Ag colloidal nanoparticles (CNP) were prepared by reduction of AgNO₃ (Sigma- Aldrich, ≥ 99.0%) with NH₂OHHCl (Sigma) as described by Leopold and Lendl (Leopold and Lendl, 2003). Briefly, a 90 mL NH₂OHHCl solution with a concentration of 1.6 * 10^–3 M was prepared. 300 μL of 1 M NaOH (Fisher Chemical) was then added to this solution. The solution was stirred continuously at 350 rpm as 10 mL of 1*10^–2 M AgNO₃ was added dropwise to the solution. The solution was stirred for an additional 45 min. See Supplementary Figure S1 for UV- Vis spectra of CNP syntheses used in this study. Dynamic light scattering (DLS) of a prepared silver colloidal suspension following our experimental conditions was performed using a NanoBrook particle analyzer (Brookhaven). The average effective particle diameter was measured to be 56.42 nm with an average polydispersity index of 0.312. See Supplementary Figure S2 and Supplementary Table S1 for the differential distribution of particle diameter, instrument parameters, and summary statistics.

Powder TFV and adenosine were provided by CONRAD at Eastern Virginia Medical School and manufactured by Gilead Alberta ULC. Serial dilutions from an aqueous stock of 100 μg/mL TFV and adenosine were performed for subsequent concentration gradients. An aluminum plate containing 40 machined wells in an 8 × 5 well-plate, each having a volume capacity of 20 μL, was used as our SERS surface. Certified 1,100 aluminum was used for the well plate construction. In between experiment replicates, the plate was cleaned by soaking in a 1 M NaOH bath and hand milling with a drill bit, followed by deposition of 4 M HClO₄ into the wells and thorough DI H₂O rinsing.

All samples presented in this study underwent the preparation and spectra acquisition parameters described in the preceding section. A 1:1 mixture of Ag CNP and analyte was prepared in a microcentrifuge tube and vortexed for 5 seconds. 20 μL aliquots of this mixture were added into each well where each TFV concentration was deposited into five aluminum wells and taken to dryness in a chemical hood prior to spectra acquisition. Since the evaporative process is stochastic, the total volume for all reaction mixtures remained constant (i.e., equivalent volumes of CNP suspension and aqueous TFV or adenosine standards, where each standard is a different concentration). For the TFV double-aliquot experiment, this mixture, deposition, and evaporation process was repeated for one additional aliquot of 20 μL in each well. A high-resolution “square” scan was acquired of each well in the well plate following parameters described in Table 1. The platform by which statistical analysis is performed was developed by our group at Old Dominion University. A computerized XY stage (ThorLabs) holding the Wasatch Photonics 785 nm Raman spectrometer (Kent Lawson Wise and Lee Schoen, 2002) was used for automated spectra acquisition of each well in the aluminum well plate. A raster scan pattern was used for acquisition. See Figure 2 for images of the sampling apparatus. Additional acquisition software setup parameters are listed in Supplementary Figure S3.

For all experiments, 1806 spectra were acquired of each sample well. Statistical analysis was done on acquired spectra to generate calibration curves of TFV and adenosine. Adenosine was investigated in this study for the purpose of comparison with TFV analysis (see Figure 3; Figure 4) because the adenine ring-breathing peak was used for quantification of both analytes and TFV’s inherent derivation from adenosine. See Supplementary Figure S7 and Supplementary Table S2 for peak assignments from acquired TFV data compared to literature assignments. Calibration curves were also generated for the double aliquot TFV experiment (see Figure 5) to probe the benefit of evaporating our reaction mixture of TFV and CNP to dryness then depositing an additional aliquot. Spectra were statistically analyzed using quality indexes (Q_i) calculated from the intensity of the adenine ring-breathing peak in the spectral region 739–745 cm^-1.

The calculation of spectra quality indexes in general terms is described in the preceding paragraph. The subscript indexes, positions of cursors in the data processing program, p: peak; b₁: baseline 1; b₂: baseline 2. Baselines are defined in Eqs 2, 3. b 1 = p − 2 F W H M (2) b 2 = p + 2 F W H M (3)

Where FWHM is defined as the full width and half max of the selected peak(s) of interest. To maximize S/N of the selected peak(s) of interest, a summation of intensities at respective points, generally defined as point j, I_j, can be calculated to determine the average intensity about each peak at a user-defined range of points to the right of the selected cursor. The difference between these intensity summations with respect to peak(s) and their respective baselines are defined as a quality index (Q_i) used for spectral ranking in this study, defined in Eq. 4. Q i = ∏ k = 1 k = t 1 2 n + 1 ∑ j = p − n p + n I j − ∑ j = b 1 − n b 1 + n I j × ∑ j = p − n p + n I j − ∑ j = b 2 − n b 2 + n I j k 1 t (4) Q i < 0 ≝ Q i = 0

Where t is defined as the total number of peaks factored into the Q_i. The Q_i value is also conditional such that any calculated Q_i less than 0 is equal to 0, defined in Eq. 5. Supplementary Figures S4–S6 show scatter diagrams of all spectra Q_i values corresponding to each concentration for these experiment sets. All Q_i data corresponding to acquired spectra was then written into a new file and uploaded to a second data processing program.

