Silver nitrate (AgNO₃) was purchased from Sinopharm Chemical Reagent. Trisodium citrate (TSC) was obtained from Shanghai Lingfeng Chemical Reagent. Rhodamine 6G (R6G) was purchased from Aladdin Chemical Reagent. Gymnodimine (GYM) was acquired from National Research Council Canada. All purchased reagents were all analytic grade and directly used without purification. Deionized water was utilized throughout this research.

UV–vis spectra of Ag colloids were obtained by T6 spectrophotometer (General Instrument Co. Ltd., Beijing, China). Shellfish meat homogenate was prepared by Wiggens D130 homogenizer (Germany). Scanning electron microscopy (SEM) (Scios 2 HiVac, FEI) and transmission electron microscopy (TEM) (F20, FEI) are used to characterize the morphology of the SERS substrate. A laser confocal Raman spectrometer (LabRAM HR Evolution, Horiba) coupled with an objective lens (numerical aperture (NA) of 0.75) was employed to obtain SERS spectra, where a 532 nm laser light source and 600 g/mm grating were chosen for measurements.

All DFT calculations were performed in Gaussian 09 software package. Vibrational mode assignments of the calculated peaks were made by observations of molecular animations in the Gaussview program. In this study, B3LYP method and ef2svp basis set were used to optimize the molecular geometrical structure and simulate the theoretical Raman spectroscopy of gymnodimine. Given solvent effects, all calculations were performed in a methanol environment.

Silver nanoparticle colloids were synthesized by heating silver nitrate solution with trisodium citrate at high temperatures [28] . First, 45 mg silver nitrate was dispersed in 250 mL of deionized water, and heated to boiling in a three-neck flask. Then, 5 mL of 1% trisodium citrate (TSC) solution was immediately added. The mixture was heated for another 1 h to form a silver-gray solution and cooled to room temperature. After being washed and centrifugated 5 times, respectively, the product was redispersed in deionized water and stored in the dark at −4˚C.

Rhodamine R6G was introduced as a Raman probe and prepared at concentrations ranging from 10⁻³ M to 10⁻¹¹ M in deionized water. The stock solution of GYM with 100 μM is diluted with methanol stepwise to prepare analyte concentrations from 10 μM to 0.1 μM. For SERS measurements, 10 μL of AgNP colloids were dropped on the silicon wafer, which was then overlaid with 20 μL of gymnodimine solutions of different concentrations and dried in air. SERS signals were performed on a LabRAM HR Evolution system with a 532 nm laser wavelength, 50 mW of nominal power, and a 50× objective. Signals of 20 regions on the substrate were randomly collected on a silicon wafer to investigate the uniformity of the substrate. The spectra were recorded in the range of 540 - 2000 cm⁻¹, with an acquisition time of 10 s and the accumulation of 2 times. The peak position calibration of Raman spectroscopy was operated on the silicon wafer at 520.7 cm⁻¹ before detection.

The effect of the individual factors on the GYM detection process was determined by the volume of TSC, the concentration of AgNPs, and the ratio of GYM/AgNPs. Response surface methodology (RSM) was used to evaluate the collective influence of the above three variables and establish a model to optimize the response.

Based on the results of previous single-factor experiments, a three-variable Box-Behnken design (BBD) experiment was designed, considering the volume of TSC (Factor A), the concentration of AgNPs (Factor B), and the bulk ratio of GYM/AgNPs (Factor C). The Design-Expert software 8.0.5b (Stat-Ease, Inc., Minneapolis, USA) was used for experiment design and statistical analysis. Table 1 shows the level of factors in the BDD model. According to BBD design, the results of variables on the SERS signal were composed of 17 experiments, which included three repeated measurements of crucial points to eliminate system errors. Data from BBD were fitted to the quadratic polynomial model as equation (1):

Y = β 0 + ∑ ​ β i i X i 2 + ∑ ​ β i j X i X j (1)

where Y is the response factor, X_i and X_j are independent variables, and β_i, β₀, β_ii, and β_ij are the coefficients of linear, constant, quadratic, and cross-product, respectively.

The statistical significance of the BBD model was assessed, and parameters that significantly affected responses were determined by using analysis of variance (ANOVA) and Fisher’s statistical test (F-test) at the 95% confidence level (p-value < 0.05). The three-dimensional response surface was presented to explain the influence between variables on the SERS signal of gymnodimine.

The mussels were purchased from local seafood markets, weighing about 500 g and evenly sized. The water samples were taken from a local lake. 200 mL of lake water was filtered with quantitative filter paper and 0.22 μm microporous membrane to remove impurities, then stored at room temperature for future use. Then, 100 mL of methanol/water (1:1, V/V) was added to the homogenate of shellfish. The solution was ultrasonically treated for 20 min for further extraction and centrifugation. The supernatants were removed from impurities similarly and stored at −4˚C for further use.

