Peptide arrays of three collections of human sera from patients infected with mosquito-borne viruses

Peptide preparation and microarray printing

A subset of the previously predicted diagnostic peptides²², representing multiple mosquito-borne virus species and subtypes, were synthesized at the Center for Protein and Nucleic Acid Research at The Scripps Research Institute (TSRI)^23,24. This selected collection of peptides consisted of surface-exposed 15-mers with sequences that represented the consensus amino acid sequence among strains belonging to each of our six target taxa including: CHIKV, DENV1, DENV2, DENV3, WNV, and ZIKV. Peptides on the array that represented mosquito-borne virus taxa for which there were no serum samples were ignored in downstream quantification and computation. As such, a total of 25, 51, 28, 34, or 70 peptides in the E protein as well as 15, 19, 15, 23, or 70 peptides in the NS1 protein (all derived from DENV1, DENV2, DENV3, WNV, or ZIKV sequences, respectively) were evaluated in these experiments. A set of 25 peptides spanning portions of the CHIKV E2 protein that had previously been reported as relevant for detecting anti-CHIKV antibodies were also included²⁵. Synthesized peptides were suspended in 12.5 μL DMSO and 12.5 μL of ultra-pure water. Immediately prior to printing, suspended peptides were diluted 1:4 in a custom protein printing buffer [saline sodium citrate (SSC): 300 mM sodium citrate, pH 8.0, containing 1 M sodium chloride and supplemented with 0.1% Polyvinyl Alcohol (PVA) and 0.05% Tween 20], in a 384-well non-binding polystyrene assay plate. Two positive control peptides, hemagglutinin A (HA) (YPYDVPDYA) and FLAG tag (DYKDDDDK), together with a dye that permanently fluoresces at 488 nm (Alexa Fluor 488) were included in the print to guide proper grid placement and peptide alignment, as well as to serve as printing controls as well as controls to quantify the maximum fluorescence for the assays.

Quadruplicate sets of all peptides were printed onto N-hydroxysuccinimide ester (NHS-ester) coated NEXTERION Slide H (Applied Microarrays) slides at an approximate density of 1 ng/spot, using a Microgrid II (DigilabGlobal) microarray printing robot equipped with solid steel (SMP4, TeleChem) microarray pins. Humidity was maintained at 50% during the printing process. Immediately prior to interrogating the arrays, slides were blocked for 1 h with ethanolamine buffer to quench any unreacted NHS-ester on the slide. All slides were used within 2 months of printing and were stored at -20°C²³.

Serum sources

Spent diagnostic serum samples were provided by collaborators working under three separate clinical studies in Honduras, the United States, and Nicaragua. These sera were collected from a total of 137 consented human patients under IRB supervision and were characterized as positive for antibodies against at least one of: ZIKV, DENV1, DENV2, DENV3, WNV, and/or CHIKV.

A total of 32 deidentified plasma samples from patients suspected of Zika, chikungunya or dengue in Honduras were obtained at the discretion of health care providers at the Hospital Escuela Universitario from patients (ages 6–73 years old). These acute-phase samples were sent to the Centro de Investigaciones Geneticas at the Universidad Nacional Autonoma de Honduras in Tegucigalpa, Honduras for ZIKV, CHIKV and/or DENV molecular testing. Of these patients, 23 had infection with DENV and nine had infection with ZIKV confirmed by RT-qPCR during the acute phase. Convalescent samples were collected from these patients 10–30 days post-onset of symptoms between June 1 to November 30, 2016.

A total of 73 de-identified human serum samples were obtained from the Vanderbilt Vaccine Center Biorepository. Sera from individuals with previous history of natural infection with DENV, WNV, CHIKV, or ZIKV (confirmed by serology for convalescent samples or RT-qPCR for acute-phase samples) while traveling in the Caribbean, Central or South America, or West Africa were included on arrays. For WNV, sera were from individuals with confirmed previous history of natural infection contracted during an outbreak in 2012 in Dallas, TX. The samples were collected in the convalescent phase, months to years after post-onset of symptoms.

