Micro-Fourier-transform infrared reflectance spectroscopy as tool for probing IgG glycosylation in COVID-19 patients

It has been reported that patients diagnosed with COVID-19 become critically ill primarily around the time of activation of the adaptive immune response. However the role of antibodies in the worsening of disease is not obvious. Higher titers of anti-spike immunoglobulin IgG1 associated with low fucosylation of the antibody Fc tail have been associated to excessive inflammatory response. In contrast it has been also reported that NP-, S-, RBD- specific IgA, IgG, and IgM are not associated with SARS-CoV-2 viral load, indicating that there is no obvious correlation between antibody response and viral antigen detection. In the present work the micro-Fourier-transform infrared reflectance spectroscopy (micro-FTIR) was employed to investigate blood serum samples of healthy and COVID-19-ill (mild or oligosymptomatic) individuals (82 healthcare workers volunteers in “Instituto de Infectologia Emilio Ribas”, São Paulo, Brazil). The molecular-level-sensitive, multiplexing quantitative and qualitative FTIR data probed on 1 µL of dried biofluid was compared to signal-to-cutoff index of chemiluminescent immunoassays CLIA and ELISA (IgG antibodies against SARS-CoV-2). Our main result indicated that 1702–1785 \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {cm}^{-1}$$\end{document}cm-1 spectral window (carbonyl C=O vibration) is a spectral marker of the degree of IgG glycosylation, allowing to probe distinctive sub-populations of COVID-19 patients, depending on their degree of severity. The specificity was 87.5 % while the detection rate of true positive was 100%. The computed area under the receiver operating curve was equivalent to CLIA, ELISA and other ATR-FTIR methods (\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$>0.85$$\end{document}>0.85). In summary, overall discrimination of healthy and COVID-19 individuals and severity prediction as well could be potentially implemented using micro-FTIR reflectance spectroscopy on blood serum samples. Considering the minimal and reagent-free sample preparation procedures combined to fast (few minutes) outcome of FTIR we can state that this technology is suitable for fast screening of immune response of individuals with COVID-19. It would be an important tool in prospective studies, helping investigate the physiology of the asymptomatic, oligosymptomatic, or severe individuals and measure the extension of infection dissemination in patients.


Methods
Sampled population. Serum samples from 82 healthcare workers volunteers of the "Instituto de Infectologia Emilio Ribas", São Paulo, Brazil were included in this study. These samples are part of a cohort study to evaluate the seroprevalence of SARS-CoV-2 infection in healthcare population. Samples were collected from July to November of 2020, before the beginning of the vaccination program in Brazil. Levels of illness were mild or oligosymptomatic. This study was conducted in accordance with the Declaration of Helsinki, and the protocol was approved by Research Ethics Committee of the "Instituto de Infectologia Emílio Ribas", São Paulo, Brazil, protocol number CAAE 32264120. 5.2001.0061. Invited volunteers were informed about the objectives, propositions and conditions of this project, and those who agreed to participate in the research signed the free and informed consent term. A volume of 5 mL of whole blood was collected by peripheral vein puncture and stored in a 10 mL dry tube for each patient. Then it was centrifuged at 2052 g for 10 min to separate the serum. Two mL of serum were stored in a −20 • C freezer until performing the immunoassays and FTIR tests.

Detection of antibodies anti-SARS-CoV-2. Chemiluminescent immunoassay (CLIA). A CLIA (IgG
Antibodies against SARS-CoV-2; reagents pack # 619 9919; VITROS Immunodiagnostic calibrator # 619 992; Ortho Clinical Diagnostics, Raritan, NJ, USA) immunoassay was performed to detect immunoglobulin G anti-Spike protein from SARS-CoV-2. This assay does not differentiate binding IgG antibodies from virus-neutralizing IgG antibodies 15 . The results were expressed in terms of ratio of the sample signal to a calibrator-assigned www.nature.com/scientificreports/ cutoff signal with threshold of 1.0. Negative diagnosis were considered for outcomes with reactivity index < 1 otherwise they were considered positive 16 . The test sensitivity was reported to be 90.0% ( ≥ 8 days) and specificity of 100.0% 16 . This CLIA diagnostic test uses S antigens for SARS-CoV-2 detection 17 .
