FluChip‐8G Insight: HA and NA subtyping of potentially pandemic influenza A viruses in a single assay

Abstract Background Global influenza surveillance in humans and animals is a critical component of pandemic preparedness. The FluChip‐8G Insight assay was developed to subtype both seasonal and potentially pandemic influenza viruses in a single assay with a same day result. FluChip‐8G Insight uses whole gene segment RT‐PCR‐based amplification to provide robustness against genetic drift and subsequent microarray detection with artificial neural network‐based data interpretation. Objectives The objective of this study was to verify and validate the performance of the FluChip‐8G Insight assay for the detection and positive identification of human and animal origin non‐seasonal influenza A specimens. Methods We evaluated the ability of the FluChip‐8G Insight technology to type and HA and NA subtype a sample set consisting of 297 results from 180 unique non‐seasonal influenza A strains (49 unique subtypes). Results FluChip‐8G Insight demonstrated a positive percent agreement ≥93% for 5 targeted HA and 5 targeted NA subtypes except for H9 (88%), and negative percent agreement exceeding 95% for all targeted subtypes. Conclusions The FluChip‐8G Insight neural network‐based algorithm used for virus identification performed well over a data set of 297 naïve sample results, and can be easily updated to improve performance on emerging strains without changing the underlying assay chemistry.

human and animal populations to better guide pandemic vaccine production and mitigate the potential impact of the global public health threat of pandemic influenza. [3][4][5][6][7] Routine surveillance of influenza viruses often involves molecular assays based on real-time RT-PCR typically performed by clinical and public health laboratories. Though PCR-based assays are sensitive, they typically rely on amplifying short portions of HA and NA, both of which exhibit high rates of genetic drift, 8 with real-time RT-PCR assays having shown to be susceptible to reduced performance as genetic drift occurs. [9][10][11][12] Many real-time RT-PCR assays require numerous singleplex assays run serially, increasing the complexity and time to result for HA and NA subtyping. Microarray-based approaches to influenza detection are also available. In addition to commercially available microarraybased assays such as the ePLEX and eSensor RVP assays (both planar microarray assays, GenMark Diagnostics, Inc), the Verigene RV+ (planar microarray, Luminex Corp.), and the xTAG RVP assay (solution phase bead-based microarray, Luminex Corp.), a variety of other microarraybased approaches to influenza detection for both human and avianorigin have appeared in the literature [13][14][15][16][17][18][19][20][21] but are not commercially available. Alternatively, next-generation sequencing (NGS) of influenza viruses is routine in the three National Influenza Surveillance Reference Centers (NIRC) and at the Centers for Disease Control and Prevention (CDC), but has not been adopted for surveillance and other clinical microbiology applications in clinical and public health laboratories in part due to availability, cost, and complexity of data analysis. 2,22,23 In this work, we report on a new assay called FluChip-8G Insight for the detection and characterization of influenza viruses, including HA and NA subtyping for key potentially pandemic subtypes.
FluChip-8G Insight was developed with the goal of aiding public health, government, and academic laboratories conducting surveillance for pandemic influenza preparedness. The ability to differentiate influenza A viruses with high-risk pandemic potential from currently circulating strains of influenza A is an important factor for influenza response preparedness and could enable more efficient detection of influenza viruses warranting immediate follow-up analysis. The technology utilizes multiplexed RT-PCR in which multiple full gene segments are amplified, followed by hybridization to a DNA microarray containing capture sequences that represent a significant fraction of the influenza A genome, and ultimate application of a modular, easily updateable neural network-based pattern recognition to the microarray data. The benefits of this approach are universal amplification and detection of all influenza A and B viruses that provide robustness against genetic drift and shift, completely automated data interpretation, and the ability to rapidly update the underlying analysis algorithm to identify newly emerging viruses and to optimize performance as additional strains are obtained for inclusion in algorithm training.
FluChip-8G Insight provides typing and subtyping of seasonal influenza A viruses and HA (H1, H3, H5, H7, and H9) and NA (N1, N2, N7, N8, and N9) subtyping of non-seasonal A viruses in a single assay. Herein, we describe the verification of the HA and NA subtyping algorithm and subsequently assess performance in a validation study of 297 non-seasonal influenza A samples.

