Association analyses of large-scale glycan microarray data reveal novel host-specific substructures in influenza A virus binding glycans

Influenza A viruses can infect a wide variety of animal species and, occasionally, humans. Infection occurs through the binding formed by viral surface glycoprotein hemagglutinin and certain types of glycan receptors on host cell membranes. Studies have shown that the α2,3-linked sialic acid motif (SA2,3Gal) in avian, equine, and canine species; the α2,6-linked sialic acid motif (SA2,6Gal) in humans; and SA2,3Gal and SA2,6Gal in swine are responsible for the corresponding host tropisms. However, more detailed and refined substructures that determine host tropisms are still not clear. Thus, in this study, we applied association mining on a set of glycan microarray data for 211 influenza viruses from five host groups: humans, swine, canine, migratory waterfowl, and terrestrial birds. The results suggest that besides Neu5Acα2–6Galβ, human-origin viruses could bind glycans with Neu5Acα2–8Neu5Acα2–8Neu5Ac and Neu5Gcα2–6Galβ1–4GlcNAc substructures; Galβ and GlcNAcβ terminal substructures, without sialic acid branches, were associated with the binding of human-, swine-, and avian-origin viruses; sulfated Neu5Acα2–3 substructures were associated with the binding of human- and swine-origin viruses. Finally, through three-dimensional structure characterization, we revealed that the role of glycan chain shapes is more important than that of torsion angles or of overall structural similarities in virus host tropisms.

Scientific RepoRts | 5:15778 | DOi: 10.1038/srep15778 could allow these animals to be susceptible to avian-origin and human-origin influenza A viruses; thus swine have been proposed as "mixing vessels" for influenza viruses 12 .
By presenting multiple glycans or glycoconjugates printed on a single slide, the glycan microarray technique has offered high-throughput analyses of the glycan-binding profile of influenza viruses 15,16 . Glycan microarray has become a routine experimental tool for characterizing the receptor-binding profiles of influenza viruses 17 . To date, > 500 influenza virus-related glycan microarray data entries have been deposited in the Consortium for Functional Glycomics (CFG) glycan microarray database 18 , and this number is still increasing. Glycan microarray profiling of influenza virus pandemic strains has shed light on the receptor-binding specificities of their hemagglutinin (HA). For example, such analyses revealed that the 2009 influenza A(H1N1)pdm09 pandemic virus bound to α 2,6-linked and to a large range of α 2,3-linked sialyl sequences [19][20][21][22] . Moreover, glycan microarray analysis has been widely used to study receptor recognition and host tropism of influenza virus mutants [23][24][25][26][27][28] . In addition to providing data on Neu5Acα 2-3Gal and Neu5Acα 2-6Gal, glycan microarray analysis also provided data on other complicated glycan substructures. In one study, structural topology (i.e., two glycan chain shapes, one cone-like and the other umbrella-like) was reported to be related to SA2,3Gal and SA2,6Gal during influenza virus-receptor interactions 29 . On the other hand, during a glycan microarray screening, influenza A viruses were shown to bind receptors other than SA2,3Gal and SA2,6Gal, although such bindings have not been confirmed by interventional experiments 22,30 . For example, influenza A (H1N1) virus can bind α 2,8-linked polysialyl sequences 22 . Nevertheless, it is still unclear what specific substructures or moieties in host receptors determine influenza virus host tropisms.
To better understand structural specificities for glycan binding, Cholleti et al. 31 developed an algorithm called GlycanMotifMiner, or GLYMMR, that is frequently used with subtree mining to identify motifs for protein-glycan interactions for a given glycan microarray data entry. Porter et al. 32 applied a clustering algorithm to identify glycan substructures with high intensities in the glycan array data. More recently, we developed a novel quantitative structure-activity relationship (QSAR) method to analyze glycan array data; the method focuses on glycan substructure features by applying PLS regression and selection functions to the glycan microarray data 33 . Another frequent glycan structure mining of influenza virus data also detected sulfated glycan motifs increased viral infection 34 . However, none of the above methods were designed for large-scale glycan microarray data analysis that integrates multiple microarray data entries for a particular research interest. Particularly, statistic-based motif identification methods rely on pre-defined hypothesis and could not discover unexpected and infrequent ones. Feature selection strategies for the regressions of glycan microarray data have not considered modeling multiple microarrays. Thus, a computational method is needed to characterize glycan substructure motifs by utilizing the information across multiple datasets, especially glycan microarray data across various platforms, and this method must be able to tolerate the noises within and across glycan microarray data.
