Advancing SERS Diagnostics in COVID‐19 with Rapid, Accurate, and Label‐Free Viral Load Monitoring in Clinical Specimens via SFNet Enhancement

This study presents an integrated approach combining surface‐enhanced Raman spectroscopy (SERS) with a specialized deep learning algorithm, SFNet, to offer a rapid, accurate, and label‐free alternative for COVID‐19 diagnosis and viral load quantification. The SiO2‐coated silver nanorod arrays are employed as the SERS substrates, fabricated using a reliable and effective glancing angle deposition technique. A dataset of 4800 SERS spectra from 120 positive and 120 negative inactivated clinical human nasopharyngeal swabs are collected directly on the SERS substrates without any labels. A SFNet algorithm is tailored to adapt to the unique spectral features inherent to SERS data, achieving a test accuracy of 98.5% and a blind test accuracy of 99.04%. Moreover, an optimized SFNet algorithm unveils the capability of estimating SARS‐CoV‐2 viral loads, accurately predicting the cycle threshold values (Ct values) of the three vital gene fragments with a root mean square error (RMSE) of 1.627 (1.3 for blind test). The methodology is substantiated using actual clinical specimens and completed in <15 min, thereby strengthening its real‐world point‐of‐care applicability. This rapid and precise yet label‐free modality competes favorably with classical reverse‐transcription real‐time polymerase chain reaction (RT‐PCR) and marks an advancement in SERS‐based sensor algorithms.


Introduction
Recently, surface-enhanced Raman spectroscopy (SERS) has been extensively explored as a potential diagnostic platform for detecting SARS-CoV-2, owing to its remarkable sensitivity, ability to provide unique "signature" spectral features for different viruses, inherent simplicity, and capability for a point-of-care detection device. [1,2]The overarching goal of SERS-based SARS-CoV-2 detection is to establish a portable alternative to the reverse-transcription real-time polymerase chain reaction (RT-PCR) technique.Various detection strategies have emerged, such as direct detection of viral particles, [3][4][5] RNA hybridization, [6,7] spike protein capture, [8] as well as labeled approaches. [9,10]Notably, direct detection (label-free) methods have gained prominence because of the simple and cost-effective process, and their performance reported in the literature is summarized in Table S1 (Supporting Information) These methods are developed for three distinct aspects: classification, quantification, and a combination of classification and quantification.
Classification, a pivotal aspect of COVID-19 management and intervention, involves determining the SARS-CoV-2 status of specimens.As specimens are derived from either spiked buffer solutions or body fluids, inherent background SERS signals are presented in all measured spectra, leading to significant interference.][13][14][15][16] A noteworthy example is the successful differentiation of non-infectious lysed SARS-CoV-2 using support vector machine (SVM) analysis based on SERS spectra. [17]The distinction between SARS-CoV-2, A/influenza (H1N1), Marburg, and Zika viruses in spiked saliva has been achieved through a random forest (RF) algorithm, yielding varying accuracies from 85.4% to 95.6%. [4]Recent advancements showcase the potential for SERS spectra-based SVM classification, attaining >99% accuracy in detecting thirteen distinct respiratory viruses in saliva, including two SARS-CoV-2 variants. [3][13][14][15][16] For instance, SERS spectra collected from patients' saliva specimens were subjected to an SVM classifier, achieving a prediction accuracy of 95% for differentiating positive and negative COVID-19 cases. [13]Similarly, a residual neural network-based approach was used for the detection of the SARS-CoV-2 S antigen from clinical throat swab or sputum specimens, demonstrating an 87.7% accuracy. [14]Notably, the majority of these studies showcased classification performance by gathering SERS spectra from fewer than 30 patients, yielding accuracy levels spanning from 87.7% to 95%, requiring the research in tailoring the SERS substrates and algorithms to more accurate diagnosis.
[20] Typically, the SERS intensity of a specific peak, uniquely linked to a particular virus, is plotted against its concentration.Ansah et al. presented a calibration curve for the SARS-CoV-2 detection in saliva on the intensity of SERS peak at 732 or 964 cm −1 . [20]However, when quantifying clinical specimens, at least two challenges must be addressed.First, based on our recent study on 13 respiratory viruses, [3] most spectral features of SARS-CoV-2 are shared with other viruses (refer to Table S11, Supporting Information of Ref. [3]).[23] Second, the interference in the SERS spectra from background, stemming from buffers or body fluids, is significant.These backgrounds tend to overshadow spectral features from viruses, depending on the relative composition and components of virus and medium.Our previous SERS investigation on virus-spiked saliva with varied virus concentrations did not yield a straightforward correlation between SERS peak intensity and virus concentration. [3]A thorough examination of spectra across diverse virus concentrations showed subtle spectral alterations.To surmount these challenges and gain quantification understanding, MLAs and DLAs were explored to interpret the complex relationship between the viral load and signal.
