Plasmon-enhanced stimulated Raman scattering microscopy with single-molecule detection sensitivity

Stimulated Raman scattering (SRS) microscopy allows for high-speed label-free chemical imaging of biomedical systems. The imaging sensitivity of SRS microscopy is limited to ~10 mM for endogenous biomolecules. Electronic pre-resonant SRS allows detection of sub-micromolar chromophores. However, label-free SRS detection of single biomolecules having extremely small Raman cross-sections (~10−30 cm2 sr−1) remains unreachable. Here, we demonstrate plasmon-enhanced stimulated Raman scattering (PESRS) microscopy with single-molecule detection sensitivity. Incorporating pico-Joule laser excitation, background subtraction, and a denoising algorithm, we obtain robust single-pixel SRS spectra exhibiting single-molecule events, verified by using two isotopologues of adenine and further confirmed by digital blinking and bleaching in the temporal domain. To demonstrate the capability of PESRS for biological applications, we utilize PESRS to map adenine released from bacteria due to starvation stress. PESRS microscopy holds the promise for ultrasensitive detection and rapid mapping of molecular events in chemical and biomedical systems.


Introduction
Raman spectroscopy is a versatile analytical tool providing information about the native fingerprint vibrational states of a sample determined by a molecule's structure and its environment.
Non-electronically resonant spontaneous vibrational Raman scattering cross-sections are typically 10 −30 cm 2 sr -1 and intrinsically small cross-sections on this order result in detection limits only as low as milli-molar (mM) levels. By placing a molecule close to a plasmonic nanostructure, plasmon-enhanced Raman spectroscopy pushes the detection sensitivity to the single-molecule level, [1][2][3][4][5][6][7][8][9] yet the speed in spectral acquisition is still not sufficient for ultrasensitive chemical mapping of molecular events in a dynamic and complex system. 10 Owing to the development of advanced lasers and electro-optic instruments, nonlinear Raman microscopy has been shown to provide label-free chemical imaging, based on either coherent anti-Stokes Raman scattering (CARS) or stimulated Raman scattering (SRS), for a broad range of biomedical applications. 11 Early developments of CARS and SRS microscopy relied on picosecond pulses for detection of a single Raman peak. 12,13 Intra-pulse broadband CARS, developed by Cicerone and coworkers, allowed recording of a whole Raman spectrum within 3.5 ms. 14 Multiplex SRS microscopy developed by Cheng and coworkers 15 , is able to acquire a Raman spectrum covering a 200 wavenumber spectral window within 5 µs, which allowed highthroughput chemical analysis in a flow cytometry setting. 16 Yet, the imaging sensitivity of SRS microscopy is limited to ~10 mM for chemical bonds such as the C-H vibrations in cell membranes. 17,18 . Min and coworkers recently reported electronic pre-resonance SRS achieving sub-M-sensitivity detection for chromophores having a Raman cross-section over 10 3 or 10 4 times larger than endogenous biomolecules. 19,20 To push coherent Raman detection sensitivity further, plasmon-enhanced CARS has been reported [21][22][23][24][25][26] and single-molecule sensitivity has been proved. 23,24 While, the CARS signal carries a non-resonance background which complicates quantification and distorts the CARS spectrum.
Additionally, the CARS signal displays a nonlinear dependence on the concentration of analytes. 27 The SRS signal, on the other hand, exhibits identical spectral profile as spontaneous Raman and a linear dependence on the concentration of analytes. 13 The Van Duyne group reported reproducible surface-enhanced femtosecond SRS spectra from molecules embedded in a gold nano-dumbbell sol. 28, 29 Yet, plasmon-enhanced SRS at single-molecule detection sensitivity has not been reported.
Major hurtles of achieving single-molecule SRS detection include the damage of plasmonic substrates by the ultrafast pulses 30 and a large pump-probe background, arising from plasmoninduced photothermal and/or stimulated emission process.