Enzyme-linked immunosorbent assay (ELISA). Anti-SARS-CoV-2 IgG ELISA (Euroimmun Medizinische Labordiagnostika, Lübeck, Germany; # EI 2606-9601 G; Indirect ELISA) assay was performed to detect IgG antibodies against SARS-CoV-2. In this diagnostic test IgG antibodies against SARS-CoV-2 spike protein subunit 1 (S1) are detected in human serum or plasma. Following the instructions of the manufactures, serum samples were 1 : 101 diluted, added to wells coated with recombinant SARS-CoV-2 antigen and incubated for 60 min at 37 • C. Then, wells were washed three times and followed by the addition of HRP-conjugated anti-human IgG and subsequent incubation for 30 min at 37 • C. Wells were washed three times and a chromogen solution was added. After 30 min of incubation at room temperature, the reaction was stopped and the absorbance at 450 nm with reference at 620 nm was read on a microplate reader. A ratio between the extinction of the sample and calibrator on each plate were calculated. According to the recommendations of the manufacturer, a signal-to-cutoff ratio smaller than 0.8 is considered negative, while a positive one if greater than 1.1. The borderline region falls into 0.8-1.1 interval. The sensitivity and specificity of this assay were reported to be 90.0% and 100%, respectively 17 .
Fourier-transform infrared absorption (FTIR) reflectance spectroscopy. All samples were brought to room temperature prior to preparation for micro-FTIR measurements. Aliquots of 1 µL of serum (1 : 3 in ultrapure water) were deposited in platinum sample holder and dried at 80% relative moisture under a desiccator with NaCl saturated solution in order to avoid coffee ring effect and obtain a homogeneous biofilm 18 . The final droplet presented an average radius of 1000 ± 150 µm. A Varian 610 FT-IR micro-spectrometer equipped with a linearized MCT detector with a 100 × 100 µm detector element (InfraRed Associates, Inc.) coupled to a 640-IR FT-IR spectrometer was used in reflectance mode for spectra acquisition. After optical focusing, the microscope aperture in the Cassegrain collecting lens ( 60× magnification) was reduced to 150 × 150 µm 2 to set the area of interest for measurement on the sample. The spectral resolution was set to 2 cm −1 in order to obtain the maximum throughput in the IR microspectrometer. The number of scans was 32 per sample (equivalent to 30 s time of acquisition). The scheme of methodology is shown on Fig. 1.

Statistical analysis. Principal components analysis (PCA). The classical Principal Components Analysis
(PCA) 19 was performed on mean centered raw data to extract outliers and identify possible experimental bias. All spectral analysis steps were implemented in the ChemSpec vignette available in the software R 20 . Outliers were identified using the Q and T 2 Hotelling's statistics. The Q statistics indicates how well each observation matches to the PCA model and the Q-residuals measure the residual between a sample and its projection on the factors retained in the model. Large residual outliers can be detected by inspection of Q-residuals. On the other Figure 1. Scheme for micro-FTIR reflectance measurements for human serum. One µL of diluted (1 : 3 in ultra-pure water) serum sample solution (1) was transferred to a circular platinum sample holder (2). Then the sample holder was installed in a desiccator with saturated solution of NaCl (3) which controls the relative moisture in 80%. The drying time was 10 min. After this period the sample holder was installed on the FTIR reflectance accessory of the micro-FTIR spectrometer. The IR beam (4) passing through the IR 60× magnification Cassegrain lens (5) focuses the light on a given sample (6). The reflected light is collected by the same lens and analyzed by the spectrometer. www.nature.com/scientificreports/ hand, Hotelling's T 2 value represents a measure of the variation in each sample within the model, indicating how far each sample is from the center (scores = 0 ) of the model. It is a quantifier for scores outliers. The T 2 Hotelling's versus Q-residuals (reduced) plot were inspected raw spectral data.