The relationship between host receptors (glycan substructures) and influenza A viruses (e.g., viruses with different host origins) can be naturally formulated as a computational problem of data integration plus association rule mining. Therefore, in this study, we firstly applied a PLA regression on individual glycan microarray data entries as normalization and then used association rule analysis on extracted glycan substructure features to identify motifs for influenza virus host tropisms. In addition to SA2,3Gal and SA2,6Gal, results showed that glycan substructures with SA2,8SA, non-sialic acid saccharides (Galβ and GlcNAcβ terminal substructures), and sulfated SA2,3Gal could contribute to influenza host tropism differently. Additional computational modeling demonstrated that, for trisaccharide substructures, a shape angle formed by mass centers of three residues, instead of linkage torsion angles, may determine the overall glycan chain shapes and thus distinguish glycans with SA2,3Gal from those with SA2,6Gal or SA2,8SA. These findings may imply a more general property caused by glycan terminals than just by sialic acid with different linkages during influenza -glycan binding. Figure 1 shows a simplified flowchart of the computational strategy we used, with glycan microarray data, to identify host-specific glycan substructures. In brief, we collected and integrated glycan microarray datasets (Fig. 1A), defined and extracted substructures from glycans (Fig. 1B), and applied association rule mining to identify the influenza viruses' specific glycans and their substructures (Fig. 1C).

Datasets. Collection of influenza A virus-specific glycan microarray data.
A computational script was written to automatically retrieve glycan microarray datasets from CFG 18 by using the keyword "influenza. " A total of 542 glycan microarray entries were retrieved, of which 324 were excluded from the final dataset: 31 entries for influenza B viruses, 182 for mutant viruses, 51 for mouse-adapted strains, 53 for HA recombinant proteins, and 7 for microarrays with incomplete binding affinity values. The remaining 218 entries were for influenza A virus-specific glycan microarray datasets with complete binding affinity values. These datasets, which were used for further analyses (Table 1), consisted of influenza A viruses of human origin (n = 154), waterfowl origin (n = 17), terrestrial bird origin (n = 13), canine origin (n = 6), and swine origin (n = 21). The metadata associated with these datasets, including CFG entry identification codes, investigators' names, influenza virus sample names, glycan array version, raw array binding files, and host species, are listed in Table S1 in the supporting information.
Integration of glycan microarray datasets. In the CFG database, the datasets were generated by using 11 versions of glycan microarrays, each of which had a different number of glycan entries. For example, version 1 had 200 glycans, whereas version 5.1 had 611 glycans. However, most glycans present in earlier glycan microarray version are present in later versions. To facilitate the data analyses across different datasets, we merged all microarray versions into one with 936 unique glycans (Table S2) and generated a single matrix for the 211 data entries (211 viruses × 936 glycans, Table S3). Glycan-binding affinities (i.e., fluorescent signal values) in our dataset were assigned to corresponding elements in the matrix. The elements for which there was no corresponding affinity value among the 936 glycans were assigned a "not available" value and excluded from the glycan substructure feature selection. Glycan substructure feature extraction. Glycan substructures were defined as described elsewhere 33 . Specifically, mono-, di-, tri-, and tetrasaccharide substructures were extracted from 936 glycans as features. These extractions resulted in 249 monosaccharide, 738 disaccharide, 1,198 trisaccharide, and 1,477 tetrasaccharide substructures (Tables S4-S7). The fluorescent signal value for the corresponding glycan on the array was assigned as the binding affinity for each individual substructure. Only fluorescent signal values ≥ 2,000 were considered as effective numbers in regression, and those < 2,000 were treated as background noise. Next, a partial least squares (PLS) regression and feature selection algorithm (QSAR 33 ) were adapted to select the features predominating glycan binding from an influenza virusspecific glycan microarray dataset (see details in Supplementary Information). This PLS regression was performed four times for each single data entry from our 211 glycan microarrays by using four sets of substructure feature definitions (mono-, di-, tri-, and tetrasaccharides). Each feature vector was finally labeled according to the host origin of the influenza A viruses used in the glycan microarray experiments (i.e., human, swine, canine, waterfowl, or terrestrial bird [chicken, quail, and turkey] host).