Here, we report plasmon-enhanced SRS (PESRS) microscopy (Fig 1a, instrument in SI1) and its application to ultrasensitive imaging of biomolecules released from a cell. We reached singlemolecule detection sensitivity by incorporating several innovations. First, we used chirped laser pulses at 80 MHz repetition rate for spectral-focusing hyperspectral SRS imaging. The pulse energy on the sample was on the level of pJ. Such low pulse energy together with chirping to picosecond duration effectively avoided sample photodamage, while the high repetition rate allowed fast chemical mapping of molecules adsorbed on gold nanostructured surfaces. Second, we employed a penalized least squares (PLS) approach and successfully extracted the sharp Raman peaks from a spectrally broad non-Raman background largely contributed by the photothermal effect. 31 Third, harnessing a block-matching and 4D filtering (BM4D) algorithm to denoise a hyperspectral stack, we were able to generate high-quality single-pixel SRS spectra for statistical analysis of single-molecule events. By a bianalyte method, [32][33][34][35] we used two isotopologues of adenine that offer unique vibrational signatures and verified PESRS detection of single molecules with Raman cross-section as low as 10 -30 cm 2 per molecule. Furthermore, we demonstrated PESRS imaging of adenine resulting from nucleotide degradation as a stress response of S. aureus cells to starvation.

Plasmon-enhanced stimulated Raman scattering (PESRS) spectroscopy. Adenine adsorbed on
Au NPs aggregation substrates (see Methods) was selected as a proof-of-principle system for the demonstration of PESRS. Adenine is one of the four constituent bases of nucleic acids. The Raman band at 723 cm -1 of adenine powder, which has a cross-section of 2.9×10 -30 cm 2 , 36 has been studied for a single-molecule detection by surface-enhanced Raman spectroscopy (SERS) 35,36 and surfaceenhanced CARS. 23 As shown in Fig. 1a, a pump laser centered at 969 nm and a Stokes laser centered at 1040 nm were employed to induce a PESRS spectrum covering a window ranging from 550 to 850 cm -1 . 10 µL of a 5 mM aqueous adenine solution was added to 2~4 µL of a concentrated Au colloid suspension which induced the aggregation of Au NPs. A representative extinction spectrum of an adenine-induced Au NPs aggregation substrate is shown in Fig. 1b. The plasmonic band of the aggregated Au NPs is broad and peaked at 1040 nm, which allows PESRS for the pump and Stokes laser wavelength used here. The resulting PESRS spectrum (Fig. 1c, black) from the adenine-adsorbed Au NPs aggregates consists of a narrower feature at 733 cm -1 (highlighted by green) on top of a strong and broad non-resonant background. This sharp feature is close to the prominent adenine ring-breathing mode frequency observed in the normal SRS spectrum of adenine powder (Fig. 1c, blue) and identical to the corresponding 733 cm -1 peak observed in the SERS results on Au substrates (SI2). 37 The blank result (Fig. 1c, red) was independently measured from the Au NPs substrate without adenine adsorption. The background could arise from three different non-Raman processes: a photothermal effect, cross-phase modulation, and transient absorption, 31 all due to laser interactions with the Au nanostructures. The spectral shift between the substrate with/without adenine may related to the different extent of aggregation with/without adenine. These backgrounds are spectrally overlapped with the SRS signal, but are largely independent of the Raman shift. 31 In contrast, the SRS signal originates from a vibrational resonance that has a sharp spectral feature. A PLS approach was used to fit the broad spectral background. The resulting fitting backgrounds of PESRS are shown in Fig. 1c as the same color dash lines for the corresponding observed PESRS spectra. Fig. 1d shows the vibrationally resonant component of the PESRS spectra resulting from subtraction of the fitting backgrounds from the observed PESRS signals. The PESRS spectrum of adsorbed adenine shows a dominated peak at 733 cm -1 . Only a noisy baseline is evident after background subtraction from the pure substrate spectrum. Compared with the SRS spectrum of adenine crystal (blue line in Fig. 1d), a 10 cm -1 blue shift of the peak is observed in the PESRS spectrum. This blue shifted frequency (733 cm -1 ) is consistent with the strongest vibrational feature observed in SERS spectra of adenine (SI2). 38 These results collectively indicate that the observed vibrational PESRS signal component originates from the surface adsorbed adenine. To verify that the SRS signal is due to the adenine vibrational resonance, we varied the pump wavelength while keeping the Stokes wavelength fixed. The pump laser centered at 972 nm as well as the previous 969 nm wavelength encompass the adenine Raman resonance for a 1040 nm Stokes pulse, and both generated SRS spectra showing a pronounced peak at the expected wavenumber (Fig. 1e, black and red). In contrast, the 942 nm is off-resonance for the 733 cm -1 band. Accordingly, the measured spectrum does not exhibit such a peak as shown in Fig. 1e (blue).
