Understanding diversity of human innate immunity receptors: analysis of surface features of leucine-rich repeat domains in NLRs and TLRs

Multiple alignments and phylogenetic trees (see Figure 1) were built using ClustalX2 [35]. The most divergent sequence, NAIP, was used as an outgroup to root the tree. The tree reliability assessment was conducted by means of the bootstrap method using 1,000 tree samples. Several other multiple alignment and tree building strategies produced essentially identical results (data not shown). Protein sequences and gene organization information were obtained from the Ensembl database [36].

In order to visualize structures and surface features of LRR domains from human NLRs (see Figure 2), homology models were built based on pairwise alignments of NLR amino acid sequences with the sequence of porcine ribonuclease inhibitor (RI), whose X-ray structure (PDB id: 2BNH) was used as a template. The two proteins for which the models were not built are NOD4 and NALP10. NOD4 LRR domain consists of 43 LRRs, and all LRR-containing proteins with available structures are too small to be used as a template. The second protein, NALP10, lacks an LRR domain.

Amino-acid sequence alignments were manually refined using the Swiss-Pdb Viewer [37] to account for additional information obtained from the gene structure and from structural comparisons between LRR proteins (see below). Because of the high sequence similarity between the template and the prediction targets (50-60%), homology modeling was performed by means of the SWISS-MODEL server [38].

Electrostatics and hydrophobicity calculations were carried out for NLR homology models as well as for experimental crystal structures of two TLRs. They were conducted utilizing MOE software (Chemical Computing Group).

In order to conduct a quantitative comparative analysis of LRR domain surface features in the absence of experimental structure data, we mapped physicochemical features of residues on the surface using existing knowledge about LRR sequence-to structure-mapping [31‐33].

The typical LRR sequence contains a conserved motif LxxLxLxxNxL, where L is leucine, isoleucine, valine, or phenylalanine, and N is asparagine, threonine, serine, or cysteine. In NLRs there is a striking correlation between their gene organization and amino acid sequence. Specifically, in NALPs, two consecutive LRR motifs XLaaNaLaaaaaaaaaaaaaaaaaLaaLXLaaNaLaaaaaaaaaaaaaaaaaaIaaL of length 28 + 29 = 57 residues are encoded by a single exon [3], while we have found that in NODs each LRR motif XLaaNaLaaaaaaaaaaaaaaaaaLaaL of length 28 residues is encoded by a single exon. This unusual genomic structure has interesting structural consequences, which would be a subject of a separate publication [39]. These two facts have allowed us to make reliable predictions for positions and boundaries of LRR motifs in human NLR sequences. In TLRs such an unusual structure-exon correlation is lacking, and the whole LRR domain is typically encoded by one large exon.

Finally, from the available structures of other LRR-containing proteins it is known that the aaLXLaa pattern corresponds to a beta-strand in the inner concave surface, which is then followed by a C-terminal side loop, by an alpha-helix on the outside convex surface, and by the N-terminal side loop. This property is very well conserved in numerous LRR-containing proteins, and we have utilized it to improve mapping between LRR sequence and structure. The residue X in the aaLXLaa motif was used as reference residue to define positions of all other residues on the LRR surface (Figure 3A). The next 5 residues are located on the C-terminal side (usually loop conformation), the next 12 residues are located in the outer convex part (usually alpha-helical or loop conformation), and finally the next 5 residues represent the N-terminal side (usually loop conformation). In TLRs, the typical length of the outer part is 9 residues, but there is much variation.

After we performed the sequence-to-structure mapping as described above, we were able to partition the surface into four parts and analyze distributions of amino acid properties such as charge and hydrophobicity within these parts separately. Distribution of potential glycosylation sites was also examined for extracellular TLRs.

Amino acid charge and hydrophobicity distributions within LRR domains were calculated for each of the four surfaces shown in Figure 3B. The amino acid charge was calculated assuming the pH of 7.0, and the hydrophobicity was quantified by hydropathy index ranging from -4.5 to +4.5 [40] with negative (positive) values corresponding to hydrophilic (hydrophobic) properties.