Partial least squares: discriminant analysis (PLS-DA). All spectra were pre-processed to become comparable for statistical analysis. The baseline was corrected using the least-squares polynomial curve fitting method as described by Lieber and Mahadevan-Jansen 21 . All spectra were normalized and scaled using probabilistic quotient normalization 22 . Then PLS-DA analysis was performed. It is a multivariate supervised method that uses linear regression of original variables to predict the class membership. In our case the PLS regression was performed using the plsr function provided by R pls package 20,23 . The classification and cross-validation were performed using the corresponding wrapper function using the caret package 20 . A permutation test was performed to assess the performance of class discrimination. In each permutation, a PLS-DA model was built between the data and the permuted class labels using the optimal number of components determined by leave-one-out cross validation for the model based on the original class assignment. The class discrimination performance was measured using classification accuracy, R 2 , and Q 2 parameters. The first one is based on prediction accuracy. The R 2 parameter is the "goodness of fit" or explained variation which is based on the ratio of the between group sum of the squares and the within group sum of squares. On the other hand, Q 2 is the "goodness of prediction", or predicted variation, calculated from cross validation. In each round, the predicted data are compared with the original data, and the sum of squared errors is calculated and then summed over all samples (Predicted Residual Sum of Squares or PRESS). For convenience, the PRESS is divided by the initial sum of squares and subtracted by 1 to scale to R 2 . Good predictions will have low PRESS or high Q 2 while negative Q 2 means the model is not predictive at all or is overfitted 8,24,25 . Two quantifiers were used to measure the vibrational band frequency importance in PLS-DA model. The first, Variable Importance in Projection (VIP) is a weighted sum of squares of the PLS loadings taking into account the amount of explained spectral intensity-variation in each dimension. The other importance measure is based on the weighted sum of PLS regression. The weights are a function of the reduction of the sums of squares across the number of PLS components. For multiple-group analysis, the same number of predictors will be built for each group and the average of the feature coefficients were used to indicate the overall coefficient-based importance. The receiver operating characteristic (ROC) analysis was used to evaluate the discriminating performance and the area under the ROC curve (AUC ) as its summary index. In general tests with excellent discriminating capability will furnish AUC > 0.80 26,27 .
Normality and F tests. The Kolmogorov-Smirnov test of normality 28 was applied to test the hypothesis that IgG data does not differ significantly from that which is normally distributed. The F-test was used to test the hypothesis of equality of averages of two sets of data populations of unequal size (IgG and demographic data) 29 .

Results and discussion
The main clinical and demographic pieces of information about volunteers are presented on Table 1. Percentages of 40.2% ( n = 33 ) and 59.8% ( n = 49 ) of individuals tested positive and negative, respectively. Female individuals predominated in the sampled groups ( > 72 % of total number of patients). However we notice that this unbalance did not induce a bias on the positive or negative rates within the confidence value of 5%. The average age of patients in positive and negative groups was 44 ± 14 years and 49 ± 12 years, respectively. Again these differences were not statistically relevant for COVID-19 diagnostic purposes. Likewise the body mass index did not represent a statistically relevant variable comparing negative and positive groups. On the other hand, the presence of comorbidities presented a well-defined increased risk for positive test against COVID-19. A percentage of 39.4% of individuals which tested positive presented some kind of comorbidity (rheumatoid arthritis, asthma, diabetes, systemic arterial hypertension, chronic obstructive pulmonary disease, obesity, or hypothyroidism) while this rate is only 16.3% in negative group. The outcome of F-test indicated that there is enough evidence against the hypothesis that the population sampled with positive and negative test has the same average ( p = 45.4%). The correlation between comorbidities and prevalence of COVID-19 is well reported in literature and our findings give one more piece of evidence concerning this important aspect of etiology of COVID-19 30 . We notice that CLIA and ELISA tests were 100% concordant about the patient's diagnosis in our cohort. Average FTIR spectra for negative (black line) and positive (red line) classes of samples in the fingerprint spectral window (880-1800 cm −1 ) are shown on Fig. 2. Assignments for the main vibrational bands (vertical Table 1. Demographic data of volunteers, a set of 82 healthcare workers from "Instituto de Infectologia Emilio Ribas", São Paulo, Brazil. The list of comorbidities includes rheumatoid arthritis, asthma, diabetes, systemic arterial hypertension, chronic obstructive pulmonary disease, obesity, and hypothyroidism. www.nature.com/scientificreports/ lines in Fig. 2a) are presented on Table 2. Proteins, amino acids, and nucleic acids vibrational bands dominate the spectra in accordance with the average chemical composition of the blood serum 31 . Prior data processing and multivariate statistical analysis, a quality check evaluation was performed on raw spectral data to identify anomalous spectra, outliers and/or biased patterns. The PCA was calculated on meancentered raw data and then Q residuals (reduced) versus T 2 -Hotelling's plot 8,25 was checked (Fig. 2b) in order to find residuals and scores outliers. Data outside of the confidence limits of 97% for scores and 3% for residuals were considered outliers (indicated by "*" in Fig. 2b) and removed in further analysis. PLS-DA analysis was then performed on processed, normalized, and scaled spectra. The optimal number of factors was determined by cross-validation after inspection of accuracy, R 2 , and Q 2 . Figure 3 presents the pairwise scores up to 5th PC (4.7% of explained variance) for positive and negative groups. At first glance, combinations including components 1 and 2 are prone to discriminate betwenn positive and negative samples.