After subtraction of the background in Fig. 1e (corresponding fitted backgrounds were shown as same color dash lines), the PESRS spectra of adenine excited by both Raman resonance wavelengths show a Raman peak at 733 cm -1 (black and red, Fig. 1f), whereas the off-resonance spectrum only shows a noisy featureless baseline (blue, Fig. 1f). Moreover, as shown in SI3, the intensity of the 733 cm -1 peak linearly depends on the pump power and the Stokes power before it reaches saturation. The reproducible spectra recorded at the same location demonstrate that the laser power in our experiment did not damage the substrate or induce molecular photodegradation.  The single-pixel spectrum from spot 1 shows a broad background and a weak Raman peak around 733 cm -1 . After pixel by pixel subtraction of the fitting background, the area of the resulting vibrational band at 733 cm -1 at each pixel is shown in Fig. 2c revealing a clear spatial contrast between regions of adsorbed adenine and blank areas. The single-pixel background-removed spectra from spot 1 and 2 are displayed in Fig. 2d. It remains challenging to obtain high-quality single-pixel spectra due to the noisy non-Raman background. To address this challenge, we employed a BM4D algorithm which was widely used for 3D data denoising. 39,40 The reconstructed peak area image and the single-pixel spectra after BM4D denoising and background removal are shown in Figs. 2e and 2f, respectively. By employing BM4D, we achieved a signal-to-noise ratio of 33 for the single-pixel spectra at spot 1, ~4 times better than that without denoising. Good Raman reproducibility in terms of peak frequency in different locations of the imaging area was found (see SI5) even given an expected inhomogeneous intensity distribution due to the Au NPs randomly aggregated hot spots. PESRS was also demonstrated for epi-detection of molecules on a non-transparent plasmonic substrate, since such substrates are often used in plasmon-enhanced spectroscopy applications. The experimental setup is shown in Fig. 3a. We used a sol-gel-derived SiO2 substrate covered by immobilized aggregates of monodispersed-sized Au NPs (AuNPs-SiO2 substrate). 41 10 L of a 100 M adenine solution were dropped on this plasmonic substrate and dried the sample in air.
The spectrally integrated image (Fig. 3b) reveals the distribution of NP clusters on the SiO2 chip.
After BM4D denoising and background subtraction, the distribution of hot spots is evident in Fig.   3c. Single-pixel spectra extracted from spots 1 and 2 are indicated in Fig. 3d. After denoising, a signal-to-noise ratio of 48 is achieved for these single-pixel spectra. These data collectively show the high sensitivity of epi-detected PESRS.
To estimate the relative enhancement factor of a local hot spot, we assumed a monolayer surface coverage of adenine and a monolayer NP cluster under the laser focus. Based on the measured local PESRS intensity (spot 1) and the average SRS intensity of 5 mM adenine solution, the power-and concentration-averaged local enhancement factor of PESRS relative to normal SRS is estimated to be ~ 710 7 (see details in SI6). Consistent with this result, enhancement factors of 10 4 ~10 6 and 10 5 ~10 8 were reported for surface-enhanced femtosecond SRS 28 and surfaceenhanced CARS spectroscopy, 22, 23, 25 respectively.  (Fig. 4a). 35,36 The PESRS spectrum of an equimolar solution of 14  To evaluate the sensitivity of PESRS, we prepared a mixture solution of 14 NA and 15 NA at 500 nM concentration each with Au NPs and dried the colloid on a cover glass under vacuum. A hyperspectral PESRS image of this mixture sample, consisting of 40000 spectra, was acquired for statistical analysis. Fig. 4b shows typical single-pixel spectra after denoising and background subtraction. Spectrum 1 has a single peak at 726 cm -1 , matching the spectrum taken from the reference sample of isotopically pure 15 NA. Spectrum 3 has a single band at 733 cm -1 , corresponding to the PESRS spectrum of 14 NA. Spectrum 2 has a single peak at 730 cm -1 that corresponds to the spectra of mixed molecules. Two distinct features could not be observed in the PESRS spectra of these adenine isotopologues due to the spectral resolution of the hyperspectral SRS system (c.a. 14 cm -1 ). Notably, however, spectra with single peaks at 726 and 733 cm -1 were observed at multiple pixels (see SI7). These data allowed the statistical analysis described below.