Within a given protein, for each LRR motif i, we calculated average charge (q _i ) and hydropathy (h _i ) separately for each of the four surface regions. For an LRR domain with n motifs, this produced 4 n values for charge and the same for hydropathy. Next, in order to characterize the distribution of these values within each surface, we have calculated the first four moments, m _k , (k = 0...3) of q _i and h _i distributions:

where f _i is either q _i or h _i , and

With these definitions, the moment m₀ is simply the average value, and the moment m₂ is related to the square of the standard deviation. The moments m₁ and m₃ reflect additional details of the distributions, in particular the directional trends (N-terminus being more/less hydrophobic or charged that the C-terminus).

The distributions of charge and hydropathy within one protein are thus each characterized by a sequence of 4 surface × moments = 16 values. In order to quantify similarities of distributions within different proteins, we used Spearman's rank-order correlation coefficient, r _S (it is similar to the usual Pearson's correlation coefficient, but reduces the effect of outliers). The similarity between distributions was then defined as (1 - r _S ). Clustering analysis was then performed by a complete linkage method utilizing Java Treeview software [41].

Consensus N-linked glycosylation sites were predicted by searching for patterns NxxS and NxxT in TLR sequences. In globular proteins, such predictions greatly overestimate the number of glycosylation sites, assuming that all sites matching the pattern above are accessible to oligosaccharyltransferases in order for glycosylation to occur, while in reality most of these sites are buried inside protein structure. In our case, however, such predictions can be much more accurate because we could predict which of the potential glycosylation sites were exposed to solvent.

As a first step, we have performed a phylogenetic analysis of LRR domains in NLRs and TLRs. Extreme sequence length diversity among LRRs from these families makes conducting such an analysis in a meaningful way nontrivial. This is further complicated by the fact that full length multiple alignment of complete sequences contains numerous large gap regions, which makes its truncation difficult. To circumvent the latter problem, we built a multiple alignment using a 10-fold increased gap extension penalty and truncated it from N- and C-terminal ends to make its length similar to that of the shortest sequence (see Additional File 1). The resulting alignment was then used to build a rooted phylogenetic tree (see Figure 1 and Methods). Among the NLR and TLR sequences, we also included a sequence of human ribonuclease inhibitor (RI) consisting of a highly-regular LRR domain closely homologous to that of LRRs in the NRL family.

The sequences group into three distinct clusters: (i) NALPs and RI; (ii) CARD-containing proteins and NLRX1; and (iii) TLRs. (Similar clustering was found when the multiple alignment was built using a default gap extension penalty and was used without truncation to infer a tree; data not shown.) Surprisingly, the sequence similarity-based phylogeny of the C-terminal LRR domain well agrees with the classification of all these proteins based on the N-terminal effector domain. Within NALPs, there are three distinct subgroups: (a) NALP2, NALP5, NALP7, NALP8, NALP14; (b) NALP4, NALP9, NALP11, NALP13; (c) NALP3, NALP12, and RI. It is interesting to note that IPAF and CIITA, as well as NLRX1 whose effector domain lacks definitive classification, group together with CARD-containing NODs. This suggests that NLRX1 effector domain might be CARD-like. This analysis supports known phylogenetic relationships between NLR and TLR families, but it does not permit us to draw any conclusions regarding similarities of surface features between TLRs with known ligands and NLRs.

In Figure 2 we present homology models for LRR domains of twenty NLRs (except NOD4 and NALP10) along with experimentally determined structures of two TLRs. The first observation is that there is much diversity in the LRR domain size - from 7 LRRs (in CIITA) to 19 LRRs (in NALP5). An unusually large number of LRRs (43 LRRs) is found in NOD4, whose model was not built because there is no large enough template structure available (see Discussion). Second, coloring of the models according to surface electrostatic potential and hydrophobicity reveals significant diversity of their surface features.