The 100 μM of GYM solution was mixed with the above two samples, respectively, to prepare the GYM spiked samples with a concentration range of 0.1 - 10 μM. In the same way, the spectra were recorded by the optimized conditions and SERS analysis procedure.

Figure 1(a) and Figure 1(b) show the molecular structural formula and optimized

geometries of gymnodimine. As expected, gymnodimine possesses a characteristic cyclic imine structure as a toxic group and oxygen bridge. The calculated DFT spectrum and experimental SERS spectrum of gymnodimine in the 540 - 2000 cm⁻¹ range are shown in Figure 1(c), and the peak intensities of the spectra were expressed as normalization approaches, respectively. A good fit of characteristic peak bands between the theoretical calculation and the SERS spectra is obtained, mainly appearing at 620, 820, 1047, 1298, 1416, 1452, and 1735 cm⁻¹. It also indicates the consistency of the predicted RB3LYP/ef2svp method for this molecule.

A detailed assignment of the prominent peaks and vibrational mode of gymnodimine are summarized in Table S1. Due to C-C stretching vibration and -OH in-plane bending vibration, an intense band appeared at 1061 cm⁻¹, whereas being given at 1047 cm⁻¹ in the DFT calculation. The Raman peaks at 1301 and 1443 cm⁻¹ are assigned to out-of-plane and in-plane bending vibration of -CH₂, respectively. The bands at 612 cm⁻¹ are related to C-C stretching vibration and -CH₂ in-plane bending vibration. The stretching vibration of C-O and the in-plane bending vibration of C-C are mainly contributing to the SERS band at 820 cm⁻¹, while the out-of-plane bending vibration of C-H and -CH₂ is associated with the bands at 1407 cm⁻¹. However, a DFT theoretical spectrum infers that the band of 1735 cm⁻¹ is attributed to the stretching vibration of C=N, which is not observed in the experimental spectrum.

Gold, silver, and copper are the most commonly used active substrates [29] in SERS technology. SERS enhancement of the three substrates was studied first. Since the response of the metal surface plasma to the excitation light source wavelength is different, a specific range of laser wavelengths can effectively excite the surface plasmon resonance of the substrate material [30] . Therefore, the 532 nm laser is chosen for signal enhancement. The Raman spectral results are shown in Figure S1. Compared with AuNPs and CuNPs, it is evident that the SERS spectrum of AgNPs is most consistent with the DFT theoretical calculation. Moreover, AgNPs substrate enhances the SERS intensity of the characteristic peaks of gymnodimine. However, AuNPs show shoulder peaks in 1330 - 1530 cm⁻¹, which obscured the relevant characteristic peaks. Meanwhile, for CuNPs, the single Raman peak at 1443 cm⁻¹ can be enhanced with significant background noise.

Figure S2 shows the SEM and TEM images of silver colloids. The sliver colloids exhibit uniform quasi-spherical particles with a 60 - 70 nm diameter. The UV-Vis absorption spectra of AgNP colloids are shown in Figure 4, and it indicates a clear absorption peak at 422 nm, confirming the formation of AgNPs. The value of full width at half-peak (FWHM) depends on the dispersion of the nanomaterial. The larger the FWHM, the stronger the dispersion of the material is [31] . Figure S3 also shows a small shoulder around 350 nm, caused by the multilevel transition of the surface plasmon due to the increased size of AgNPs [32] .

We have explored the SERS performance of Ag colloids by using rhodamine R6G as a probe molecule. Figure S4(a) shows the Raman spectrum and comparative SERS spectra of rhodamine R6G at a concentration of 10⁻⁶ - 10⁻¹¹ M on silver colloids, and it can be found significant enhancement effect of spectral signal. The characteristic peaks of R6G at 612, 722, 1183, 1361, 1510, and 1650 cm⁻¹ are demonstrated, respectively, similar to the literature data [33] . The SERS spectra show that R6G solution can be detected at concentrations as low as 10⁻¹¹ M. The equation for calculating the Raman enhancement factor (EF) of the substrate material according to the method of McFarland et al. as Equation (2) [30] :

E F = I S E R S / N S u r f I R S / N V o l (2)

where I_RS and I_SERS are the normal Raman signal intensity of the rhodamine R6G solution and the SERS signal intensity at the lowest detectable concentration on the silver colloidal substrate. N_Surf and N_Vol are the average number of molecules detected on R6G solutions with or without SERS substrate materials. The EF calculation results of the silver colloidal base at each characteristic peak of R6G are shown in Figure S4(b). The EF values of R6G at 1510 and 1650 cm⁻¹ are measured to be 1.23 × 10⁸ and 1.17 × 10⁸, respectively, which are greater than the value of others.