A multivariate curve resolution (MCR) method was used for the statistical analysis of the PESRS spectra. The hyperspectral data (containing 40000 single-pixel spectra) was unmixed by MCR into concentration contributions of pure 14 NA (C14) and pure 15 NA (C15) spectra (Fig. 4c) and a relative concentration ratio of 14 NA (C14/(C14+C15)), defined as the fraction of the average number of 14 NA molecules contributing to the total signal, was thus determined. We selected 4172 of the total spectra acquired that displayed the desired Raman bands and had an intensity above a threshold value (maximum values > 0.01) were selected for this statistical analysis. The threshold requirement helped reduce inclusion of noise events and avoid the artificial counting of molecular events. 33 The histogram of relative contributions to the total signal produced by 14   PESRS mapping of adenine generated from bacteria. The investigation of dynamic living samples requires imaging at a high speed. Compared to SERS, the dramatically improved speed of PESRS microscopy makes it a potentially useful tool for imaging the chemical dynamics of a complex living system. To demonstrate such capacity, we studied the metabolic response of S. aureus to starvation, as shown in Fig. 5a. Following enrichment in a nutrient-rich environment, the S. aureus sample was washed and centrifuged in pure water. After 1 hour, 1 L of a bacterial suspension was placed on the Au NPs-SiO2 plasmonic substrate and once the water evaporated (~ 5 min) the PESRS signal was acquired. A control sample was similarly prepared but in contrast a PESRS spectrum was obtained without a 1-hour delay. PESRS spectra of S. aureus under starvation conditions for 1 hour are displayed in Fig. 5b top, and the observed spectra closely resemble the Raman spectrum of adenine. In contrast, the spectra of S. aureus obtained immediately (no waiting period) (Fig. 5b bottom) do not exhibit an adenine-like Raman band.
These results are consistent with the SERS data (SI9). 37 These data imply that adenine, a purine degradation product, is secreted from S. aureus as a response to the no-nutrient, water-only environment. 37 As shown previously, 37 these molecular species are secreted from the bacterial cells under starvation conditions and appear most heavily concentrated in the pericellular region near the outer cell wall. Fig. 5c shows the PESRS image of starved S. aureus on the plasmonic substrate, which presents the distribution of the secreted adenine. The two representative single-pixel PESRS spectra of S. aureus are present in Fig. 5d. These results collectively demonstrate that PESRS has the potential for the study of the bacterial exogenous metabolome.

Discussion
Through plasmonic enhancement and hyperspectral recording, the detection sensitivity of SRS microscopy reached the single-molecule level. The Au plasmonic nanostructures provided an extraordinary SRS intensity enhancement relative to normal SRS of about 10 7 . Such large enhancements allow the detection of single molecules with a Raman cross-section as low as 10 -30 cm 2 . Single-molecule PESRS detection of adenine molecules was verified by a bianalyte method.
A potential biomedical application of PESRS was demonstrated through mapping of adenine released from stressed bacterial cells.
In this work, single-molecule detection sensitivity in PESRS was achieved through a combination of several strategies. First, we created and employed a nanostructured Au substrate with a plasmon resonance peak overlapping with our pump and Stokes laser wavelengths. Second, we used chirped pJ laser pulses at 80 MHz repetition rate. Such pulse energy, which was 2 to 3 orders of magnitude lower than previous surface-enhanced femtosecond SRS work, 28,29 significantly decreased potential photodamage. 29 The high repetition rate also allowed high-speed data acquisition. Third, a PLS method was used to distinguish the Raman peak from the broad non-Raman background in the hyperspectral dataset. Finally, a BM4D approach denoised a hyperspectral cube and allowed high-quality single-pixel spectra to be obtained for statistical analysis of single molecules events.
A portion of our recorded PESRS spectra show a dispersive vibrational line shape (Fig. 1f).
Such line shapes were reported in previous surface-enhanced femtosecond SRS works. 28,29,42 Three possible explanations have been proposed: Fano-resonance of the molecule-nanostructure system, 42 interference between the PESRS signal and the plasmon enhanced aggregated NP emission, 43 and the effects of the complex character of the plamonic field amplitude. 44 The distribution and character of this PESRS line shape character will be described in subsequent work.
PESRS microscopy opens a new window for fast vibrational spectroscopic imaging of lowconcentration molecules with high sensitivity. PESRS microscopy can sensitively detect metabolites secreted from a live cell. Thus, PESRS may be used to distinguish bacteria types and investigate metabolic changes linked to the development of microbial populations or to the exposure to antibiotics. In addition, this method can be extended to study the dynamic processes in surface chemistry, such as the mapping of the solid electrolyte interface membrane in a lithium cell or imaging the heterogeneity of catalyst.