The models in Figure 2 are helpful in visualizing LRR domain structures and could be used for explicit protein-ligand docking; however they are not suitable for quantitative assessment of surface similarity. In order to quantitatively analyze the similarity of surface features of LRR domains, we resorted to their reduced-resolution representation. First, we adopted a simple approach to implicitly infer LRR domain three-dimensional structures from their amino acid sequences (see Figure 3A and Methods). Second, we have partitioned the total surface of the LRR domain of each protein into four regions (see Figure 3B and Methods): concave inner surface (mainly beta-strands), N-terminal side, C-terminal side, and convex outer surface (mainly alpha-helices in NLRs and RI, and loops or beta strands in TLRs). Before conducting a quantitative comparison of NLR and TLR surface features we had to consider possible perturbations to TLR surface properties due to glycosylation.

It is well known that LRR domains in all ten human TLRs are N-glycosylated due to their extracellular localization, and in some of them glycosylation is essential for proper function [42, 43]. This is not surprising as glycosylation can dramatically alter surface properties and thereby affect ligand binding. Since NLRs are all known to be cytosolic proteins their N-linked glycosylation is unlikely, which complicates a direct comparison of their surface features to those of TLRs. To understand the potential role of TLR glycosylation in their ligand binding, we have performed a search for consensus N-linked glycosylation sites within TLR sequences, and analyzed their distribution over the domain surfaces. The results are summarized in Table 1.

The TLRs that contain many consensus glycosylation sites are TLR3, TLR7, TLR8, and TLR9. In TLR3, all surfaces except the C-terminal side that is known to bind dsRNA contain several glycosylation sites. In TLR7, inner and outer surfaces are likely to be glycosylated while side ones are not, which is consistent with binding of ssRNA via one of its side surfaces. Glycosylation of TLR8 and TLR9 resembles that of TLR3 in that only the C-terminal side surface is glycan-free. In the rest of the TLRs, each of the four surfaces contains no more then three potential glycosylation sites. As an additional step, we have also analyzed NLR sequences for potential glycosylation (data not shown), and found that they contain 0-2 consensus glycosylation sites per surface, with the only exception being an N-terminal surface of NOD3, which interestingly contains 10 such sites.

These findings indicate that while glycosylation of some TLRs certainly modulates their surface properties, in most cases the ligand binding sites are located in non-glycosylated regions and, in addition, the overall surface area affected by glycosylation is relatively small, so that global surface comparisons such as those performed here are not affected.

The two physico-chemical amino acid properties that critically affect ligand binding and whose distributions we have thus analyzed are charge and hydrophobicity. Figure 4 shows an example of our clustering analyses of hydropathy distribution moments within each of the four domain surfaces. The similarity measure in this analysis is Spearman's correlation (see Methods). Based on clustering results, LRR domain pairs that have particularly similar hydropathy distributions and LRR numbers are: TLR5 and IPAF (both respond to bacterial flagellin, although for IPAF this is under debate [25]); NOD2 (ligand: MDP) and NALP8 (ligand unknown); RI (ligand RNase) and NALP7 (ligand unknown); TLR6 (ligand: diacyl lipopeptide) and NAIP (ligand unknown); NALP4 (ligand unknown) and NALP14 (ligand unknown); TLR1 (ligand: triacyl lipopeptide) and TLR10 (ligand unknown); TLR3 (ligand: dsRNA) and NALP6 (ligand unknown).

Figure 5 presents results of a similar clustering analysis of charge distributions. The domain groups with similar hydrophobicity distributions are: RI (ligand: RNase) and NALP9 (ligand unknown); NALP3 (ligands: RNA, small molecules) and NALP2 (ligand unknown); TLR8 (ligands: ssRNA, imidazoquiniolines) and NAIP; TLR3 (ligand: dsRNA), TLR7 (ligands: ssRNA, small molecules), and NALP6 (only 8 LRRs; ligand unknown).