In our previous work, SVM-based regression algorithms were employed to establish quantitative calibration curves for eleven respiratory viruses, accurately estimating unknown virus concentrations in buffer and saliva within a detection range of ≈195 -1 × 10 5 PFU mL.−1 [3] In a parallel effort, Hwang et al. developed a DLA-based autoencoder, followed by the targeted elimination of non-discriminatory SERS features of spike proteins, facilitating the quantification of 10 1 −10 4 PFU mL −1 SARS-CoV-2 lysates in aerosols with an accuracy surpassing 98%. [24]o bring SERS-based diagnostic methodologies in line with the accuracy of RT-PCR, it is essential to establish three crucial attributes while capitalizing on the inherent benefits of SERS: First, one needs to achieve effective classification of SARS-CoV-2 positive or negative specimens; second, it is important to accurately quantify SARS-CoV-2 viral load, i.e., the cycle threshold value (Ct value) from RT-PCR, which is defined as the number of cycles required for the fluorescent signal to surpass the threshold; and lastly, successfully detect the virus within clinical specimens.Therefore, there is an emerging need to develop a strategy that combines SERS with MLA or DLA to classify and quantify SARS-CoV-2 infection from real clinical specimens.In addition, this strategy should be able to perform a direct comparison with RT-PCR results based on Ct-values.
This study directly compares the use of SERS and specialized DLA for detecting and quantifying SARS-CoV-2 in clinical human nasopharyngeal swab (HNS) specimens to the results from RT-PCR.The process involves mixing the HNS specimen with a virus inactivation buffer, placing a droplet on a SiO 2 -coated silver nanorod array (AgNR@SiO 2 ) SERS substrate, and collecting multiple SERS spectra without labels from different substrate locations after drying.Two new DLAs called SFNets are custombuilt, one for classification, and the other for regression.The classification model distinguishes positive and negative HNS specimens with a remarkable 98.5% test accuracy, while the regression model predicts RT-PCR Ct values with an average root mean square error (RMSE) of 1.627.Notably, both performances depend on inherent SERS spectral differences in viral components.Blind tests on 104 unknown HNS specimens show that the SERS-DLA approach achieves 98.28% test accuracy for positive specimens and 100% test accuracy for negative ones, an overall accuracy of 99.04%.Ct values are predicted with a small RMSE of ≈1.3.These outstanding outcomes demonstrate the favorable performance of SERS-DLA compared to RT-PCR, providing a direct, rapid, and reliable point-of-care COVID-19 diagnostics platform and novel insights for tailoring algorithms for SERS spectra.

General Detection and Classification Strategy
The procedure using SERS and DLAs to directly differentiate and quantify SARS-CoV-2 positive and negative HNS specimens is shown in Figure 1.The detection strategy consists of background and forward efforts.The background effort includes establishing SERS spectral database from known SARS-CoV-2 positive and negative HNS specimens and developing DLAs according to the spectral database.The forward effort is to use the established DLAs to validate the SARS-CoV-2 detection, i.e., a blind test.The background effort consists of four steps: first, HNS specimens are collected in viral inactivation buffers to allow all the subsequent procedures to be performed at BSL-2.Second, the inactivated HNS specimens are dispended on AgNR@SiO 2 substrates for SERS measurements.After repeated SERS spectra are collected from different locations of the substrates and from known positive and negative specimens, an appropriate baseline correction method and spectral normalization are applied to all spectra.Finally, two DLAs for classification and quantification are developed, the corresponding model parameters are optimized, and the models are cross-validated and tested.These DLA models with optimized parameters are used for blind tests.

The SERS Spectra of HNS Specimens
The plasmonic property of the substrates can be characterized by measuring the UV-Vis reflection spectra since the AgNR array was grown on a 100 nm-Ag film.Figure S1B (Supporting Information) shows the typical reflection spectra of both the AgNR and AgNR@SiO 2 substrates.Before and after SiO 2 coating, the spectra appear quite similar: the reflection increases with wavelength  from 320 to 800 nm.At  = 785 nm, the reflection is 73.2% for the AgNR substrate and 71.3% for the AgNR@SiO 2 substrate.To characterize the SERS enhancement, we compared the SERS signals of BPE (trans-1,2-bis (4-pyridyl) ethylene) from the AgNR and AgNR@SiO 2 substrates under the same conditions, as shown in Figure S1C (Supporting Information).The peak intensities of the average BPE spectrum of AgNR@SiO 2 substrate are ≈60% of those from AgNR substrate.Since the AgNR substrate is a well-studied SERS substrate with an enhancement factor close to 10 9 , [25] the estimated enhancement factor for AgNR@SiO 2 substrate was determined to be ≈6×10 8 .Based on our previous studies, the SERS enhancement mechanism is mainly due to the hot spot between the nanorods as well as on the tip of the nanorods. [26][29][30][31] In addition, according to the SEM image of AgNR@SiO 2 substrate shown in Figure S1A (Supporting Information), the surface of tilted nanorods has a wide range of morphologies, such as corrugations, bifurcations, protrusions, as well as random irregularities, which could become additional SERS "hot spots" locations.The uniformity of the AgNR@SiO 2 substrate was characterized by mapping the BPE peak intensity (Δv = 1200 cm −1 ) as shown in Figure S1E (Supporting Information).The intensity mapping shows a uniform distribution of BPE intensity, with an average intensity of 28 000 a.u. and a variation of 7%.And the bath-to-batch variation of AgNR is ≈10% as shown in Figure S1F (Supporting Information).Other detailed characterizations of AgNR@SiO 2 substrates can be found in ref. [32] and SI of ref. [3] Although the enhancement factor of AgNR@SiO 2 substrates decreases compared to AgNR substrates, the SiO 2 coating improves the long-term stability and reliability of AgNR substrates in biomedical applications.
