Surface enhanced Raman scattering artificial nose for high dimensionality fingerprinting

Label-free surface-enhanced Raman spectroscopy (SERS) can interrogate systems by directly fingerprinting their components’ unique physicochemical properties. In complex biological systems however, this can yield highly overlapping spectra that hinder sample identification. Here, we present an artificial-nose inspired SERS fingerprinting approach where spectral data is obtained as a function of sensor surface chemical functionality. Supported by molecular dynamics modeling, we show that mildly selective self-assembled monolayers can influence the strength and configuration in which analytes interact with plasmonic surfaces, diversifying the resulting SERS fingerprints. Since each sensor generates a modulated signature, the implicit value of increasing the dimensionality of datasets is shown using cell lysates for all possible combinations of up to 9 fingerprints. Reliable improvements in mean discriminatory accuracy towards 100% are achieved with each additional surface functionality. This arrayed label-free platform illustrates the wide-ranging potential of high-dimensionality artificial-nose based sensing systems for more reliable assessment of complex biological matrices.


Supplementary Discussion 1. Fabrication of SERS-active FASERS substrates
The production of the FASERS Au-nanopillars substrates was achieved via colloidal lithography and plasma etching adapted from the reported protocols for silicon nanopillars 1, 2, 3 . Supplementary Figure 1a shows the overall fabrication process comprising four steps: spin coating of polystyrene beads (PS), reactive ion etching (RIE) of the Si3N4 coated silicon (Si) wafer substrates for the pillar formation, thermal evaporation of a Cr/Au film, and subsequent SAM functionalization. The change in the surface morphology of the substrates following each step was evaluated with scanning electron microscopy (SEM) images (Supplementary Figure 1b). 300 nm PS beads were used to generate a hexagonal array of beads to act as a mask for RIE. The RIE step enabled etching of the Si3N4 layer (~120 nm), which yielded PS capped Si3N4 nanopillars (PS-Si3N4). It was observed that the RIE step resulted in a roughness change on the surface of the PS beads and their subtle size reduction (~30 nm); both phenomena consistent with reported literature 3 . The subsequent Cr (10 nm) and Au deposition (70 nm) on the PS bead capped Si3N4 nanopillars yielded the Au-nanopillars exhibiting gaps of several nanometers between each adjacent dome. It was observed in the SEM images that the Au was successfully deposited on the sidewalls of pillars. (f-j) SERS mapping (50 µm x 50 µm, pixel size of 1 µm 2 , peak at 1074.7 cm -1 ) of corresponding location: (f) Au-Si, (g) Au nP Site 1, (h) Au nP Site 2, (i) Au nP Site 3, (j) Au nP Site 4. Scale bar represents 10 µm. Each spectrum was smoothed, baseline subtracted and normalized by the area under the curve. The maximum value of the intensity bar corresponds to the maximum intensity out of all the measurements performed on all of the different sites. (k) Normalized SERS intensity of peak at each pixel (50 x 50) of the scan. Significant SERS enhancement of 4-MBA signals across the entire scanned area were observed compared to those of Flat Au surface (Au-Si). Importantly, we have observed no/little apparent mean signal variation between the different spatial location, indicating a good spatial reproducibility of the enhancement.

Tentative peak assignment of SERS spectra from self-assembled monolayers
The C-C stretching region (1000-1500 cm -1 ) of the monolayers contains information about the conformational behavior of the alkyl chain, arising from the gauche and trans conformers of C-C bonds.
The prominent peaks around 1100-1130 cm -1 correspond to (C-C) trans-conformers, the positions of which agree well with a number of previous studies 8,9,10,11,12 . Regarding the relatively strong intensity of (C-C) stretching of 1-propanethiol and 1-undecanethiol compared to other short thiolates, this is attributed to the 1-alkanethiol molecules having the lowest symmetry amongst the Cs point group, therefore the vibrations of the adsorbed molecule exhibit a component perpendicular to the surface in all orientations 9,11 . As the vibration components along the z-axis (i.e. zz tensor component) are enhanced to a larger extent than those with xz or yz components, the C-C components of 1alkanethiol show a large contribution of stretching vibration of trans-conformers 9,11 . On the other hand, the short-chain alkanethiol with a substituted terminal group can also interact with the metal surface 10,13,14,15,16,17 . In particular, the terminal carboxylic and amino group are known to exhibit a relatively higher concentration of gauche conformers via double bonding to the metal surface 10,13,14,15 , which leads to minimized stretching vibration of trans-conformers when compared to 1-propanethiol.