The electromagnetic enhancement mechanism is closely related to experimental factors on the SERS signal intensity [34] [35] [36] . Thus, three factors, including the amount of trisodium citrate, the concentration of AgNPs, or the bulk ratio of AgNPs/GYM, are changed to achieve a superior signal response. The amount of reducing agent can lead to different sizes, although the mechanism of Raman signal enhancement is not unified [37] . As shown in Figure 2, TSC amount affects the SERS signal response on GYM, and the best enhancement effect is achieved at 9 ml TSC set. Considering the TSC influence, 7, 9, and 11 mL of TSC were used as three levels in the BBD model. In addition, the AgNPs concentration effect was investigated in Figure 3. The SERS response for gymnodimine improved when the concentration increased from 2 to 5 mg/mL, due to the extended hot spot areas [38] . However, the SERS signal declined when the concentration exceeded 5 mg/mL, probably explained by lower adsorption of the analyte or the obstruction of excitation light irradiation according to the AgNps aggregation at higher concentrations [39] . Thus, the concentration of AgNPs with ranges from 4 to 6 mg/mL can be considered as three levels in the BBD model.

Figure 4 shows SERS spectra and signal intensity of gymnodimine based on AgNPs/GYM bulk ratio. The characteristic peaks of GYM began to appear when the ratio was 1:4, but the low efficiency of the laser through GYM molecules

induced surface plasmon resonance in AgNPs caused by the low ratio, which resulted in a minor enhancement of GYM. With the increase of the AgNPs/GYM bulk ratio, the number and the signal intensity of characteristic peaks also increased gradually. However, the SERS signal becomes ambiguous as the ratio increases, which probably generated characteristic peaks about methanol solvent, residual reducing agent, or AgNPs substrate. Thus, only the suitable ratio could better reflect the SERS spectrum of low-concentration gymnodimine [40] , and the AgNPs/GYM bulk ratio was regarded as a meaningful variate in the BBD model.

Based on BBD from RSM, 17 experiments were designed and done to optimize the Raman tests further. The coding levels of the solitary factors used in the experimental design and results are shown in Table S2. Design-expert 8.05b software was used to perform quadratic regression analysis on the data, and the quadratic polynomial regression equation was shown as Equation (3):

Y 1 = 673.40 − 16.50 X 1 + 41.12 X 2 − 63.13 X 3 + 5.75 X 1 X 2 − 5.25 X 1 X 3 − 16.50 X 2 X 3 − 220.45 X 1 2 − 190.70 X 2 2 − 242.20 X 3 2 (3)

where Y₁ was the Raman intensity of GYM at 1443 cm⁻¹, and X₁, X₂, and X₃ were the volume of TSC, the concentration of AgNPs, and the GYM/Ag NPs bulk ratio, respectively.

The results of ANOVA revealed that the experimental data were suitable for the quadratic models (Table 2). The p-value was employed to evaluate the significance of each coefficient and indicate the interaction pattern between variables. For GYM, X₂, X₃, X 1 2 , X 2 2 and X 3 2 were significantly (p < 0.05), while X₁, X₁X₂, X₁X₃ and X₂X₃ were not (p > 0.05). The one-way ANOVA and Duncan’s multiple range tests indicated that the models were highly significant (p < 0.0001) with high F values (82.21). The coefficient (R²) value of the regression analysis for GYM was 0.9906, which revealed that the model was significant for accurate response prediction, and the adjusted coefficient (Adj R² = 0.9716) was also a positive value to certify the significance of the model. Therefore, it inferred that the proposed model was appropriate for analyzing the response tendencies.

Regarding the effects of linear and interactive factors, three-dimensional response surface plots (Figure 5) were established. While keeping other variables at optimal levels, the response (Intensity) was plotted on the Z-axis for either of the two independent variables. For GYM, the experimental data determined the following optimum operating parameters: volume of TSC of 10.32 mL, concentration of Ag NPs of 5.34 mg/mL, and GYM/Ag NPs ratio of 1.17. Triple parallel experiments were conducted to show that the intensity of the GYM was 687 (relative standard deviation of 3.41%), which was consistent with the predicted value. These results confirmed that the response model sufficiently reflected the expected optimization.

Gymnodimine is a cyclic imine marine toxin, which is of increasing concern because of its low content and long-term toxicity in living organisms. Since gymnodimine is trace and challengable to extract [41] , it is vital to establish an accurate and rapid method for determining GYM. SERS has been proven to be a prospective tool to realize this goal. To study the sensitivity of SERS substrate material, a volume of 10 μL of gymnodimine methanol solution was mixed with

*Significant, ^NSnot significant.