A total of 32 de-identified human sera were collected from the Pediatric Dengue Cohort Study (PDCS) in Managua, Nicaragua^26,27. Early convalescent-phase samples were collected 15–17 days post-onset of symptoms from 9 Zika cases that were confirmed as positive for ZIKV infection by real-time RT-qPCR between January and July, 2016. Late convalescent samples were obtained from 21 DENV-positive cohort participants after RT-qPCR confirmed DENV1 (n=7), DENV2 (n=8), or DENV3 (n=6) infection and 2 DENV-negative subjects, all in 2004–2011, prior to the introduction of ZIKV to Nicaragua. Samples were analyzed by inhibition ELISA^28,29 and neutralization assay^30,31. The PDCS was approved by the IRBs of the University of California, Berkeley, and Nicaraguan Ministry of Health. Parents or legal guardians of all subjects provided written informed consent; subjects 6 years old and older provided assent.

High-throughput screening and quantification of characterized patient sera

Once the peptide microarrays were printed, aliquots from a subset of samples were used to optimize the screening and detection processes. Specifically, dilutions ranging from 1:50 to 1:1000 were evaluated to determine the optimal dilution level for subsequent screening. A 1:200 dilution was selected to achieve an optimal balance between the available aliquot volumes and assay sensitivity.

The 137 characterized sera were separately subjected to high-throughput screening using the synthesized peptide arrays. Sera were tested for IgG reactivity using the custom peptide array at TSRI. For immunolabeling, the incubation area around the printed grids was circumscribed using a peroxidase anti-peroxidase (PAP) hydrophobic marker pen (Research Products International Corp) and the subsequent steps were performed in a humidified chamber at room temperature on a rotator. Control anti-HA (mAb 12CA5, Scripps Research, mouse IgG, RRID:AB_514505) and anti-FLAG monoclonal antibodies (Invitrogen, MA1-142-A488, RRID:AB_2610653) were assayed at a concentration of 10 μg/ml while 10 μl of human sera were diluted 1:200 in PBS buffer containing Tween (PBS-T) and incubated for 1 h followed by three washes in PBS buffer. The arrays were then incubated for 1 h with goat anti-human IgG tagged with Alexa Fluor® 488 (Invitrogen, cat. #: A-11013, RRID: AB_2534080) as a secondary antibody. Arrays were washed three times in PBS-T, two times in PBS, and another two times in deionized water and centrifuged to dry at 200 × g for 5 mins.

The fluorescence of the processed slides was quantified using a ProScanArray HT (Perkin Elmer) microarray scanner at 488 nm and 600 nm, and images were saved as high-resolution TIF files. Imagene® 6.1 microarray analysis software (BioDiscovery; ImageJ could be used as an open-access alternative) was used to calculate the fluorescence intensity of the area within the printed diameter of each peptide as well as the fluorescence of the same diameter directly outside of the area occupied by each peptide. The mean and median fluorescence signal and background pixel intensities, as well as other data for each antigen, spot were calculated, digitized, and exported as individual rows in a comma-delimited file for subsequent analysis.

Data processing to identify immunodominant epitopes

A custom script³² was written to implement a previously described array processing workflow²⁴ with a minor change to use the median foreground and background values instead of mean values to minimize outlier effects (available on GitHub). Negative background values were interpreted as zeroes. Briefly, background correction was calculated by subtracting the median background from the median foreground measurements for each spot on each array. Normalization was performed by dividing the background-corrected values for each spot on each slide by the non-control spot having the largest fluorescence value on each slide as has been described previously²⁴. The quadruplicate spots for each peptide on each array were then summarized into a single value by calculating the median value of the quadruplicate spots for each peptide to further reduce the effects of any outliers. The normalized relative fluorescence intensity values for all peptides and all samples were output as a separate file together with summarized quantitative values indicating how well each peptide was recognized by each of the polyclonal serum samples.

A separate script was used to transform all relative fluorescence intensity values for each peptide into Z-scores, and separate tables were constructed to contain the summarized Z-score values for all peptides (as columns) representing each of the viral taxa and all samples (as rows) that were tested with the peptide array. A random forest algorithm (randomForest version 4.6-12 package in R) was applied to each of these tables in order to identify the peptides that were best able to differentiate between each of the viral taxa. In this case, the number of trees generated in the random forest for each species was 100,000, and the number of variables randomly sampled as candidates at each split was equal to the square root of the number of columns present in each table.