Figure 2A (with complete spectra in Figure S3, Supporting Information) plots the baseline-corrected and normalized SERS spectra for 120 positive and 120 negative HNS specimens, as well as the inactivation buffer.The SERS spectra of the inactivation buffer exhibit clear and consistent peak features at specific wavenumbers (Δv = 930, 1005, 1187, 1303, 1452, and 1598 cm −1 ), attributed to various molecular compositions like guanidinium-HCl, Tris-HCl, and EDTA within the buffer.Such consistency is reflected in Figure 2B, where the primary peaks of the inactivation buffer align well with these composition-related SERS peaks.While the SERS spectra of both positive and negative HNS specimens share similar overall features with the inactivation buffer, especially when considering the average spectra (shown in Figure S4, Supporting Information), they exhibit more fluctuations.The inactivation buffer (1 m guanidine hydrochloride, 0.2% Triton ×-100, 1 mM EDTA, and 2 m Tris-HCl) is commonly used in viral RNA purification by dissociating the virus into nucleic acid and protein fragments. [33,34]The chaotropic agent, guanidinium-HCl, unfolds proteins and breaks into polypeptide chains.The viral proteins are dissolved by Triton X-100 micelles as the viral lipid envelope is destroyed. [35]Therefore, inactivation of SARS-CoV-2 viruses in the HNS specimens results in various compositions including proteins, RNA, etc., as shown in Figure 1 Step 1.To ensure that the AgNR@SiO 2 substrate indeed has the SERS enhancement for the inactivated virus specimen, a Raman measurement was carried out for an inactivated positive specimen dispensed on a Si wafer substrate, and the spectrum obtained was compared with that obtained from a AgNR@SiO 2 substrate under the same sample preparation and Raman measurement conditions.As shown in Figure S5 (Supporting Information), there are no obvious peaks in the Raman spectrum measured from the Si wafer substrate (black curve), while multiple strong peaks are observed in the spectrum measured from the AgNR@SiO 2 substrate (red curve).This result clearly demonstrates the SERS enhancement effect for the AgNR@SiO 2 substrates.
Thus, the SERS spectra obtained from the positive and negative HNS specimens may have three contributions: 1) inactivation buffer, 2) HNS compositions, and 3) viral fragments in positive specimens.For the negative specimens, the physical origin of signal I Negative can be expressed as: I Negative = I NS + I buffer + I noise , where I NS is the SERS intensity originating from nasopharyngeal swabs, I buffer is the SERS signal of inactivation buffer, I noise is electronic noise inherent to the Raman instrument, independent of the instrument's optical response.For the positive specimens, the physical origin of signal I positive can be expressed as: I positive = I RNA + I proteins + I lipid + I other + I NS + I buffer + I noise , where I RNA , I proteins , I lipid , and I other are the SERS intensity from different components of SARS-CoV-2 viruses: RNA, proteins, lipids, and others.The molecules in inactivation buffer, being small, readily adsorb to SERS "hot spots", while viral fragments and HNS compositions, larger in size, face challenges in adsorption by these "hot spots", and the buffer quantity offsets other constituents during sampling.Therefore, the SERS spectra of inactivated specimens are dominated by the spectral features from the buffer, resulting in high-similarity spectral shapes from positive and negative specimens as well as the buffer.Consistent with this, a comparison between the SERS spectra from positive and negative HNS specimens indicates that spectra from negative specimens have greater uniformity and fewer variations.Conversely, spectra of positive specimens show considerable fluctuations in the 600 -900 and 1300 -1425 cm −1 ranges.Given the striking similarity among these three sets of spectra, as illustrated in both Figures 2A and Figure S4 (Supporting Information), traditional chemometric techniques like principal component analysis (PCA) and t-distributed stochastic neighbor embedding (tSNE) are difficult to discern the minute differences.Figure 2C,D present the PCA and tSNE plots, and all three clusters due to 3 sets of SERS spectra are intertwined together.Hence, simple classification methods are inadequate to differentiate the SERS spectra from inactivated HNS specimens.This prompts the introduction of MLAs and DLAs.