Association rule mining for selected glycan substructures. We formulated the detection of host-specific glycan substructures as an association-mining task (see more details in Supplementary Information), where we let items = , , , represent a set of items and let = , , , be a set of transactions forming a database. An association rule, ⇒ X Y, where , ⊆ X Y I, is usually interpreted to mean that when the items in X exist, those in Y also occur at a certain confidence level 35 . Here, for our glycan microarray dataset, transactions T were the data derived from influenza virus-specific glycan microarray entries, so m = 211; the substructure features X derived from glycans on the array by previous PLS-β selection and the labeled features Y with host origin will form I. Given a rule ⇒ X Y, the confidence is defined as The support was defined as the proportion of transactions in the dataset, which contains the item set. Another measurement, lift, is the ratio of the observed support and was defined as 35 . Therefore, we expected to obtain interesting association rules with high confidences (≥ 80%), high lifts (has a lift value ≥ 1 36 ), and low supports (≥ 0.005, infrequent but potentially interesting) to supply highly probable, unexpected, and infrequent conclusions. We adapted the Apriori algorithm implemented in R 37 to infer these host substructure-specific associations. Moreover, during the mining process, redundant rules were also removed by defining super rules as redundancy. A super rule is a rule with the same or lower lift value, where the left hand side, X, contains more items than a previous rule, but still results in the same right hand side, Y. Last, we kept only satisfied rules, which were filtered by leaving only those with terminal saccharides on the substructure features.

Three-dimensional structural modeling and analysis. Structural characterization for terminal gly-
can saccharides. To understand the structural determinants for a specific glycan associated with certain influenza A virus, we compared the spatial relationship between six terminal trisaccharide features derived from data mining. These six features were (Neu5Acα 2-6)-(6Galβ 1-4)-4GlcNAc (PDB 38 accession number:   (1) The angle formed by the mass centers of three saccharides. We calculated the angle formed by the mass centers of three saccharides as a measurement of the glycan chain's turning shape. (2) The root-mean-square deviation (RMSD). Given two glycan substructures, each containing the terminal saccharide, we superimposed the corresponding atoms on the six-membered rings of the terminal saccharides. From there, while keeping the terminal saccharides superimposed, the following two values of RMSD were measured: RMSD2 and RMSD3. Using the standard formula of calculating RMSD from two sets of six-membered ring atoms v and w: where n = 6, RMSD2 was calculated for the two saccharides linked directly to their respective terminal saccharides. If both glycan substructures had a third saccharide, RMSD3 was then calculated for the third pair of saccharides. (3) φ and ψ torsion angles. We calculated the φ and ψ torsion angles for each linkage between two adjacent residues. Glycosidic torsions were defined by only heavy linkage atoms as φ = O5-C1-O n -C n and ψ = C1-O n -C n -C n-1 44 . Accordingly, a trisaccharide substructure has two linkages with two sets of torsions.

Results
Influenza virus-specific features derived from glycan microarray data by PLS regression and feature selection. Certain saccharide residues are enriched at glycan substructures contributing to influenza virus binding. In the integrated dataset of glycan microarrays with 936 unique glycans, the glycans with influenza virus-binding affinities mostly consist of sialic acids (Neu5Ac and Neu5Gc), galactose (Gal), N-acetylgalactosamine (GalNAc), glucose (Glc), N-acetylglucosamine (GlcNAc), fucose (Fuc), and mannose (Man). Table 2 summarizes the number and percentage of glycan substructures that have these saccharides by each one of the four substructure feature definitions on the glycan microarrays (i.e., mono-, di-, tri-, and tetrasaccharides) and thus reflects their existence according to the microarray design. Table 3 lists the same distribution values of influenza virus-specific substructures selected by PLS regression and illustrates that only a small portion of glycan substructures (73/249 monosaccharides, 230/738 disaccharides, 322/1,198 trisaccharides, and 320/1,477 tetrasaccharides) was determined to contribute to a binding signal of ≥ 2,000 with influenza viruses. All PLS-selected substructure features are also summarized in Table S8.