XPS analysis
In all 8 SAM-functionalized surfaces, clear S 2p peaks were identified and assigned to boundthiolate (Au-S) which is reported to be located in a ~1.5 eV lower binding energy region than that of unbound thiol (S-H) 18 . It was observed that the short thiolates were bound with no detectable contaminants, while most of the long thiolates showed a signal correlating to small fraction of nonbound thiolates, which is commonly detected for ethanolic immersion methods 4,18 . In particular, the C 1s peak of the alkyl-terminating SAM (Supplementary Figure 4) was composed of C-C bonding while the hydroxyl-terminating SAM exhibited an additional peak arising from C-O bonds (Figure 2h). The carboxyl end-group was identified in emerging peaks related to C-O, C=O, C-OH, and O-C=O bonds, while the peak corresponding to the amine end-group was fitted to C-C and C-N bonds with C-O possibly from surface-bound water 19 . The bare Au surface C 1s peak showed a small amount of non-volatile hydrocarbon contaminants (C-C bond), which are commonly detected due to the high surface energy of gold 20, 21 .

Supplementary Figure 4. High-resolution XPS spectra of C 1s, S 2p and Au 4f for non-functionalized and SAM-functionalized Au-Si.
Each SAM is referred to as: non-functionalized (bare), 1-undecanethiol (11CH3), 11-mercapto-1-undecanol (11OH), 11-mercaptoundecanoic acid (11COOH), 11-amino-1-undecanethiol (11NH2). * Refers to the weak peaks that are expected from the chemical structure of the molecules used. The peak shifting in the 3-carbon SAMs is likely attributed to attenuation of X-ray intensity due to differing degrees of overlayering on top of the SAM.

Supplementary Discussion 3. Tentative peak assignment of SERS spectra from model analytes
The prominent SERS bands of p-PDA were located at 1169 cm -1 , 1496 cm -1 , 1601 cm -1 with a shoulder at 1627 cm -1 , which are consistent with the previously reported SERS spectrum of p-PDA on gold nanoparticles 22 . Each band can be assigned to C-N stretching/in-plane NH2 wagging, in-plane C-N stretching, and in-plane NH2 scissoring, respectively. The spectra of 4-APA have two prominent bands at 1175 cm -1 and 1601 cm -1 , which can be tentatively assigned to C-N stretching 22 and the ring breathing mode of the aromatic ring, 23 respectively. Figure 3d and 3e showed resultant signatures of R6G and FA from FASERS. The peak positions of R6G are in accordance with previously reported spectra, obtained both experimentally 24,25,26 and computationally 27,28 . Several prominent bands include 1186 cm -1 , 1312 cm -1 , 1602 cm -1 , and 1646 cm -1 , which were tentatively assigned to in-plane ring deformation, xanthene ring breathing, asymmetric C=C stretching mode and xanthene ring stretching, respectively. The characteristic bands of FA were also in good agreement with previously reported SERS spectra of FA adsorbed on silver and gold substrates 29,30 . The SERS bands correspond to NH2 scissoring/C-O stretching (1625 cm -1 ), aromatic ring breathing (1595 cm -1 ), inplane N-H vibration (1563 cm -1 ), CH2 wagging/in-plane C-H deformation (1321 cm -1 ), and CH2 aliphatic twist (1181 cm -1 ), respectively 31 . Further details regarding peak assignments for these molecules are summarized in Supplementary Table 2-3.