Deep Learning Model to Classify SERS Spectra of Positive and Negative Specimens
A SFNet deep learning model was developed to predict the SARS-CoV-2 status (positive or negative) of HNS specimens based on their SERS spectra.Recurrent neural network (RNN) architecture in the SFNet model allows for cyclic connections between nodes and enables outputs from certain nodes to influence subsequent inputs to those same nodes. [36]This property is particularly useful for handling variable-length sequence inputs, such as SERS spectra.Figure 3A shows the architecture of the SFNet model.A convolutional layer (Conv Layer) and a maximum pool (Max Pool) were first used to preprocess the input spectra.Then three consecutive blocks composed of a convolutional block (Conv block) and two identity (ID) blocks with shortcuts were connected to the previous Max Pool.The Conv block consisted of an initial Conv layer, followed by a batch normalization (BN), a corrected linear (ReLU) transform, and a Max Pool.The ID block had a similar framework without the Conv Layer and BN after the input.The Conv block and ID block can solve the problem of gradient disappearance of deep structures by marking shortcut connections from previous input data to output data to maintain previous gradient information, reduce the computational power, and increase the training efficiency.Two LSTM layers and a dense layer were The 4800 SERS spectra of positive and negative HNS specimens were trimmed, excluding the spectral features from 960 to 1080 cm −1 .This exclusion was driven by two factors: substantial fluctuations in normalized peak intensity (illustrated in Figure S3, Supporting Information); and the dominant presence of the peak at Δv = 1005 cm −1 due to the inactivation buffer.So only 982 1D float spectral data from 600 -960 and 1080 -1700 cm −1 per spectrum, were retained.In the binary classification, positive and negative HNS specimens were denoted as 0 and 1, respectively.In any neural network, the overfitting issue is one of the major concerns due to excessive correspondence to specific datasets. [36]To mitigate overfitting, 70% of the entire spectral set was employed for training, 15% for validation, and another 15% for testing.Different RNN units in the SFNet model, including simple RNN, gated recurrent unit (GRU), and bidirectional RNN, were initially compared to select the optimal unit.Subsequently, the chosen SFNet model underwent optimization involving hyperparameter tuning (e.g., various optimizers, learning rates, loss functions, and fully connected layers), as detailed in Table S3 (Supporting Information).With optimized hyperparameters, the spectral set was trained with increasing epochs.
Figure 3B shows the evolution of the loss function and accuracy against epochs for the test spectral set.The loss function rapidly decreased from 2.3 to 0.2 until ≈100 epochs, subsequently stabilizing at ≈0.025 after 200 epochs.Meanwhile, the accuracy of the test spectral set rose from 12% to 80%, eventually converging to 95%.At ≈650 epochs, accuracy reached 98.5%, marking the conclusion of training.The SFNet model's performance was evaluated using a confusion matrix (Figure 3C) based on the test spectral set.It achieved a 97.1% accuracy for positive spectra and a perfect 100% accuracy for negative spectra.The false-negative prediction for positive specimens could be due to non-uniform distribution of viral components on the SERS substrate from the trace amount of virus in the specimens.The receiver operating characteristic (ROC) curve (Figure 3D) yielded an area under curve (AUC) of 0.9921, indicating outstanding classification performance close to unity.In comparison to other MLAs (SVM and RF) and DLAs (MLP and CNN, details in Section S4 Supporting Information), the SFNet model excelled.As summarized in Figure 3E, the SFNet model exhibited the highest training and testing accuracies, reaching 99.8% and 98.5%, respectively.This demonstrated its superior classification performance relative to other models like SVM, RF, CNN, and MLP.