The best performance of PLS-DA classification was observed considering 2 components (Fig. 4a). In this case the observed accuracy on groups discrimination was 76% while R 2 = 0.39 and Q 2 = 0.34 (Fig. 4a). Regression coefficients are shown in Fig. 4b). Both curves for discriminating positive and negative groups appeared smooth, showing no random fluctuations around positive and negative values, which would be a symptom of overfitting. However, the coefficients for positive and negative classes appeared superimposed in many spectral windows, indicating greater similarity among them. The calculated response is shown in Fig. 4d). Sensibility and specificity were 53.1% and 87.5%, respectively in this case. Interestingly, there is a distinct subgroup of misclassified samples (labeled here as "mix" group). This subgroup showed a clear spectral signature, as seen in the heatmap in Fig. 5.
The performance of the PLS-DA model with 3 possible groups (positive, negative, and mix) is shown in Fig. 4c). R 2 and Q 2 presented a consistent increase of ∼ 0.1 while the accuracy was around 76%. The coefficient and response obtained are indicated in green in Fig. 4b,d, respectively. Specificity was 87.5% while the detection rate of true positive and true mix increased to 100%. However, one outstanding result was observed contrasting the IgG Signal-to-Cutoff index from CLIA of each positive, mix, and negative groups to FTIR PLS-DA www.nature.com/scientificreports/ classification data. Histograms in Fig. 6a-c summarize this finding. The positive group presented a broad IgG reactivity index distribution between 1 and 17 while the mix group presented reactivity ranging from 1 to 8. Obviously, the negative group presented a narrow histogram centered on 0.1 value. The outcome from the Z-test ( p < 5 %) indicated that these IgG distributions come from distinct populations. Interestingly the histograms of signal-to-cutoff from ELISA (Fig. 6e,f,g) showed no evidence that positive and mixing groups belongs to distinct populations ( p > 5 % for Z test). Thus, the question that arises concerns the characteristics that gives rise to these distinct populations. Important vibrational bands that contribute to discrimination can help find the answer to this question. The important vibrational frequencies (VIP) identified by PLS-DA are represented in Fig. 6d,h. For positive and negative discrimination ( Fig. 6h and Table 2), β-sheet structure of Amide I of proteins (1689-1698 cm −1 ) and deoxyribose from DNA bands (840-856 cm −1 ) contributed to the greatest over-expression in the positive group. Immunoglobulins are heterodimeric proteins composed of two heavy (H) and two light (L) chains. They can be functionally separated into variable (V) domains that bind antigens and constant (C) domains that drives its functions, such as complement activation or binding to Fc receptors 38 . The basic H2L3 structures consist of the Fc region (fragment, crystallizable) and Fab region (fragment, antigen binding), both composed mainly of β-pleated sheets 38 .Thus the increased intensity of β-sheet band in the positive group is completely correlated with antibody recruitment in COVID-19 disease. Likewise, the increase in cellular activity due to Sars-Cov-2 inflammatory response explains the observed increase in DNA band intensity. However, for 3 groups (positive/ mix/negative) discrimination, a very narrow set of vibrations confined to 1702-1785 cm −1 spectral window is the most important (Fig. 6d). The 1735-1785 cm −1 region had been reported as spectral marker for lipids, C = O cholesteryl esters, and triglycerides 35 . Besides, the C = O bond for the methyl-esterified carbonyl groups also presents a relatively strong absorption peak at 1730-1760 cm −136 . Also it had been reported that the FTIR spectrum of glycated human serum albumin presents a new peak carbonyl group at 1737 cm −134 . Also 1739 cm −1 and 1781 cm −1 bands were reported as predictors for classifying anti-neutrophil cytoplasmic antibodies in sera samples 39 .