AgNPs colloidal solution on the Si wafer for Raman measurement. SERS spectra of different concentrations (10, 8, 6, 4, 2, 1, and 0.1 μM) are collected for detection quantitatively.

Figure 6(a) illustrates the SERS signal of prominent peaks with different concentrations of gymnodimine. As the concentration decreases, the intensity of two characteristic peaks located at 1301 cm⁻¹ and 1443 cm⁻¹ decreases simultaneously, suggesting a positive cooperation between GYM concentration and peak intensity (Figure 6(a) and Figure 6(b)). Thus, the spectral intensity can be used to quantify GYM in a methanol solution. Table S3 shows the linear relationship and goodness of fit (R²) in different matrices regarding the intensity of different characteristic peaks versus GYM concentration. Notably, the most excellent fitting equation based on peak intensity at 1443 cm⁻¹ in Figure 6(b) is y = 86.224x + 46.771 (where y is SERS intensity, x is GYM concentration) with R² of 0.9873. Compared with the blank signal, characteristic peaks of GYM based on AgNPs colloidal can still be observed clearly in a level of as low as 0.1 μM, revealing an efficient sensitivity of SERS substrate. In addition, the limit of detection (LOD) value of GYM in methanol solution was calculated as 0.085 μM and 0.105 μM at 1301 and 1443 cm⁻¹ according to the equation of LOD = 3 δ/k [42] [43] . In the formula, δ stands for the standard deviation of spectral intensity of AgNPs colloidal substrates at 1301 and 1443 cm⁻¹, and k represents the slope of the traced calibration curve. The above results prove that the detection of shellfish toxin GYM is feasible on prepared AgNPs colloidal substrate.

It is renowned that the reproducibility of SERS signals is also of great importance that ought to be concerned to demonstrate the prepared nanomaterial as an outstanding SERS substrate [44] . To investigate the reproducibility of GYM based on AgNPs active-substrate, a set of twenty SERS spectra of GYM molecules from different areas on a silicon wafer is acquired under the same experimental conditions as shown in Figure 6(c). The characteristic peak bands collected at different random areas have slight variations in the intensity of Raman

spectra and no significant change in position, indicating that AgNPs colloidal as the SERS substrate displays well reproducibility. The corresponding intensity and relative standard deviation (RSD) of correlative characteristic peaks at 1301 and 1443 cm⁻¹ are verified in Figure 6(d). Generally, RSD values controlled to be less than 20% are regarded as encouraging results [45] . The calculated RSD values of 10.7% and 8.3%, respectively reveal the excellent performance and feasibility in real sample detection as a suitable substrate.

To further illustrate the practicability of the AgNPs colloidal substrate, the GYM spiked to the lake and shellfish matrix were detected, respectively. As shown in Figure S5, the SERS signal of GYM in real samples was recorded to compare with that of methanol solutions and AgNPs colloidal. As labeled, six typical Raman peaks are enhanced in terms of standard GYM solution at 612, 834, 1061, 1301, 1407, and 1443 cm⁻¹. However, characteristic peaks of real samples at 612, 1061, and 1407 cm⁻¹ overlapped with that of methanol solutions. Similarly, the Raman peak at 834 cm⁻¹ corresponds to that of the proposed substrate. Therefore, characteristic peaks of standard GYM solution at 1301 and 1443 cm⁻¹ were extracted for the quantitative analysis of GYM residues in real samples, which were shifted to 1310 and 1447 cm⁻¹ in the lake sample, while 1295 and 1445 cm⁻¹ in the shellfish matrix, as specific “fingerprint-like” peaks. It is worth noting that the position of characteristic peaks was changed, probably attributed to the competitive adsorption of unknowns in complex samples with the target molecules and the obscuration of the original signal by clutter peaks. Figure 7(a) recorded the SERS spectra of the GYM with different concentrations in treated lake samples. With the increase of GYM concentration from 0.1 to 10 μM, the related characteristic peaks intensity enhanced. The linear relationship of SERS

intensity at 1310 and 1447 cm⁻¹ with the concentration of GYM is illustrated in Figure 7(c). The superior calibration curve at 1447 cm⁻¹ was y= 117.555x + 74.237 with R² of 0.9913, while the LOD value was measured to be 0.148 μM.

In the same way, the detection of GYM with different concentrations in the treated shellfish matrix, as shown in Figure 7(b), exhibited a reduction in relevant peak intensity with the decrease of GYM concentration. Figure 7(d) shows the linear formula of shellfish matrix samples based on the 1295 and 1445 cm⁻¹ bands. The preferable formula at 1445 cm⁻¹ was y= 72.658x + 86.679, R² was 0.9897, and the LOD value was calculated to be 0.170 μM. The recovery experiment in Table S4 demonstrates that Ag colloid based SERS is suitable for gymnodimine detection in the lake and shellfish matrix samples.