Given the striking similarity between positive and negative spectra in Figure 2A, comprehending the source of such high classification accuracy in the SFNet model becomes crucial.A method to achieve this understanding is through a "feature importance map (FIM)", which assigns higher importance values to specific wavenumbers based on their contributions to the SFNet model's identification of SARS-CoV-2.The calculation of the FIM is based on a full-gradients algorithm discussed in Section S5 (Supporting Information). [5,37]Figure 4A shows the resulting FIM based on the SFNet model, and spectral features spanning 699 -717 and 1219 -1678 cm −1 ranges emerge as key signatures in the SFNet model's classification process.By cross-referencing the FIM to existing Raman spectra of known biomolecules and chemical functional groups such as lipids, proteins, nucleic acids, amino acids, and amides, we can have a better understanding of which composition contributes the most in the SFNet model to differentiate SARS-CoV-2 positive specimens.A "matching score" is proposed for this purpose [5] : First, an important wavenumber range contributing to the SERS spectra's classification (above the 50% threshold) was extracted from the FIM, denoted as R SFNet .The purple lines beneath the FIM in Figure 4A indicate the significant R SFNet range surpassing the 50% threshold.Then, the wavenumber ranges of the Raman peaks of lipid, amide I, amide III, RNA, tyrosine, and phenylalanine were obtained from previous reports, [38,39] and the corresponding wavenumber ranges of the Raman peaks are designated as R K , which are presented as different colored segmented lines in Figure 4B.The extent of overlapping wavenumbers for each biomolecule between R SFNet and R K was calculated as R RNN ∩R K , where "∩" denotes overlap.The matching score was then defined by the ratio of R SFNet-K and R K , i.e., matching score = R SFNet ∩R K R K . A higher matching score indicates a greater likelihood of the specific biomolecule's Raman signatures contributing to distinguishing negative and positive SERS spectra during testing.The calculated matching score of each biomolecule is indicated as a percentage in Figure 4B.Notably, amide III and amide I exhibit high matching scores of 86.84% and 62.91%, signifying their crucial role in distinguishing SARS-CoV-2 spectra, followed by lipid (56.39%), phenylalanine (54.48%), and RNA (52.79%).These results are consistent with the composition of SARS-CoV-2.Specifically, the viral proteins consist of membrane proteins (M), envelope proteins (E), spike proteins (S), and N proteins, up to a total of 29 different kinds, [14,40] and the ratio of M, E, and S is ≈1:20:300, [41,42] compared to one strand of RNA.Both lipid and RNA are the major components of the virus but with substantially less amounts.The matching scores based on FIM indicate that the efficiency of the SFNet model is linked to the major genetic components of the SARS-CoV-2 virus as well as the host response to differentiate positive and negative spectra.The phenylalanine is a COVID-19 immune-responsive metabolism representing the severity of the infection, [43][44][45] which may vary from patient to patient.

The Quantification of Viral Load
A common output from RT-PCR is the Ct value which is defined as the number of cycles required for the fluorescent signal to cross the threshold, i.e., exceeds background level.The Ct value inversely correlates with the nucleic acid quantity, i.e., a lower Ct indicates a higher nucleic acid amount [46] and can be used as a semi-quantitative measure of the amount of viral RNA in the specimen, as discussed in Section S6 (Supporting Information).Therefore, to compare SERS-DLA based detection with RT-PCR, it is important to predict the corresponding Ct value in positive specimens.Regression models were constructed for predicting Ct values of three viral gene fragments: ORF1ab, N gene, and S gene, all obtained from positive specimens.Regression results for three viral gene fragments (ORF1ab, N gene, and S gene) of the predicted Ct value and actual Ct value from the SFNet regression model are plotted in Figure 5D.The predicted and actual Ct values of all three viral gene fragments follow the relationship Ct pre = Ct act , with an average small RMSE of 1.627 and the coefficient of determination R 2 of 0.955.These performance parameters underline the accuracy of the SFNet regression model.Such a good performance in the SFNet regression is possibly due to the viral components, such as viral proteins, lipids, and RNA, as well as biomarkers related to SARS-CoV-2 infection in HNS specimen, e.g., inflammatory mediators, breath volatile organic compounds, etc. [47,48] The predictive capabilities of Ct values for three viral gene fragments using five regression models -SFNet, CNN, MLP, SVR, and RF -are compared: the MLP and CNN results are presented in Figure 5D, while those for RF and SVR are shown in Figure S10 (Supporting Information).The corresponding RMSEs and R 2 values are summarized in Table S6 (Supporting Information) and Figure 5C.SVR's R 2 is ≈0.299 with an RMSE of 6.544, while RF's R 2 is roughly 0.328 with an RMSE of 6.072.For the two DLAs, the MLP model yields an R 2 of ≈0.226 with an average RMSE of 6.879, while the CNN regression model demonstrates an R 2 of ≈0.570, accompanied by an average RMSE of 5.128.Notably, the SFNet regression model stands out with the highest quantification performance, as shown in Figure 5C,D.
SVM and RF, both falling under the MLA category, exhibited weaker predictive performance compared to DLAs like MLP, CNN, and SFNet.This difference results from DLAs' capacity to effectively learn intrinsic data features and optimize loss functions.DLAs excel in fitting intricate non-linear relationships, possessing a higher-dimensional hypothesis space and enhanced representation abilities.Conversely, MLAs rely on predetermined features and exhibit reduced generalization capabilities.In the realm of DLAs, the order of performance is as follows: MLP < CNN < SFNet.The SFNet model's superior capabilities lie in its ability to process diverse input sequences through internal memory, allowing the hidden layer to correlate current and past data points for enhanced feature identification, particularly changes in SERS spectra trends.CNN and MLP models, focusing on local intensity at distinct wavenumbers, disregard correlations among spectral features in neighboring wavenumbers.CNN's improved performance over MLP arises from its reduced parameter requirements, lower model complexity, and optimized connection weights.Ultimately, the SFNet model achieves the highest predictive accuracy, making it the preferred choice for blind tests of patient specimens.