| Viruses and characterization
All samples utilized in neural network algorithm optimization and subsequent testing were either archived original specimen material, grown viral isolates, or extracted nucleic acid (for highly pathogenic viruses or when no other material was available) of known type, subtype, and strain via either sequencing, real-time RT-PCR analysis, or via certificate of analysis obtained from a commercial vendor. All respiratory specimens obtained and utilized in this testing were completely de-identified and provided without any protected health information. For several human clinical specimens, only subtype information was available. Contrived samples were prepared by spiking stock material into individual or pooled influenza-negative human nasopharyngeal swab material stabilized in universal transport medium (UTM), and subsequently characterized by the CDC real-time RT-PCR influenza A typing assay to estimate concentration prior to analysis. All samples were stored at −70°C or below prior to executing the FluChip-8G Insight assay.

| FluChip-8G Insight assay procedure
The FluChip-8G Insight assay procedure was executed on all samples described herein according to the instructions for use. In addition, for samples containing human genetic material a segment of the 18S rRNA found in eukaryotic cells was amplified as an internal control to assess sample integrity. RT-PCR products were subsequently heat fragmented and hybridized to the microarray containing 458 influenza-targeted short oligonucleotide capture sequences designed to target subtype-and lineage-specific sequences of the 7 amplified gene segments. The microarray was subsequently washed, labeled, and imaged using the fluorescencebased FluChip-8G Imaging System (InDevR, Inc).  Supplementary Tables S1 and S2). Tier 2 neural networks were trained to identify "non-seasonal A" subtypes H1, H3, H5, H7, H9, N1, N2, N7, N8, and N9 as well as general "Hx" and "Nx" categories that included all other subtypes not specifically listed.

| Neural network training
Training of tier 2 networks was performed using 1479 microarray images (samples) representing 140 unique influenza A strains (see Supplementary Tables S3 and S4).
The neural networks were trained in 3 steps: First, the samples were split into 10 groups of approximately equal size and composition. Second, 10-fold cross-validation was completed by training neural networks utilizing sample groups 1-9 with group 10 reserved for testing the newly trained networks. This process was repeated 9 times with a different group held out for testing each time. The results of the 10-fold cross-validation were then examined to assess the ability of each of the groups of trained neural networks to predict the expected sample result (see Results section). Lastly, final training was completed using all ten 10 groups to produce the optimized neural network algorithm. Upon successful training, tier 1 and tier 2 network parameters were coded into the FluChip-8G Insight Software which was subsequently loaded onto the FluChip-8G Imaging System for validation testing using a completely naïve sample test set.

| Naïve test set composition
The test set used for evaluation of Tier 2 network performance was made up of 297 microarray images each representing a sin-  Table S3) and were comprised of human samples and grown isolates or extracted nucleic acid of avian, porcine, and equine strains from North America, South America, Europe, and Asia. All 180 unique strains included were samples known to be "non-seasonal" influenza A via the methods described above. Of the 297 total images, 280 were identified as "influenza A, non-seasonal" by the tier 1 networks and therefore were analyzed by the tier 2 networks for subtype identification. Performance data for tier 2 networks therefore represent only sample images correctly identified as "non-seasonal" influenza A by the tier 1 networks.

| Algorithm optimization via 10-fold crossvalidation
All of the virus samples included in the tier 2 neural network training set (see Supplementary Table S3) were processed by the FluChip-8G Insight assay, with the microarray images subsequently utilized to perform 10-fold cross-validation of the tier 2 neural network algorithm. Performance of the 10-fold cross-validation of the training set is shown in Table 1

| Assay performance assessment on naïve sample test set
To independently validate performance of the FluChip-8G Insight assay and algorithm, the naïve sample test set of contrived samples (see Supplementary Table S5) was processed using the FluChip-8G Insight assay and the data subsequently analyzed using the optimized neural network algorithm. The result of each specimen processed was compared to the expected result to assess the positive and negative percent agreement shown in Table 2.
For HA and NA subtyping, positive percent agreement for all classifications targeting specific subtypes exceeded 93%, except for H9 which resulted in 87.9% positive percent agreement. Similar to the results seen for the 10-fold cross-validation shown in Table 1, the "Hx" and "Nx" categories demonstrated lower positive percent agreement of 76.3% and 79.7%, respectively. Negative percent agreement for HA and NA subtypes targeted exceeded 95%.