A comparison of data in Table 2 with that in Table 3 shows that Neu5Ac, Neu5Gc, Gal, and GlcNAc were more abundant in the glycan substructures contributing to influenza virus binding. For example, Neu5Ac appeared in 9.59% of the monosaccharides, 11.3% of the disaccharides, 18.0% of the trisaccharides, and 31.9% of the tetrasaccharides when QSAR was used to select significant glycan substructures for influenza virus binding (Table 3), compared with 4.82%, 6.10%, 7.68%, and 10.9%, respectively, of all the glycan substructures from microarrays ( Table 2). Similar differences were observed for Neu5Gc, Gal, and GlcNAc. These findings suggest that influenza virus-specific glycan substructures are prone to have these four saccharides. Nevertheless, when QSAR was used, glycan substructures with GalNAc, Glc, Fuc, and Man were equally or less frequently correlated with influenza virus binding than those on the glycan Scientific RepoRts | 5:15778 | DOi: 10.1038/srep15778 arrays (Tables 2 and 3), which indicates a limited contribution to influenza binding by substructures with these saccharides.
Host-specific glycan substructures derived from association rule mining. To understand the specific substructures associated with each influenza A virus, we performed associate rule analyses across 211 influenza virus-specific glycan microarray data. On the basis of their host origins, the 211 influenza A viruses were categorized into human (n = 154), canine (n = 6), swine (n = 21), waterfowl (n = 17), terrestrial (i.e., chicken, quail, and turkeys, n= 13), and avian (waterfowl plus terrestrial birds, n = 30). The association analysis results (summarized in Fig. 2 and Table S9) illustrate the specific substructures being associated with each of six host origins; these associations aid in our understanding of the key substructures that determine influenza host and tissue tropisms.

Consensus among the influenza virus-specific glycan substructures.
To identify common features from the substructures associated with different hosts, we compared the structural similarity among them by calculating, as described in Methods, the angle formed by three mass centers of all residues for trisaccharide substructures, RMSD2 and RMSD3, and φ and ψ torsion angles of linkages for the six representative glycan substructures (Fig. 4). In addition, superposition images of these glycan substructures are shown in Fig. 5. Fig. 4A,D, we obtained four trisaccharide three-dimensional structures, of which one has SA2,6Gal terminal, one has SA2,8SA terminal, and two have SA2,3Gal terminals. Three observations were made from the substructures. First, the residue mass centers for the SA2,6Gal substructure formed acute angles (63.1°), for the SA2,8SA substructure formed an angle of 91.1°, and the mass centers for both α 2,3-linked substructures formed obtuse angles (142.9° and 132.2°). This observation suggests that SA2,6Gal and SA2,3Gal substructures are fundamentally different from each other on saccharide chain shapes and thus could lead to virus host tropism, in which human influenza viruses recognize glycans with SA2,6Gal terminals, canine and avian viruses recognize glycans with SA2,3Gal terminals specifically, but swine viruses can recognize and bind to both shapes. Moreover, the α 2,8-linked polysialyl substructure with a right angle shares a more similar turning shape to the one of SA2,6Gal and then may cause the human-origin influenza virus binding. Second, the all-against-all RMSD values for these glycan substructures indicate that none of the substructures with sialic acid terminals are similar on the basis of both RMSD2 and RMSD3 values, if we define similar saccharide structures by using RMSD2 smaller than 3 Å and RMSD3 smaller than 5 Å (Table 4, Fig. 5). This finding shows that shape angles formed by residue mass centers are not the sole factor for glycan structural diversity.