The values representing the mean decrease in Gini index were calculated separately for samples obtained from each of the three collections as well as all possible combinations of two or more collections. These data were then used to identify the top 30 peptides according to their usefulness in identifying the correct virus taxon. The BepiPred algorithm was then used to predict the number of residues that are frequently present in B-cell epitopes, and would therefore contribute to increased affinity and binding by antibodies in downstream assays³³. The peptides were then assigned a cumulative rank based on the epitope prediction and Gini values, and the 10 highest-ranking peptides across the E and NS1 proteins for each viral taxon, as well as 8 peptides in the E2 region for CHIKV, were categorized as the most likely to have high immunodominance and therefore be recognized by antibodies in sera collected from previously infected patients in the western hemisphere. Statistical comparisons of quantitative differences between the Gini and normalized fluorescence values for sets of peptides were performed using Student’s t-test.

Peptide validation using ELISA

Each peptide was synthesized (LifeTein, LLC) and 2 ng of peptide was diluted in 50 μL of ddH2O. Natural human IgG protein (abcam, cat. # ab91102), complement component C1q from human serum (sigma, cat. #: C1740), and labelled secondary antibody (ThermoFisher, cat. #: A18817, RRID: AB_2535594) were used as additional controls. Pools of two peptides were used to coat duplicate wells on a 96-well Immulon 4HBX plate (ThermoFisher, cat. # 3855) and incubated at 4°C overnight. Next, 100 μL of blocking buffer (PBS+5% BSA) was added to each well and incubated for 2 hours at room temperature prior to three washing steps with washing buffer (PBS + 0.05% Tween 20). Human serum was diluted 1:25 in blocking buffer and 50 μL of this solution was added to each well prior to incubation for 2 hours at room temperature. Each plate was then washed four times with washing buffer and 50 μL of HRP-conjugated anti-human IgG antibody (ThermoFisher, cat. #: A18817, RRID: AB_2535594; 0.1 mg/mL diluted 1:20,000) was added to each well, followed by incubation at room temperature for 2 hours. Each plate was then washed four additional times before incubating at room temperature for 30 minutes with 75 μL of TMB substrate (abcam, cat. # ab171523). Then, 75 μL of stop solution (abcam, cat. # ab171529) was added to each well and a BioTek-synergy HT plate reader was used to quantify the fluorescence in each well at 450nm within 15 minutes.

ELISA data processing

A normalization process was implemented that adjusted fluorescence values in each well based on the control wells included on each ELISA plate³⁴, which enabled the downstream comparison between plates. Briefly, the average background value from all negative control wells (i.e. no bound peptide) was calculated and subtracted from each set of duplicate wells on the plate. The normalized value was calculated by dividing the background-corrected value for each set of duplicate wells by the background-corrected average of the positive control wells (i.e. bound secondary antibody). Wells with normalized values of greater than 2.5, between 1.5 and 2.5, or less than 1.5 were categorized as putative positive, borderline, or negative, respectively, for the target viruses. A downstream quality control method was also implemented to ignore results from ELISA plates that displayed high levels of background, inconsistent signal from multiple control wells, or samples observed to have at least two wells for each taxon with higher than expected signal.

Human subject approval

All samples evaluated on the peptide arrays and ELISA plates were acquired from patients under informed consent and approved by the Ethical or Institutional Review Board at each participating institution, including: Universidad Nacional Autonoma de Honduras (IRB 00003070), Vanderbilt University (IRB 8675), University of California, Berkeley (Committee for Protection of Human Subjects 2010-09-2245), and the Comite Institutional de Revision Etica (NIC-MINSA/CNDR CIRE-09/03/07-008.ver19).

Data records

Overall, we screened 137 unique serum samples for their reactivity against a panel of viral peptides (Figure 1 and Underlying data³⁵). These samples, together with the clinical diagnosis, were collected from patients with known past exposure to at least one of the viruses targeted by our peptides (Table 1). Also contained within the Underlying data are files describing the metadata of peptides included on the array and each experimental sample³⁵.

Figure 1. Bioinformatics and laboratory workflow diagram.

A graphical depiction of the processes for predicting, screening, processing, and validating peptide array experiments.