According to the record of Ct values from RT-PCR, the highest Ct values from RT-PCR results for the current deep learning model training are ≈40; therefore, the limit of detection (LOD) is determined to be a Ct value of 40.Additionally, considering specimen dilution, sample volume, and laser beam size used in SERS measurement, it is expected that the highest Ct value that SERS-DLA method can predict is 54 (see Section S6, Supporting Information), compared to 45 of conventional RT-PCR.Such an estimation shows the potential higher sensitivity of SERS-DLA compared to RT-PCR.

Blind SARS-CoV-2 Diagnosis with SFNet Model
The applicability of the established SFNet model was further assessed to predict the status of 104 additional deidentified HNS specimens (46 negatives and 58 positives) in a blind test.The SARS-CoV-2 status of these specimens, as determined by RT-PCR tests, remained concealed from the SERS operator.A total of 21 SERS spectra were measured from each specimen at different locations.These newly obtained SERS spectra were used as input in the previously trained SFNet models to predict the infection status of each specimen.A ratio  ( ) is defined to classify the SARS-CoV-2 infection status and the subsequence Ct value of the positive specimen is predicted by the SFNet regression model using the average of the 21 spectra.Here n + (n − ) is the number of positive (negative) predictions among the 21 spectra obtained from one specimen.A threshold of  = 0.7 is used, which means that the specimen is SARS-CoV-2 negative when  ≥ 0.7 and positive when  < 0.7.This threshold value is based on Figure 3C, where the SFNet model exhibited a false-negative prediction rate of 2.9% and a false-positive prediction rate of 0%, and can be varied based on future training datasets.The obtained classification results from the SERS spectra were then compared to the RT-PCR-determined status of each corresponding blind specimen.Table S7 (Supporting Information) summarizes the original data, while Figure 6A plots the ratio  against the specimen number listed in Table S7 (Supporting Information).Impressively, all negative specimens were accurately predicted with a 100% success rate.Among the positive specimens, all were correctly identified, except for Positive-3.This specimen has n -and n + values of 15 and 6, respectively, with  (= 0.71) slightly surpassing 0.7, indicating that Positive-3 was incorrectly predicted as negative.This discrepancy might be attributed to a positive specimen with a high Ct value (low virus concentration).By expanding the training spectral set and adjusting the  threshold value (say 0.75), improved prediction performance can be expected.In total, the constructed SFNet model achieved an overall classification accuracy of 99.04%, with a 98.28% accuracy for positive specimens and a perfect 100% accuracy for negative specimens.Subsequently, the SFNet regression model used the averaged 21 spectra per positive specimen for Ct value prediction, as shown in Figure 6B.The prediction plots show an R 2 of ≈0.94-0.95 and RMSEs ≈1.1-1.3,demonstrating commendable quantification performance.Compared to Figure 5D, Figure 6B shows a relatively better result because the spectra used for the regression in blind test are the average of the 21 measured spectra.
Unambiguously, the SERS-DLA direct detection strategy has comparable results to RT-PCR.In comparison to other reported methods, this approach has superior accuracies.For example, rapid antigen detection achieves ≈66% accuracy, [49] breath detection via gas chromatography-mass spectrometry reaches 91.2% for positive and 99.3% for negative specimens, [48,50] nucleic acid tests by various commercial products attain ≈95% accuracy, [51] and dipstick detection using a Palm Germ-Radar achieves 97.2% accuracy. [52]Remarkably, the entire detection process takes just 15 min.These findings indicate the potential of the AgNR@SiO 2 array SERS-DLA as a rapid and promising point-of-care COVID-19 diagnostic platform.

Conclusion
This study heralds a paradigm shift in rapid SARS-CoV-2 diagnostics by synergistically integrating SERS with DLAs encapsulated in the SFNet framework.The entire process-from specimen deactivation and preparation to SERS measurement and SFNet classification-requires <15 min.Utilizing a customized SFNet model, classification accuracies reached 97.1% for positive and 100% for negative specimens.Importantly, the model lends itself to interpretability; SFNet-selected features correlate closely with the molecular signatures of proteins, lipids, and other crit-ical biochemical entities within the virus.Moreover, the SFNet model precisely predicts RT-PCR Ct values, demonstrating dual capabilities in classification and quantification based solely on label-free SERS spectral variances.In a blind test, the methodology achieved a diagnostic accuracy of 99.04%, reinforcing its reliability and practicality as a point-of-care diagnostic tool.
The scalability and adaptability of this technology could potentially advance field diagnostics and contribute substantially to signal processing for SERS-based sensors.