3D structural characterization for glycan substructures with sialic acid terminals. As shown in
The third observation involves the linkage torsion angles of these four representative host-specific trisaccharides ( Table 5). It is shown that, although most torsions of linkage 2 share similar values and hence do not contribute much to virus host types, both φ and ψ angles of linkage 1 distribute variously and indicate the shape-forming roles of α 2,6, α 2,8 and α 2,3 linkages with terminal sialic acids. In particular, the linkage 1 φ angle values of these four trisaccharides, combined with their shape angles formed by residue mass centers, could shed some light on the relationship between glycan geometric shapes and influenza virus host types. On one hand, when trisaccharides with SA2,6Gal or SA2,8SA have angles of acute shapes (Fig. 4A,B), a positive φ angle (e.g., 71.32° of Neu5Acα 2-6Gal or 55.05° of Neu5Acα 2-8Neu5Ac in Table 5) might be necessary to make the glycan associated with human host type; however, the association might not be unique because the positive φ angle might also result in an association with swine viruses. On the other hand, when trisaccharides with SA2,3Gal have angles of obtuse shapes (Fig. 4C,D), different terminal residues (i.e., Neu5Ac and Neu5Gc) form φ angles with different values (e.g., a − 59.47° of Neu5Acα 2-3Gal and a 50.95° of Neu5Gcα 2-3Gal in Table 5). This observation illustrates that an obtuse angle of α 2,3-linked trisaccharides is sufficient, but not necessary, for a glycan to associate with non-human-origin viruses and that a positive φ torsion angle at linkage 1 may make the trisaccharides associated with canine-and avian-origin viruses only. Furthermore, all four trisaccharides, except Neu5Acα 2-8Neu5Acα 2-8Neu5Ac, have linkage 1 ψ angles of similar negative values and therefore do not show a clear relationship with virus host types.
In summary, the structural characteristics of glycan trisaccharides with sialic acid terminals might be associated with influenza virus host tropisms. For example, it seems that the shape angle formed by residue mass centers plus the linkage 1 φ torsion angle, not just torsion angles themselves, might suggest certain glycan structural patterns associated with influenza virus host tropism.

Structural conservation of receptor binding pocket in influenza A viruses. In
In Fig. 6C,D, we docked Neu5Acα 2-8Neu5Ac and Galβ 1-4GlcNAcβ 1-3Gal to the receptor binding pocket of the human-origin HA (PDB 3LZG) and the avian-origin HA (PDB 2FK0) separately by using a HA-glycan structural complex as the template (see Methods). Previous association results suggested   a relationship of Neu5Acα 2-8Neu5Ac and Galβ 1-4GlcNAcβ 1-3Gal with the binding for influenza A viruses; thus, their comparable binding poses are expected to occur at the virus HA binding pockets (Fig. 6C,D).

Discussion
The objective of this study was to characterize the host-specific glycan substructure responding to influenza A virus infections. Glycan microarray data provide an opportunity to systematically study the factors that determine virus-glycan binding. However, such analyses have several limitations. The first limitation is that glycan microarray data are not quantitative because values from batch to batch are highly variable. The variability is caused by spot intensities dependent on immobilization efficiency and results in the misleading use of fluorescence intensities to quantify binding affinities 51 . The second limitation is that the glycans on microarray do not represent all glycans or all substructures in the natural hosts, and they are also distributed differently from those in nature. The last limitation is that the number of datasets for influenza A viruses from viruses of different host origins are not equal. For example, we have 155 datasets for human-origin influenza A viruses but only 7 for canine-origin influenza A viruses.
In this study, we expected association analysis to detect significantly nonrandom, but possibly infrequent, substructure features contributing to influenza A virus binding. To ensure better coverage of all potential substructures, hierarchical clusters (mono-, di-, tri-, and tetra-) of substructure profiles were characterized and integrated into data mining, and our analyses focused on the terminal structures. To minimize the potential noise across different datasets due to variations in glycan microarray versions and experiments, we integrated the significant substructures extracted from each individual dataset by PLS regression. To identify the host-associated glycan substructures, we categorized 211 data entries into five categories (human, swine, canine, waterfowl, and terrestrial birds) and then formulated glycan substructure problems as a typical association mining problem, where we treated glycan substructure features as products, virus host types as the only label of customs, and the glycan-virus binding signals in the dataset as transactions. Comparing to other methods, either statistical or mining strategies, our formulation of the problem benefits the novel observations in this study in two following ways. First, after the PLS regressions on individual glycan microarray entries, the binding transection definition was used to integrate all of them for a cross-array analysis, by which we overcame the challenges from the varying numbers of glycans on different version of arrays. Second, the association mining strategy avoided particular hypothesis before analyses and were able to detect rare but potentially significant rules.