Additional improvements may be envisioned for this technology. For example, the imaging speed can be further improved by using a multiplex SRS method, 15,16 advanced delay tuning approach, 18 or a wide-field SRS system. Secondly, harnessing the rational-designed reproducible plasmonic nanostructure fabricated by lithographic methods, our method can pave the way for reproducible and quantitative molecular imaging platform. Such improvements will invoke the integration of coherent Raman imaging techniques and novel nanostructure designs to open new avenues towards ultra-sensitivity, ultra-fast, label-free chemical imaging.

Methods
Hyperspectral plasmon-enhanced stimulated Raman scattering microscope. Figure S1 presents the scheme of a hyperspectral SRS microscope. Briefly, an 80 MHz tunable femtosecond laser (InSight DS+, Spectra-Physics) provided the pump (680-1300 nm) and Stokes (1040 nm) simulating fields. The Stokes beam was modulated by an acousto-optic modulator at 2.3 MHz. The pump and Stokes beams were spatially aligned and sent to an upright microscope with 2D galvo system for laser scanning. The spectral-focusing approach was used to obtained spectral domain information. In spectral focusing, the pump and Stokes pulses were equally stretched in time by glass rods to achieve a constant instantaneous frequency difference that drives a single Raman coherence. By delaying the pump pulses, a series of Raman shifts (80 channels) were generated.
At a certain delay, all the laser energy was spectrally focused to excite a narrow Raman band. The An epi-detected SRS microscope was built for PESRS detection on non-transparent plasmonenhanced substrates. Before the microscope, a quarter wave plate was placed after a polarizing beam splitter to change the polarization of excitation and back-reflected laser light by 90. In this way, the polarizing beam splitter allowed forward light to pass through and the stimulated Raman loss signals were reflected into a photodiode to achieve epi-PESRS imaging (Fig. 3a).

Background reduction in PESRS.
A raw PESRS spectrum contains a large background signal from the photothermal effect, cross-phase modulation, and transient absorption. Cross-phase modulation originates from the optical Kerr effect, and the transient absorption and photothermal effect are due to the plasmonic resonance of the Au NPs. We minimized the background arising from non-Raman processes by using a larger numerical aperture (NA=1.4) lens for signal collection to reduced cross-phase modulation and photothermal effect. Moreover, a MHzfrequency modulation was used to further diminish the photothermal effect. With those approaches, we successfully observed a PESRS signal in the presence of a strong background.
Background subtraction in a PESRS spectrum. An adaptive iteratively reweighted penalized least squares (airPLS) algorithm, developed by Zhang et al. 45 , was employed to subtract the baseline from the raw PESRS spectrum.
Denoising of a PESRS hyperspectral data. Firstly, we used a BM4D denoising algorithm, developed by Maggioni and Foi,39,40 to process the raw PESRS hyperspectral data cube. The BM4D algorithm relies on the so-called grouping and collaborative filtering paradigm. A 3D imaging block (x-y-) was stacked into a 4D data array, which was then filtered. Thus, BM4D leveraged the spatial and spectral correlation of a hyperspectral data cube both at the nonlocal and local level. Then, we used the airPLS algorithm to subtract the background of a hyperspectral data cube pixel by pixel. In addition, the image was plotted by the peak area of the desired Raman band.

Statistical analysis of single-molecule events.
Before the statistical analysis of single-molecule events, the PESRS hyperspectral data was denoised and corrected baseline as described. A multivariate curve resolution (MCR) algorithm, developed by Tauler and Juan et al., 46,47 was used to extract the concentration maps of the two isotopically related molecules in the mixed hyperspectral dataset. Because 14 NA and 15 NA have the same Raman cross-section, we used the normalized spectra separately obtained from pure 14 NA and 15 NA samples as the initial estimation of the pure spectra. The constraints implemented during the optimization step were non-negativity for the concentration and spectrum. The outputs of the MCR treatment were the pure concentration maps (C14, C15) of 14 NA and 15 NA, respectively. Then, the 4172 spectra whose maximum values appeared at the desired wavenumber range were selected and an intensity threshold (maximum values > 0.01) was defined for removal of noise events. The histogram of the relative contribution of 14 NA (C14/(C14+C15)) were plotted and the edges of the histogram, C14/(C14+C15) ≈ 0 or ≈ 1, were considered as the single molecules events. All data were processed by MATLAB.