In our case bands in 1702-1785 cm −1 spectral window could be grouped in two classes: (1) those overexpressed in the mix group and subexpressed in the positive group which are both related to IgG glycosylation and (2) those overexpressed in the positive group and underexpressed in the mix group which are both related to thymine. As long as we can argue that the mix group relates to a COVID-19 sub-population with highly glycoyilated IgG. As all patients in the present study were mild or oligosymptomatic, the above feature is in agreement with the Chakraborty et al. findings 3 which proposed that patients with severe COVID-19 are more likely to have IgG1 with afucosylated Fc glycans as signature. It is well reported that IgG glycosylation can determine whether an antibody glycoform is pro-inflammatory (such as IgG with galactose-deficient N-glycans) or anti-inflammatory (such as IgG with sialylate N-glycans) 37 . Since 1768-1786 cm −1 region is related to methyl-esterified carbonyl vibration we argue that this band is related to C=O on sialic acid on IgG with sialylate N-glycans. One important point to Table 2. Assignments for the main vibrational bands of micro-FTIR blood serum (Fig. 2a). www.nature.com/scientificreports/ clarify is the difference between CLIA and ELISA distribution of Signal-to-cutoff ratio (Fig. 6). It appears that the fluorescence yield of murine monoclonal anti-human IgG anti-body complexed to HRP of CLIA is dependent on the fucosylation of the Fc tail of human IgG. This point needs to be explored in additional experiments.
Concerning the diagnostic performance we mention that spectral windows at 820-890, 1025-1180, and 1685-1727 cm −1 presented excellent capability for discriminate negative and positive classes (AUC = 0.80-0.86, Fig. 7a). The positive and mix classes were also discriminated in excellent grade (AUC = 0.80-0.98) in spectral windows 850-940, 1015-1170, 1500-1570, 1612-1652, 1695-1735, 1755-1790 cm −1 (Fig. 7b). This curve is comparatively nosiest respect to the positive/negative case due to the low-N (only 30% of data) in this case. The representative ROC curves for negative and positive (Fig. 7c,d) and mixing and positive (Fig. 7e,f) classes with highest AUC are also shown. The ROC curves for positive/negative (Fig. 7c,d) have binormal curve shapes while mixing/positive ones have straight line shapes due to low-N. The classification boxplots for two discriminating classes presented a significant statistical difference among groups ( p < 0.05 for t-Student test).
Micro-FTIR reflectance compared to other methods. At this point it is important to compare our results to previous reports concerning blood testing for COVID-19. The usage of ATR-FTIR for analysis of www.nature.com/scientificreports/ plasma from COVID-19 patients have been recently reported in literature. This technique shares a great similarity with our methodology and for this reason a detailed comparison is important. Banerjee et al. 40 investigated a cohort of 160 clinicopathologically confirmed SARS-CoV-2 patients using ATR-FTIR. They proposed a plasma processing with 75% ethanol v/v for virus inactivation followed by vortexing and drying over ATR-crystal. The reported AUC was 0.851 when considering the set of spectral and clinical (age, sex, diabetes mellitus, and hypertension) data. The discriminating performance of the spectral data   Then only a single patient could be tested at once. This process demands usually > 5 h 35 . It is not well reported whether the proteins degradation over this time would impact on quality of diagnosis. Moreover the morphological characteristics of the deposited bio-film (heterogeneity, size, wettability, dilution, among others) is dependent of drying conditions as relative moisture and temperature. The quality of spectrum (artifacts, reproducibility) is very sensitive to the morphology of the bio-film. These points impose challenges to the standardization of the method 35 . On the other hand, our approach enables the preparation of drops of several patients over the substrate. Over an equivalent area of a microplate of 96 wells it is possible the deposition of 500 droplets. Using the microscope re-positioning the IR beam over each patient drop is relatively easy which increases the scale of www.nature.com/scientificreports/ testing. Banerjee et al. reported the fatality prediction only when considering in the statistical analysis a larger set of data of different kinds as spectra and clinical pieces of information which in some aspect would represent a bias in the statistical test. We were able to discriminate 3 possible groups in an independent way of validation.