AgNR@SiO 2 Arrays Fabrication: AgNR@SiO 2 SERS substrates were prepared by the oblique angle deposition (OAD) and salinization via hydrolysis of TEOS as described previously. [3,25,32,53]The AgNR substrates were first prepared using OAD.Piranha solution cleaned glass slides (0.5 inches × 0.5 inch) was mounted in a custom-designed electron beam deposition system.A layer of 20 nm Ti film and a layer of 100 nm Ag film were subsequently deposited at a rate of 0.2 and 0.3 nm s −1 , respectively.Then, the vapor incident angle was adjusted to be 86°, and a thickness of 2000 nm Ag film was deposited at a rate of 0.3 nm s −1 to form the AgNRs on the substrates.The entire evaporation process was conducted under a high vacuum condition (chamber pressure <3 × 10 −6 Torr).After the deposition, the AgNR substrates were immersed into a homogeneous mixture of 30 mL of EtOH, 4 mL of H 2 O, and 500 μL of TEOS for 20 min under stirring.The coating of SiO 2 was initiated after adding 560 μL of ammonium hydroxide.The substrates were removed from the reaction solution after 5 min, followed by water rinsing and N 2 drying.A 2-nm conformal SiO 2 coating on AgNR was expected under such conditions. [32]Subsequently, arrayed small wells (4 wells, with a well diameter of 4 mm and a well depth of 1 mm) on a PDMS layer were molded on the AgNR@SiO 2 array to restrict the effective sensing areas, [54] and it is referred them as AgNR@SiO 2 wells.A typical scanning electron microscopy (SEM) image of the AgNR@SiO 2 array is shown in Figure S1A (Supporting Information).
Patient HNS Specimens: Deidentified HNS specimens were obtained from the University of Georgia Veterinary Diagnostic Laboratories (GVDL) for this study, and ethics committee approval and informed written consent of all participants were not required since the specimens used were leftovers from unidentified patients.This was not considered human subject research in accordance with the Declaration of Helsinki. [55]These specimens were residual samples from the Clinical Laboratory Improvement Amendments (CLIA)-registered GVDL's confirmatory RT-PCR diagnostic testing.HNS specimens were collected using a sterile swab applicator and placed in 1 mL of saline.The GVDL determined the SARS-CoV-2 status of each HNS specimen using an Applied Biosystems Taq-Path COVID-19 Combo kit EUA assay (ThermoFisher catalog number A47814, Waltham, MA, USA) in a multiplex RT-PCR format.The multiplex RT-PCR assay had 3 target gene fragments: spike (S), nucleocapsid (N), and Orf1ab (ORF1ab) protein regions, which exhibit high specificity and low risk for mutation (except for the S gene).The RT-PCR data were analyzed and then interpreted by the Applied Biosystems COVID-19 Interpretive Software.For the positive specimens, the corresponding Ct values for three viral gene fragments were recorded.The SARS-CoV-2 viruses in the HNS specimens were inactivated.All the experiments were carried out in a biosafety level 2 (BSL-2) lab.
SERS Measurements: A 120 SARS-CoV-2 positive and 120 negative specimens were used for SERS spectra collection following the procedure: 30 μL of an HNS specimen was mixed in a 1:1 (v:v) ratio with the inactivation buffer containing 1 m guanidine hydrochloride, 0.2% Triton X-100, 1 mm EDTA, and 2 m Tris-HCl with a pH 7.8, followed by room temperature incubation for 5 min.Then, the mixture (10 μL) was diluted by 300 μL pure water without further processing.For SERS measurements, 20 μL of the inactivated specimen was dispensed onto a AgNR@SiO 2 well and was incubated for 5 min.Then the well was washed with DI water (X3) and airdried at 20 °C (the drying time varied from 2 to 5 min).The SERS spectra were acquired by using a Tec5USA Raman spectroscopy (Tec5USA Inc.), with a 785 nm excitation laser with a beam diameter of ≈100 μm, a power of 35 mW, and an acquisition time of 4 s.20 SERS spectra were collected from randomly selected locations in each well.
Statistical Analysis: All the SERS spectra were pre-processed following a procedure that included de-spiking, baseline removal, and area normalization.The detailed pre-processing of the spectra is shown in Section S2 (Supporting Information).Based on the overall spectral features of the SERS spectra obtained, a "Gaussian-Lorentzian function fitting (GLFF)" baseline removal method was applied. [3,56]This process guarantees a minimum disturbance to the raw data and avoids unnecessary information loss due to spectra pre-processing.Regarding the sample size, 120 SARS-CoV-2 positive and 120 negative clinic specimens were used for SERS spectra collection for the training and validation of different machine learning models.Additional 104 deidentified clinic HNS specimens (46 negatives and 58 positives) were used for a blind test.For data dimension reduction and visualization, principal component analysis (PCA) and t-distributed stochastic neighbor embedding (tSNE) were implemented via scikit-learn version 1.0.2. [57]achine Learning and Deep Learning Algorithms for Classification and Regression: Five different algorithms, including SVM, RF, multilayer perceptron (MLP), convolutional neural network (CNN), and SFNet, were applied to classify the patient HNS specimens based on the SERS spectra.The total spectral set consisted of 2400 SERS spectra collected from 120 positive and 2400 SERS spectra from 120 negative specimens.All the original SERS spectra were preprocessed following a procedure described in Section S2 (Supporting Information), [56] which consists of a baseline removal and a spectrum normalization.The entire spectral set was shuffled randomly to make the spectral set have both positive and negative specimen data in each batch.The entire spectral set was split into 70%: 15%: 15% for training set, validation set, and test set, respectively.The SFNet model had one convolutional layer (the size of the convolutional kernel was 5 × 1 and the number of filters was 32), one pool layer (Max Pool), three consecutive blocks, i.e., one convolutional block (Conv block), and two identities (ID) blocks (the convolutional kernel size of the two blocks were 5 × 1 and 7 × 1, respectively, and the number of filters was 32 and 64, respectively), two long-short term memory (LSTM) layers with 982 and 300 units, and one fully connected layer.The learning rate of the Adam optimization algorithm was 0.03.The other four classification algorithms, SVM, RF, MLP, CNN, are detailed in Section S4 (Supporting Information).