The data from each array is contained in a single tab-delimited text file and contains the quantitative data captured from a single serum sample on a single peptide array³⁵. A subset of the fields in each file include: location of each peptide spot on the array, peptide identifier, raw mean and median foreground fluorescence at 488 nm, raw mean and median background fluorescence at 488 nm, and other data collected from the raw image.

A matrix containing the transformed Z-score values for each peptide was then formatted for input into a random forest (RF) machine learning algorithm to assist with ranking peptides according to virus taxon. To do so, a column was added to the matrix assigning each sample to the virus taxon that was known to have infected the patient (e.g. “Zika” or “Non-Zika”). Z-score values in columns containing the predicted peptides from each taxon were then captured and input into the RF algorithm.

The benefit of the RF algorithm is that it is capable of ranking the importance of features, which are peptides in this case, based on a known classification. The ranking is based on the mean decrease in Gini index, which is a value that quantifies node impurity. In other words, the higher the Gini index value, the more important the feature is in correctly identifying the virus taxon.

In order to account for geographical, genetic, and population-based factors, we computed the mean decrease in Gini index for individual collections (e.g. Nicaragua or Honduras), all relevant pairs of collections (e.g. Nicaragua and Honduras, Honduras and United States), and the combination of all collections from our sera providers. These calculations were accompanied by a class-error rate that quantifies the number of samples characterized as being positive for ZIKV that were predicted to be ZIKV samples.

This class-error rate information for each individual or combination of collections was then used to weight the peptide rankings results. Briefly, this involved multiplying the average rank for each peptide in each comparison by the average weight and dividing it by the sum of weight. This process works to increase the rank of peptides that have consistently high Gini values. We used these rankings to identify the top 25 species-specific peptides for each virus taxon. This process was repeated for non-ZIKV samples, including WNV, DENV1-3, and CHIKV.

In addition to the class-error rate, we visualized the random forest output using a receiver operating characteristic (ROC) curve to graph the relationship between true-positive rate and false-positive rate. The area under the curve (AUC) can be calculated with higher values indicating better accuracy. We performed this analysis for the collections from Nicaragua, Honduras, the United States, and the combination of all three collections (Figure 2).

Figure 2. Receiver operating characteristic curve for random forest results.

Individual curves were calculated for A) the combination all of three collections of sera, B) the Nicaragua collection of sera, C) the United States collection of sera, and D) the Honduras collection of sera.

In order to conserve resources for the peptide array and decrease the number of peptides that would be incorporated into the future ELISA assay, we used the existing BepiPred 2.0 algorithm to predict which of our 15-mer peptides contain the highest number of amino acids that are most often recognized by antibodies³⁶. These B-cell epitope predictions were then used to reduce the 25 best peptides identified from machine learning, to the 10 best peptides that are predicted to not only be species-specific, but that are most likely to contain species-specific epitopes. In the case of ZIKV, we also reviewed the spot size and shape in the peptide array images to ensure that there were no irregularities that could negatively bias our results. The BepiPred 2.0 results enabled us to predict which peptides would be most seroreactive for each selected taxon. The mean maximum score from the BepiPred 2.0 analysis was calculated to be 0.58 (range: 0.55 – 0.63). These scores are associated with a specificity greater than 81%.

Computational validation

Given the serological cross-reactivity that has been reported among many of our targeted mosquito-borne viruses³⁷, we recognized the need to validate the results of our high-throughput screen. To do so, we not only ensured that those generating the peptide array data were “blinded” to the phenotype of each sample, but we also computationally evaluated two distinct but complementary comparative and quantitative metrics that are described below.

First, we compared two serum samples from pediatric patients that had not been infected with DENV prior to sample collection. The data from the DENV-specific peptides in these samples were then compared to those from a representative DENV-positive sample to verify the differences in signal between known positive and known negative samples. This comparison would also provide a better understanding of the contribution of cross-reactivity, which has been reported previously³⁷, on our platform (Table 2). This comparison showed that the DENV-negative samples had less than four percent of the normalized fluorescence values, well below the 10 percent that was observed in the DENV-positive sample. Transforming these raw data into Z-scores further increases the observed differences in fluorescence values and, provides additional support to the unbiased nature of the data produced in these experiments.