We have not been able to use this method to identify the specific substructures for glycan bindings when multiple terminal glycans are present. For example, glycans with different terminals (e.g. sialic acid and Gal) were observed frequently, but they may both be important players during influenza virus binding because they could bind influenza viruses simultaneously. To avoid this problem, in this study, we ignored the associated substructures with branch linkages, because they may be extracted from a glycan with other terminals and by themselves may not contribute to virus binding. To avoid such false-positives, we included in the results only terminal substructures without branches. Moreover, four substructure definitions (mono-, di-, tri-, tetrasaccharide) could lead to overlapped glycan features that were associated with the same virus host. For example, in Fig. 2, swine-associated disaccharides are all subsets of the corresponding trisaccharides, which are subsets of corresponding tetrasaccharides. Similar patterns could be observed with other host-origin categories (Table S9). To be consistent, we interpreted these overlapped rules by ignoring subset features and by keeping only substructures with the highest number of saccharide residues (see Supplementary Methods).
Our results show that (1) human-origin influenza A viruses could bind glycans with Neu5Acα 2-8Neu5Acα 2-8Neu5Ac and Neu5Gcα 2-6Galβ 1-4GlcNAc substructures; (2) Galβ and GlcNAcβ terminal substructures, without any existing sialic acid terminals, are associated with the glycan binding of human-, swine-, and avian-origin influenza A viruses; (3) Sulfated Neu5Acα 2-3 substructures are believed to be associated with the glycan binding of human-and swine-origin influenza A viruses. These observations, on one hand, are consistent with previously reported results about various types of host-origin influenza A viruses 5 . On the other hand, we also identified other substructures: α 2,6-linked Neu5Gc substructures, α 2,8-linked multiple sialic acids, substructures with a Gal and GlcNAc terminals, and sulfated α 2,3-linked Neu5Ac, which contribute to different virus bindings. These newly discovered influenza A binding moieties, particularly those with the non-sialic acidic saccharides (Gal, GlcNAc), may suggest that it is the structural pattern of acidic acids, instead of just Neu5Ac, Neu5Gc themselves, which are recognized by influenza viruses of various host origins.
The potential glycan receptors with α 2,8-linked sialic acid were reported to be associated with influenza virus binding 22 , which supports our results with Neu5Acα 2-8Neu5Acα 2-8Neu5Ac for human influenza viruses. The relatively low 3D structural similarities between this substructures and human-like α 2,6-linked sialic acid substructures (Table 4) could imply a potentially novel binding mode for Neu5Acα 2-8Neu5Acα 2-8Neu5Ac (Fig. 6C). Similarly, it has been reported that glycans with Gal terminals could play a role in some virus receptor binding 52,53 . Our association results detailed this conclusion, especially for Galβ 1-4GlcNAcβ 1-3Gal substructure, by supplying similar structural characteristics to substructures with sialic acid. Concerning the associations detected for sulfated α 2,3-linked Neu5Ac, it was reported that sulfated glycan motifs might increase influenza virus binding 34 . Our results further suggest that this sulfation process may lead to a SA2,3Gal binding for a SA2,6Gal-binding virus. It is worth mentioning that all these substructure motifs were infrequent substructures in association rules (Table S9), indicating effectiveness of association mining in this study.
Our three-dimensional structure analysis of representative host-specific substructures showed that for trisaccharides, the shape angle formed by mass centers of three residues could be the key feature that distinguishes α 2,6-linked, α 2,8-linked and α 2,3-linked glycans and their virus host tropisms (Fig. 4A-D). Although recent studies argued that the different torsion angles of residue linkages could be the reason for their diverse chain shapes 29,44 , our torsion angle values calculated from the three-dimensional structures did not support a role for torsion angle in forming the overall trisaccharide chain shapes. Hence, we argue that significant host-specific patterns related to glycan shape may become evident if shape angles are measured instead of flexible torsion angles. In addition, for trisaccharides without sialic acid terminals (e.g. Galβ 1-4GlcNAcβ 1-3Gal), neither torsion angles nor RMSD values could suggest any host-specific patterns from our results. However, since we only found a few unique such glycans associated with influenza viruses, we considered them only as individual cases of virus binding without an identifiable structural feature for host tropism.