The SFNet regression model consisted of two LSTM layers, two dropout layers, and three fully connected layers.The sizes of the LSTM layers were 700 and 300, with dropout sizes of 0.5 and 0.3, and the sizes of the fully connected layers were 500, 200, and 100, with L1 loss used as the loss function, Adam as the optimizer, and a learning rate of 0.001.The other regression algorithms based on SVM, RF, MLP, and CNN are detailed in Section S6 (Supporting Information).
SVM and RF analyses were performed in MATLAB software, using the GridSearch function for parameter optimization.The MLP, CNN, and SFNet models were trained using the TensorFlow 2.4 environment in Py-Charm Community software.For these models, the batch size was set at 20, and the number of iterations for training instances (spectra) was set at 207.All processing was conducted on a desktop computer equipped with an Intel i7-13700KF CPU @ 3.4 GHz, with 64 GB RAM, and an NVIDIA GeForce RTX 4080 GPU.

Figure 1 .
Figure 1.Schematic illustration of direct quantitative detection from HNS specimens using SERS and DLAs.

Figure 2 .
Figure 2. A) The averaged SERS spectra for individual HNS specimens: positive, negative, and inactivation buffer (not averaged).B) The SERS peak alignments of different reagents in the inactivation buffer.C) The PCA score plot and D) the tSNE plot of the clustering results based on SERS spectra of HNS specimens and inactivation buffer.

Figure 3 .
Figure 3. A) The architecture of the SFNet model.B) The loss and accuracy curves for the test spectral set.C) The confusion matrix and D) the corresponding ROC curve of the SFNet classification results of the test spectral set.E) A summary of the training and testing accuracies of different MLA and DLA models for SERS spectrum classification.

Figure 4 .
Figure 4. A) The FIM of the SFNet model to differentiate negative and positive spectra.The black dotted line marks the threshold of 50%, and the purple segmented lines under the FIM indicate the R SFNet .B) The obtained R K from Raman spectra of lipid, amide I, amide III, RNA, tyrosine, and phenylalanine and the corresponding matching scores.

Figure 5 .
Figure 5. A) The architecture of the SFNet regression model.B) The loss curves for the training spectral and test spectral sets.C) The plots of R 2 and RMSE for different regression models.D) The plots of the predicted Ct value (Ct pre ) versus the actual Ct values (Ct act ) of the ORF1ab gene, N gene, and S gene for SFNet, MLP, and CNN regression models.
Figure S9 (Supporting Information) shows near-identical Ct values for these three genes.In some instances, mutations in the spike protein made the S gene's Ct value unavailable, and the missing Ct value was substituted with the average of ORF1ab and N gene Ct values.The architecture of the proposed SFNet regression model is shown in Figure 5A.The model takes SERS spectra from specimens as input, and outputs Ct values for three viral gene fragments: ORF1ab, N gene, and S gene.Thus, a given SERS spectrum, the output is a three-element vector denoted as [C 1 , C 2 ,C 3 ].This SFNet regression model consists of two LSTM layers, two dropout layers, and three fully connected layers.The sizes of the LSTM layers are 700 and 300, with dropout sizes of 0.5 and 0.3, respectively; and the sizes of the fully connected layers are 500, 200, and 100.The model utilizes L1 loss as the loss function, Adam as the optimizer, and a learning rate of 0.001.Refer to Table S5 (Supporting Information) for details.The spectral set was divided into 70% for training and 30% for testing.Employing optimized hyperparameters, the loss functions of the training and test spectral set were recorded in Figure 5B.The training set's loss function decreased sharply from 17.3 to 3.6 over the first 2000 epochs, while the test set's loss function exhibited fluctuations.At ≈10 000 epochs, the loss functions of the training and test sets stabilized ≈0.6 and 1.1, respectively, indicating a gradual convergence of training.

Figure 6 .
Figure 6.A) The plot of ratio  based on the results from the SFNet model against specimen number (arranged particular).B) The plots of the predicted Ct pre versus the actual Ct act of ORF1ab gene, N gene, and S gene for the blind test.The dash line in each plot indicates Ct act = Ct pre .