The clinical data were used to compare and discuss the data but not as input.
Other comparison is with the gold standard method for COVID-19 testing ELISA and also with CLIA method. Both were used to validate our results. Jagtap et al. 42 evaluated the performance of spike protein antigens for SARS-CoV-2 serology compared four spike proteins: RBD, S1, S2 and a stabilized spike trimer (ST). They used indirect ELISA in serum from COVID-19 patients and pre-2020 samples. ROC curves indicate that ST is the best candidate for serological testing with the highest AUC ( AUC = 0.94 ). AUC for IgG was 0.934. Huang et al. 43 compared the serological test of SARS-CoV-2 using ELISA test and reported that AUC(IgM) = 0.812 while AUC(IgG + IgM) = 0.983 . The reported AUC for CLIA anti-Sars-CoV-2 test include 0.846 44 , 0.9143 45 , 0.98 46 , and 0.831 44 . Thus we can conclude that the performance of our micro-FTIR approach is in the same level of ELISA and CLIA.
However as summarized by Liu and Rusling 47 , in general the nucleic acid-based tests against COVID-19 have reported high false-negative rates which are dependent of the day of the symptom (38% on the day, 20% 3 days, and 66% 16 days, after of symptom onset, respectively). Factors such as sampling time, viral mutation, inadequate handling, improper storage, transportation of samples, among others had been cited to influence the result. The requirements to perform these tests evolve high workload, skilled personnel for testing and sample collection, special reagent kits, costly centralized infrastructure with specific equipment for process and measured the sample signal, and professional bio safety level (BSL)-2 lab 47 imposing elevated costs, relatively longer time to delivery of results (2-3 h) and dependence of international chain of suppliers. This last aspect would be disrupted during the pandemic period. One important advantage of the micro-FTIR method is the minimal need of sample preparation which impacts lower cost per testing and relative independence of external suppliers. The time elapsed pipetting , drying, acquiring spectra is typically 10 min. Another aspect to mention is that ELISA and CLIA tests are able to detect a single protein at once. Nevertheless FTIR has multiplexing advantage enabling comparison of several metabolites which increases the quality of prognosis being a suitable tool for precision medicine 48 . Our method enables in principle to probe 500 samples per row while ELISA and CLIA usually are limited to 96 samples. Table 3 summarizes the comparison between all techniques.

Conclusions
Our results showed that micro-FTIR was able to probe key structural aspects of serum IgG from COVID-19 volunteers. The 1702-1785 cm −1 spectral window is a spectral marker of the degree of IgG glycosylation, allowing to probe distinctive sub-populations of COVID-19 patients, depending on their degree of severity. Furthermore, the β-sheet structure of Amide I of proteins (1689-1698 cm −1 ) and deoxyribose from DNA (840-856 cm −1 ) bands presented most significant contributions to positive and negative discrimination with a specificity of 87.5% and sensibility of 100%. The computed AUC was comparable to the ELISA, CLIA, and other ATR-FTIR methods ( AUC > 0.85 ). In summary, overall discrimination of healthy and COVID-19 individuals and severity prediction could also be potentially implemented using micro-FTIR reflectance spectroscopy on blood serum samples due to direct probe of glycosylation degree of IgG. However, experimental efforts need be devoted to isolate immunoglobulins molecules from plasma samples from healthy and COVID-19-ill patients and characterize their spectral properties what would give detailed clues on IgG glycosilation and other structural changes related to pathology.  Table 3. Parameters for maintenance, management, operating, and performance of the main anti-Sars-Cov-2 clinical tests ELISA and CLIA compared to the ATR-FTIR and micro-FTIR.