We next wanted to assess the technical rigor of our approach by performing a statistical analysis of the observed experimental variation in the peptide array experiments. In this case, data was available for six of our target viruses for which sera was evaluated on the arrays. We specifically wanted to quantify the reactivity of the best-performing peptides for each sample against in a panel of comparisons (Table 3). The results from this analysis identified noticeable differences in the signals for ZIKV and WNV (Figure 3). However, we observed that the quantified values for the other four virus taxa were lower than the values for all samples combined and did not meet statistical significance when comparing known positive and negative samples (Figure 4). These results show that incorporating Gini scores and immune epitope predictions into our computational pipeline contributed to our ability to identify sets of peptides that were capable of distinguishing between past infection with a subset of our target viruses.

Figure 3. Mean of normalized ratio values across multiple viral taxa.

A bar chart depicting the mean values of the best-performing peptides for samples that were characterized as: positive for the specified taxon, negative for the specified taxon, positive or negative for the specified taxon, and all samples across all viral taxa.

Figure 4. Quantitating statistical significance between normalized ratio values.

A bar chart showing the -log10 p-value for the best-performing sets of peptides for each virus species. Comparisons include: samples that were positive for the virus vs. those that were negative for the virus (+ vs -), samples that were positive for the virus vs. all samples quantified for the virus), and samples that were positive vs. all quantified samples (+ vs. All). Values greater than 1.30 indicate p < 0.05.

It is also important to recognize that each peptide was printed at non-adjacent sites on each array in quadruplicate to minimize experimental bias due to the location of any given spot on the array. Incorporating technical replicates was an important component of the experimental design. Such an approach enables improved replication of the results and also increases the scientific rigor of the resulting dataset upstream of any data processing workflows.

Experimental validation

The number of samples that were evaluated for prior exposure to each virus was insufficient to allow the use of in silico cross-validation techniques that are generally applied to the classifier predictions. We therefore designed custom 96-well ELISA plates to validate the ability of the peptides (Figure 5). The highest predicted reactivity to accurately detect prior infection by each of the target viruses.

Figure 5. Layout of custom ELISA plate to screen sera for reactivity against predicted peptides.

Testing of duplicate pairs of each peptide for each of the target virus taxa. Each column contains pairs of peptides that were predicted to be effective at differentiating each virus based on the bioinformatics processing of the peptide array data. (row A for ZIKV; row B for CHKV; row C for WNV; row D for DENV-1; row E for DENV-2; row F for DENV-3; row G for DENV-4; row H for YFV). Negative control wells, which were not coated with peptides are in column 12 and in wells with no label. Wells containing other control material are in column 11 (green) and include: C1q, naïve human immunoglobulin G, or only secondary antibody.

These custom ELISA plates were incubated with 26 human convalescent sera that had been previously characterized as positive for at least one of our target virus taxa using complementary methods, including plaque reduction neutralization test (PRNT) from convalescent serum, IgM antibody capture enzyme-linked immunosorbent assay (MAC-ELISA) from post-acute phase serum, and/or quantitative real-time PCR (qRT-PCR) from blood collected during acute infection. These samples were obtained from public sources including: BEI Resources (5 samples), the World Reference Center for Emerging Viruses and Arboviruses (7 samples), or the United States Centers for Disease Control and Prevention (16 samples).

After processing and correcting the raw ELISA data, we found that the well-characterized samples showing a normalized absorbance ratio greater than 1.5 correlated with cases of previously confirmed Zika infection (Table 5–Table 30). Consequently, we compiled the normalized absorbance ratio data and categorized any peptide pool found to have a normalized ratio value greater than 1.5 was classified as a “borderline” result, while those having a value greater than 2.5 was classified as a putative “positive” result. In order to increase specificity, any sample with at least two wells of the ELISA plate having normalized ratios greater than 1.5 were labeled as putative “positive” for prior infection with the target virus.

Given the p-values associated with the peptide array results, we decided to especially focus on samples that were positive for ZIKV. As such, instances where excessive signal was detected for all viruses were processed in a way that still identified samples having at least 2x stronger signal for ZIKV peptides than for DENV peptides in the same sample were labeled with a “Z” to differentiate them from other categories.

The summarized results of the ELISA data revealed a fair amount of concordance with the “gold standard” methods and displayed overall sensitivity and specificity of 61.5% and 50%, respectively (Table 4). Interestingly, these values fluctuated depending on the collection that was analyzed and were affected by small sample size from two of the collections.