Supplementary Discussion 4. Molecular dynamics simulation: computational details
All-atom molecular dynamics (MD) simulations were performed to structurally characterize the eight SAM-functionalized surfaces of differing composition and explore the mechanisms of binding of the two small molecules, p-PDA and 4-APA, with the SAMs and bare gold. Simulations consisted of two stages: (i) monolayers were thermally equilibrated to characterize their interfacial structure and properties, (ii) the equilibrium SAM structures were then used to explore interactions with individual analyte molecules.
Each system contained a close-packed face-centered cubic Au(111) slab of periodic unit cell dimensions 4.1 × 4.4 nm 2 in the lateral (X-Y) directions and 8 layers (1.7 nm) of fixed gold atoms in the Z direction (surface normal). On each exposed gold slab surface, 80 monolayer molecules were initially placed upright with their sulfur head groups arranged in a hexagonal (√3 ×√3)R30° structure 32 relative to the underlying Au(111) lattice (Supplementary Figure 5). This produced a packing density of 22.33 Å 2 / chain, which corresponds to a fully saturated coverage and maximum packing density 32 . A 6 nm spacer was introduced between the two SAMs in each simulation cell (Supplementary Figure  5a-b) ensuring that the two surfaces (top/bottom) were sufficiently apart to prevent self-interactions and allowing for property statistics to be gathered from two surfaces per simulation while minimizing computational expense. The SAMs modelled are commensurate with the 8 different SAMfunctionalized Au-nanopillar substrates, which differ in terminal group chemistry and chain length, including: 1-propanethiol (3CH3), 3-mercapto-1-propanol (3OH), 3-mercaptopropionic acid (3COOH), 3-amino-1-propanethiol (3NH2), 1-undecanethiol (11CH3), 11-mercapto-1-undecanol (11OH), 11mercaptoundecanoic acid (11COOH), and 11-amino-1-undecanethiol (11NH2). The analytes were modelled in the same buffer conditions as per experiment, where p-PDA (pH 5) had one of the amine groups protonated and 4-APA (pH 7.5) had the carboxyl group deprotonated. The different pH conditions for each analyte were also reflected in the SAM composition. Carboxyl terminated SAMs 3COOH and 11COOH have surface pKa values of 5.2 33 and 5.0 34 respectively, thus SAMs composed of these ligands were modelled with two different protonation states: all carboxyl groups deprotonated (COO -, pH 7.5); and a 1:1 alternating mixed state of neutral (COOH) and deprotonated (COO -) carboxyl groups (pH 5, Supplementary Figure 5b). Amine terminated SAMs 3NH2 and 11NH2 have surface pKa values of 8.5 35 and 8.9 34 respectively, and therefore were modelled with all molecules in their protonated state (NH3 + , pH 5 and 7.5). Systems were solvated with explicit water (density of approx. 1 g/cm 3 ), counter ions (to ensure charge neutrality) and in the second stage of simulations, two p-PDA or 4-APA molecules were also added.
Computations were performed using the GROMACS 4.6.5 software package 36 in conjunction with the GolP-CHARMM 37 force field and the modified TIP3P 38 water model. The GolP-CHARMM force field was employed since it was specifically designed to capture organic species adsorption at aqueous Au(111) interfaces via a combination of experimental and first-principles data, and the force field contains explicit terms to describe the dynamic polarization of Au atoms, chemisorbing species, and interactions between sp 2 hybridized carbon atoms and Au. To emulate the strong Au-S bonding formed between thiols and gold, 39 sulfur atoms were position restrained to the X-Y plane ~2.3 Å 40, 41 above the Au(111) surface. Together with a Lennard-Jones Au-S non-bonded potential (σ = 2.05 Å, ε = 3.2 kJ/mol), this allowed for the lateral mobility and reorienting of SAM molecules on the Au surface. 42 The ParamChem server (https://paramchem.org) was used to obtain CHARMM compatible parameters 43 and atomic partial charges for the p-PDA, 4-APA and SAM molecules.
Non-bonded long-range electrostatic interactions were evaluated using the Particle Mesh Ewald (PME) method with a real space cut-off of 12 Å and a 1.2 Å fast Fourier transform (FFT) grid spacing.
