli, and RBCs with false coloring added to highlight the cells. The scale bar is 5 μm. (b) Magnified SEM of a droplet containing S. epi coated with GNRs from (a). SEM highlights that the bacteria are coated with GNRs with very few rods dispersed in the rest of the droplet. The scale bar is 2 μm. (c) Mean SERS spectra of 100 measurements each taken from single droplets printed from three cell lines (S. epi, E. coli, and RBCs) mixed with GNRs. (d) 2-component t-SNE projection across all 300 Raman spectra acquired from droplets printed with GNRs. Data is plotted after performing a 24-component PCA for dimensionality reduction. Plots show distinct clustering of our cell lines. (e) Normalized confusion matrix generated using a random forest classifier on the 300 spectra collected from single cell-line droplets of S. epi, E. coli, and mouse RBCs mixed with GNRs. Samples were evaluated by performing a stratified K-fold cross-validation of our classifier’s performance across 10 splits, showing ≥99% classification accuracy across all samples.

Figure 4

(a) False-color SEMs of droplets printed from (left to right) an equal mixture of S. epi bacteria and RBCs, E. coli bacteria and RBCs, and S. epi, E. coli, and RBCs all diluted to 1e9 cells/mL in aqueous EDTA and mixed with GNRs. The scale bar is 5 μm. (b) 2-component t-SNE projection across all 600 Raman spectra acquired from 100 droplet measurements each, taken from single droplets printed from three cell lines (S. epi, E. coli, and RBCs) and three mixtures (S. epi and RBCs, E. coli and RBCs, and S. epi, E. coli, and RBCs) mixed with GNRs. Data are plotted after performing a 30-component PCA for dimensionality reduction. Plots show clustering of our cell lines with the most overlap between droplet mixture samples. (c) Normalized confusion matrix generated using a random forest classifier on the 600 spectra collected from single-cell-line droplets of S. epi, E. coli, and mouse RBCs mixed with GNRs and our 3 cell mixtures. Samples were evaluated by performing a stratified K-fold cross-validation of our classifier’s performance across 10 splits, showing ≥87% classification accuracy across all samples. (d) Heat map highlighting feature extraction performed to determine the relative weight of spectral wavenumbers in our random forest classification. The heat map is overlaid with a plot of the mean and standard deviation of the classification accuracy (black) calculated across all trials. Wavenumbers with lower accuracies are shown to be critical features, as random perturbations are highly correlated with decreases in classification accuracy. (e) Plots of the mean SERS spectra of 100 measurements each, taken from single droplets printed from three cell lines (S. epi, E. coli, and RBCs) and three mixtures (S. epi and RBCs, E. coli and RBCs, and S. epi, E. coli, and RBCs) mixed with GNRs. Wavenumbers attributed to biological peaks found in SERS spectra of S. epi, E. coli, and RBCs are plotted as blue, green, and red vertical lines, respectively. Peak assignments can be found in Supplementary Table 1.