In order to gain further insight into similarities of LRR surface features between NLRs and TLRs, we conducted a careful manual inspection of similarity patterns across individual biologically relevant surfaces, also taking into account results of our TLR glycosylation analysis (cf. Table 1). We assessed similarities between LRR domains of all 32 proteins considered, as detailed below. (i) In NALP2, which is known to sense MDP, the concave and side surfaces have both hydropathy and charge distributions that are the most similar to NAPL3, which is known to bind a diverse set of ligands, including MDP. (ii) NALP4 and NALP14 have similar hydropathy distributions, but dissimilar charge distributions. (iii) NALP5 and NALP8 have hydropathy and charge distributions similar to NOD1 (except over the outer convex surface) that binds iE-DAP. (iv) NALP6 has the shortest LRR domain, containing only 6 LRRs. Its surface properties are similar to those of TLR3, which contains 25 LRRs. Because the hydropathy and charge distribution moments are normalized by the number of LRRs in a given domain, similarity of distributions still allows to hypothesize that NALP6 and TLR3 may bind the same type of ligands, RNA motifs. (v) NALP7 charge and hydropathy are similar to NOD2 and NALP3 both binding MDP, suggesting small molecules as probable ligands. (vi) The high similarity of NALP8 hydropathy distribution and size to those of NOD2 and also NOD1 (whose convex surface hydropathy is different) suggests that they may share the same type of ligands--peptidoglycans. This is also supported by similarities in their charge distributions, with the difference being a more positively charged convex surface of NALP8. (vii) The hydropathy and charge of NALP9 are similar to RI, which binds RNase, thus suggesting that the NALP9 ligand may be protein-like. (viii) NALP11 shows hydropathy similarity with TLR7 that binds ssRNA, most likely by one of its side surfaces. NALP11 charge distribution is similar to that of TLR7 and also TLR3 (which binds dsRNA at its side surface) over its side surfaces, suggesting that ss/dsRNA is likely to be among possible NALP11 ligands. (ix) NALP12 surface properties do not exhibit any obvious similarity to others. (x) The hydropathy distribution of NALP13 is similar to that of NALP7 and NOD4, but neither of them has known ligands. (xi) NALP14 charge distribution on its concave surface is almost identical to that of IPAF, and distributions over other surfaces are similar to IPAF. The hydropathy distributions of these two NLRs are similar within only its concave side. Assuming that IPAF binds flagellin similar to TLR5, i.e., on its concave surface [44], flagellin can be considered a likely ligand of NALP14. (xii) NOD3 hydropathy has some similarity to NALP3, but charge distribution is unlike that in any other domain. (xiii) NOD4 concave and side surfaces have hydrophobicity similar to NALP12, but the charge distribution is different. Also, NOD4 appears to have the largest ligand-binding domain of all NLRs and TLRs: it contains as many as 43 LRRs. (xiv) NLRX1 surface features do not exhibit similarity with any other domain. (xv) NAIP hydropathy distribution is similar to RI, while NAIP charge distribution is similar to TLR8.

It is interesting to note that the functional similarities inferred based on comparison of molecular surfaces correlate well with experimental data where available, while hypotheses based on evolutionary relationships shown in Figure 1 generally lack such correlation. For example, while the surface comparison and the known experimental data suggest that NALP2 and NALP3 are activated by the same type of molecules, MDP, the phylogenetic reconstruction places NALP2 and NALP3 into two distinct subgroups in Figure 1, although both subgroups belong to the NALP group. Similarly, whereas IPAF and TLR5 are known from experiment to sense a common molecule (bacterial flagellin) with which our surface analysis is in agreement, in the phylogenetic analysis they are placed into two distinct groups. The closest agreement between surface-based and phylogeny-based functional inferences appears to be between NALP5 and NALP8.

The above surface similarity considerations, which we believe are the most interesting results of this work, are summarized in Table 2. For each NLR with unknown ligands, we indicate to which other TLRs/NLRs its charge and hydropathy distributions are the most similar, and based on this infer the molecules that are likely to be its ligands. We would like to stress that because of low resolution of our approach, the suggested similarity-based predictions for putative NLR ligands should be considered with care.