The second data processing program employed a selection criterion that involved choosing a percentage of high-quality spectra prior to computing an average spectrum for each well. Each averaged spectrum corresponding to the same concentration, a total of five each because each concentration had five replicates, were then averaged, generating a single averaged spectrum for each concentration. This selection process was carried out based on the Q_i values computed in the first data processing program, allowing the second data processing program to sort the Q_i values for each well. The selected percentages of spectra from each well factored into the total averaged spectrum for each concentration was comprised of spectra corresponding to the highest Q_i values within that percent range. Using this program, three different percentages of spectra acquired for each replicate were averaged (Figures 3–5): 100% (1806 spectra), 50% (903 spectra), and 10% (180 spectra) and used for analysis in the preceding section. SERS Enhancement factors (EFs) were calculated for each dataset and are shown in Supplementary Table S3. Briefly, for each concentration, the spectrum corresponding to the maximum SERS intensity of the 741 cm^-1 peak was divided by the average 741 cm^-1 peak intensity for all spectra. These quotients were then multiplied by 10⁶, an established EF of a prominent adenine peak (733–740 cm^-1) (Sivaprakasam and Hart, 2021). From these calculations, our system of Ag CNP and analyte mixtures evaporated to dryness exhibited relative EFs between 10⁷ and 10⁶ across all concentrations.

For all three experiment sets, based on the analysis of the three percentage subsets of spectra above, using 50% of spectra (903 spectra) from each well for each concentration yielded linear correlations similar to that using 100% of the spectra for each well for each concentration (Table 2; Table 3; Table 4). However, the 50% spectral subsets showed increased S/N in the total averaged spectra compared to the 100% spectral subsets. The 10% spectral subsets generated averaged spectra with the highest S/N, however the smaller sampling size drastically increased the standard deviation and decreased the linearity across the three sample sets. Based on this result, 50% spectral subsets appeared to be the most representative of the entire data population corresponding to each concentration yet also sufficiently excluded spectra with low quality indexes and poor S/N with respect to the adenine ring-breathing peak in the 739–745 cm^-1 region. Limits of detection for all experiment sets were calculated by averaging the S/N for each TFV concentration and are shown in Supplementary Table S4.

A notable aspect of our analysis was investigating the impact of a second aliquot of the TFV and Ag CNP mixture after the initial aliquot was taken to dryness versus a single aliquot of the TFV and Ag CNP mixture with respect to spectra signal enhancement, analytical sensitivity, and the ability to differentiate between TFV concentrations 25 ng/mL, 40 ng/mL, and 50 ng/mL. From data presented in Figure 4 and Figure 5, it was shown that the double-aliquot TFV experiment exhibited superior signal-to-noise ratios in the averaged concentration spectra to the single-aliquot TFV averaged concentration spectra. Additionally, the double-aliquot TFV experiment demonstrated improved linearity in the concentration calibration curve with respect to the correlation coefficient, coinciding with improved discrimination between concentrations 25 ng/mL, 40 ng/mL, and 50 ng/mL. Furthermore, double-aliquot calibrations demonstrated higher analytical sensitivity across all subsets than single-aliquot calibrations as shown by the higher slope values (Tables 3, 4). This result is consistent with the benefit of our sampling apparatus such that evaporating the analyte and Ag CNP reaction mixture to dryness has a concentrating effect on the analyte of interest, and a second aliquot of the mixture taken to dryness enhances the SERS signal of our analyte even more, possibly from the increased amount of TFV detected at the sensitivity of the system. In addition to the improved distinction between 25 ng/mL, 40 ng/mL, and 50 ng/mL, these results suggest that with a second reaction mixture aliquot, there is better analyte/CNP well coverage, resulting in a greater amount of high Q_i spectra, highlighting another benefit of our sampling apparatus. In addition, the standard deviation for each concentration in the calibration curves also improved for the double aliquot experiments compared to the single aliquot experiments (see Supplementary Tables S5–S7). The standard deviation of the 40 ng/mL and 50 ng/mL data points were closer in value to each other in the double aliquot experiment, compared to the single aliquot experiment where the standard deviation of the 40 ng/mL data point was nearly double the standard deviation of 50 ng/mL. This is also observed between the 500 ng/mL and 400 ng/mL data points, where they are close in value in the double aliquot experiment, but the 500 ng/mL standard deviation is nearly double the 400 ng/mL standard deviation in the single-aliquot experiment. This improvement in both analytical sensitivity and precision suggests that the additional aliquot provides better well coverage and an increase in hot spots for sample wells.

Partial least squares (PLS) regression analysis was applied to the three experiments presented in this study (Figure 6; Figure 7; Figure 8). The collinear nature of each factor modeled by these regressions was used to predict responses of future factors with respect to these models whose TFV concentrations are unknown. Correlation coefficient, slope, and y-intercept values for these regressions are shown in Table 5, Table 6, and Table 7. Extending these findings into future work, our aim is to use these analytical methods of TFV and apply them to analysis of TFV in biological samples for detection and quantification. This is currently under investigation by our group.