Substrate preparation. The Au NPs colloid was prepared according to the classical citrate reduction method, 48 and resulted in particles with a diameter of ~ 50 to 60 nm, as shown in SI10.
The 0.5 mL of 0.01% (g/ml) colloidal suspension was concentrated to 2-4 µL by centrifuging, which was then added to the 10 L adenine solution. The adenine solution induced the aggregation of Au NPs. The aggregated Au NPs were dropped on a cover glass, followed by vacuum drying to obtain a substrate for PESRS detection.
Au NPs-SiO2 plasmonic substrates were prepared as described previously. 41 Immobilized clustered aggregates of 80 nm Au NPs were grown on a SiO2 chip. A 10 L adenine solution was dropped on the Au NPs-SiO2 plasmonic substrates and samples was ready for PESRS measurement after adenine solution dried under air ambient (~ 5 mins). High-purity water (Milli-Q, 18.2 Mcm) was used throughout the study.
Bacteria sample preparation: Bacteria were harvested during the log phase. Culture growth media was removed from the bacterial samples by centrifugation, and washing four times with 2 mL of deionized Millipore water. The bacterial pellet was suspended in 0.25 mL of water, and 1 μL of the resulting bacterial suspension was dropped and dried onto the Au NPs-SiO2 plasmonic substrate. Samples were dried onto the Au substrate either immediately or after 1 hour in order to demonstrate the effect of the starvation stress response.
Additional experimental results and data analysis are available in the supplementary information.
Data availability. The data that support the plots within this paper and other findings of this study are available from the corresponding author upon reasonable request.

Plasmon-enhanced Stimulated Raman Scattering Microscopy with
Single-molecules Detection Sensitivity Cheng Figure S1. The scheme of a hyperspectral plasmon-enhanced stimulated Raman scattering microscope. Laser system: 80 MHz tunable femtosecond laser. AOM: acousto-optic modulator. GM: 2D galvo mirror. PBS: polarizing beam splitter; QWP: quarter wave plate; PD: photodiode. For spectral focusing, three rods were used in combined path and one rod was used in Stokes path. In this way, the pump and Stokes pulses were chirped to achieve a constant instantaneous frequency difference that drives a single Raman coherence. A series of Raman shifts were generated by scanning the delay stage.  Figure S2. The SERS spectrum (black) of adenine adsorbed on Au NPs aggregation substrate (5 mM in solution) has a peak at 733 cm -1 . The Raman spectrum (red) of 5 mM adenine solution has a peak at 723 cm -1 . The SERS spectrum was recorded with 5 s integration time, with a 50× objective and a 0.5 mW laser power at 785 nm. The Raman spectrum was recorded with 30 s integration time, with a 40× objective and an 80 mW laser power at 532 nm. This 10 cm -1 blue shift is due to the formation of metal-adenine complex. 1 .

SI8. The simulated single-molecule PESRS data
Here, we introduced a model example to describe the statistics of single-molecule PESRS signal in a hot spot. This model was based on the previous single-molecule SERS mode developed by Eric Le Ru, PG Etchegoin, et at. 3 First, we used the boundary element method approach 4-6 to calculate a local electric field distribution on a representative hot spot. Fig S8a presented the simulated local electric field distribution of the representative hot spot (a dimer formed by two 60 Au NPs with a 1 nm gap). Then, we assumed that we had a certain number of molecules of two isotopic adenines (N14 and N15) on the hot spot. We generated random locations of molecules in the hot spot and every molecule felt a corresponding local electric field in the hot spot. 7 Then, we calculated the total intensity produced by each type of molecules (I14 and I15) by summing over the corresponding intensity of every molecules. Because the Raman cross section of 14 NA and 15 NA were the same, the ratio of I14/(I14+I15) was also the ratio of the average number of 14 NA contributing to the signal. We repeated this process for many times (as a large number of events) and obtained the histogram for relative contribution of 14 NA. Figure S8b Figure S9. SERS spectra of starved S. aureus (black) and non-starved S. aureus (red). After 1 hour starvation in pure water, the SERS spectrum of S. aureus closely resembled the SERS of adenine. This result indicates that the adenine appeared at the outer layer of S. aureus and in the extracellular metabolome as resulting from the bacterial cell stress response to the no-nutrient, water-only environment. The spectra were recorded with 1.0 s integration time with a 50× objective and a 1.0 mW laser power at 785 nm.