Van der Waals interactions were computed with a force-switch cut-off starting at 0.9 Å and ending at 10 Å. Energy minimization (EM) was performed using the steepest descent algorithm to remove steric clashes. The MD was performed in the canonical (NVT) ensemble with temperature maintained at 300 K using the Nosé-Hoover thermostat 44,45 , an integration time step of 1 fs, and the LINCS algorithm 46 to constrain bonds with hydrogen atoms. Analysis trajectories were outputted at a rate of one frame every 2 ps.
The simulation protocol for stage (i) and (ii) of the simulations was as follows. In stage (i), EM was followed by a series of 200 ps long MD simulations that sequentially increased the number of water molecules to fill the cavities formed by the relaxing SAM, until the solvent density plateaued at 1 g/cm 3 in the center of the unit cell. Subsequently, MD of each system was performed for 100 ns to determine the equilibrium structure and properties for each SAM. This process was independently repeated three times using different initial atomic velocities effectively resulting in 600 ns of statistics per SAM system (100 ns × 2 surfaces × 3 repeats). Analysis was conducted on the equilibrated part of the trajectory of each simulation (final 50 ns), determined by monitoring SAM molecule RMSD and average tilt angle convergence (not shown). The final frame from each simulation in stage (i) was then extracted (including the solvent) and used as the initial structure for the stage (ii) simulations. In stage (ii), two identical but randomly oriented p-PDA or 4-APA molecules were added to each unit cell at positions ~2 nm from each SAM-solvent interface (top/bottom) and over 3 nm apart from each other to ensure there were no immediate self-interactions and close contacts with each SAM. Any water molecules within a radius of 1.2 Å of the analytes were removed, EM was conducted, and 10 ns of NVT MD was performed. This was repeated five times per SAM configuration with random initial analyte orientations to increase statistics. A total of 300 ns per system (10 ns x 2 analytes x 5 orientations x 3 initial SAM configurations) was collected for data analysis. To investigate the interactions of p-PDA and 4-APA with gold specifically, MD simulations were also performed of individual analyte in the presence of a bare Au(111) surface, using the same approach as described above. Overall, stage (ii) resulted in a total of 5.4 μs of simulation time.
Model building, statistical analysis, and visualization of the data was performed using the GROMACS 4.6.5 suite analysis tools 36 Table 5 -6) were characterized with four parameters: analyte-Au distance (d), analyte orientation angle (φ), benzene orientation angle (ψ), and the percentage of simulation time the analyte spent in close proximity (< 0.6 nm) to the SAM/Au surface. Monolayer thickness, x, was measured as the average perpendicular separation between the gold surface atoms and SAM molecule terminal heavy atoms, i.e. methyl carbon (CH3), hydroxyl oxygen (OH), carboxyl(ate) oxygens (COOH/COO -), and amine nitrogen (NH3 + ). SAM tilt angle, θ, was determined as the angle of the molecular backbone (principal axis of inertia) with respect to the substrate normal. 11-carbon SAM molecules with a tilt angle > 45° were considered as collapsed chains/defects present on the SAM surfaces, and therefore to avoid skewed data and obtain accurate representative estimates for x and θ, measurements from these chains were excluded from the average values reported in Supplementary Table 4. Analyte-Au distance, d, was measured as the distance between analyte center-of-mass and the nearest Au(111) surface atom. The analyte molecular orientation relative to Au(111) surface, φ, was measured as the angle between the analyte molecular vector and the gold substrate normal vector. For p-PDA, the molecular vector was defined between nitrogen on NH2 and nitrogen on NH3 + , while for APA this was between nitrogen on the NH2 and carbon on COO -. The normal vector of the analyte benzene ring and the normal vector of the gold surface were used to measure the angle ψ. Average values presented in Supplementary Table 5-6 are only for simulation frames where analytes were proximate (< 0.6 nm) from the SAM/Au interface.