Aromatic clusters in protein–protein and protein–drug complexes

To analyze the structure and prevalence of aromatic interactions in protein–drug (PD) and protein–protein (PP) complexes, we surveyed the PDB and the complexes were filtered to have only one for each pair of protein–protein and protein–drug complexes. We found and analyzed 10,231 protein–drug complexes and 4837 protein–protein complexes to identify the presence of protein–drug/protein–protein aromatic clusters. Aromatic clusters are defined by the presence of at least two (i.e. a dimer), aromatic rings that are interacting according to Materials and Method based on our previous work [19]. In the present case, we looked for clusters that have at least one ring on a different molecule. For detailed description and characterization of intra-protein aromatic clusters see our previous work [19]. The number of complexes with at least one aromatic cluster is 5908 in the case of protein–drug (57% of all protein–drug complexes). Amazingly, if we consider the complexes where the drug has an aromatic ring (ca. 66% of the drugs in the dataset fulfil this criterion), 87% of them are forming aromatic clusters. In the protein–protein complexes, we found 3048 with at least one aromatic cluster, which represents 63% of the dataset. The total number of clusters was 7236 (protein–drug) and 7717 (protein–protein), this means that protein–protein interaction sites have more than twice the number of clusters compared with drug binding sites, having 2.53 against 1.22 clusters per cluster. Accordingly, the total number of interactions found was 23,303 and 15,309 for protein–drug and protein–protein, respectively, which correspond to an average of 3.22 and 1.98 interactions per cluster for each case. (Table 1) Concerning cluster size, dimers are, as expected, the predominant type but trimers, tetramers and beyond are well represented.

Interestingly, dimers are over-represented in protein–protein (59%) compared to protein–drug (31%) complexes (Additional file 1: Figure S1A). As expected, the number of interactions increases with the cluster size quite evenly for both datasets for small cluster sizes. Nevertheless, for big clusters (≥ 6), the number of interactions is higher for the protein–drug group than the protein–protein group (Additional file 1: Figure S1B).

For the sake of comparison, we also look for intraprotein clusters in the protein–drug dataset, and found 9760 clusters (shown in Table 1), which represents 95% of the dataset. As expected, the number of clusters per complex is higher than in the other two cases (protein–drug and protein–protein) because of the volume analyzed. Protein cores are bigger than protein–protein interfaces and drug binding sites. However, the average number of interactions in intraprotein clusters is comparable to that of protein–drug complex clusters, and both cases are around double that in the protein–protein clusters.

As aromatic trimmers adopt two different conformations in space, Symmetric (Sym) and Ladder (Lad) we analyzed how they behave in Protein–drug and Protein–Protein interactions (see Additional file 1: Figure S4). In particular, most of the protein–drug timers (78%) are in the Ladder structure and the same is for protein–protein Ladder trimers (79%). In both datasets, the number trimers seem to be equally balanced in terms of structural choice. (Additional file 1: Table S1).

To characterize the structure of each cluster we determined two key parameters (Fig. 1a): The distance between the centre of each aromatic ring and the planar angle (α), which is the angle formed between the planes of the two rings. We compared aromatic interactions in protein–drug and protein–protein complexes, as well as those found inside proteins, in terms of both parameters. In all cases, the plot is very similar (Fig. 1b), with a broad wide peak at ca. 5 Å, which is the optimal average distance between two aromatic rings, in a T-shape like orientation (optimal angle is ca. 75°, see Fig. 1c and d). Surprisingly, all distributions also show a minor narrow peak at ca. 3.75 Å, but the relative intensity of it is different in all three cases. It is small in intra-protein clusters, slightly larger in protein–protein and largest in protein–drug complexes. This peak corresponds to the distance of optimum π-stacking interaction according to McGaughey [26], and therefore our data suggest that stacking interactions are favoured in protein–drug complexes. This is confirmed by looking at the planar angle vs. distance plot for protein–drug and protein–protein (Fig. 1c and d) showing that π-stacking conformations, a planar angle close to zero degrees, are enriched in protein–drug compared to protein–protein complex.

In Fig. 2a, we compared the solvent exposure (or SASA) percentage for the rings containing residues in each group. As expected, aromatic residues inside the protein core are barely exposed. While residues in protein–drug are more exposed than protein core residues and the exposure of protein–protein residues is more significant and shows a wider range of solvent-accessible surface. In Additional file 1: Figures S2A and S2B, we compare the relationship between the solvent-accessible surface and the contact surface to understand how much surface is devoted to the aromatic interaction. Although we observed a solvent accessible surface spanning a wide range of values, the contact surface is mainly distributed at high values with the maximum of the distributions at 100% in both sets which means that the aromatic residues use all the available surface to establish the clusters. Figure 2b shows how the interacting rings of the clusters are distributed in the different secondary structure elements. This panel shows that aromatic residues interacting with drugs have a preference to be in loops and less likely to be in a structured region. The same is true for residues forming protein–protein complexes, residues in loops have a strong preference to be interacting with residues in loops from the partner protein.

Structural properties of aromatic interacting residue and their secondary structure preference. a Percentage of exposed surface distribution for residues in protein–drug clusters (solid), protein–protein clusters (dashed) and intra protein clusters (dotted) b Secondary structure preference. On top, the three preferences for interacting residues in protein–drug, and below, the preferences for residues in protein–protein

Moreover, in protein–protein, secondary structure matches between interaction partners. Residues in sheet or helix have a preference to contact residues in the same secondary structure (sheet with sheet and helix with helix). These residues also present an inverse preference to be in contact with residues in the other secondary structures. Although the preference is slightly higher for sheet conformation, as aromatic trimmers have been reported to stabilize the structure of beta-sheets [32].

We now turn our attention to the relationship of different aromatic cluster properties and ligand affinity in the protein–drug complexes. We looked for binding affinity information using BindingDB [31] assigning (when possible) the Ki associated with a given complex in the protein–drug dataset. This assignment resulted in 358 protein–drug complexes with annotated binding affinity which have aromatic clusters with more than 2 ring interactions and with a resolution higher than 2.5 Å. First, in Fig. 3a, we display violin plots showing how Ki distributions vary as a function of the number of clusters (1 or 2) and the total number of ring interactions. First, as expected, when we look at complexes having only one cluster, the average Ki slightly decreases as the number of interactions increases. The trend is, however, not found when two clusters are present. Thus, no improvement in affinity is observed by adding more interactions. Also, comparing complexes with two or three interactions, if these are established using two clusters (Fig. 3a) the average Ki is improved about ten times to those cases having one cluster (Fig. 3a). In other words, it is better to anchor the drug through more than one cluster (or interaction site). However, this is only possible when the drug displays more than one aromatic ring.

Aromatic clusters and binding affinity. a From left to right, violin plots are displayed increasing the number of interactions alternating (2 or 3, 4 or 5, 5 or more interactions) between drugs having 1 cluster and drugs with 2 clusters. Ki is expressed in molar units and y-axis is displayed in logarithmic scale. b Violin plot of Binding affinity of drugs interacting with a residue having less and more than 20% of the exposed surface. c Violin plot of drug binding affinity for drugs interacting with aromatic residues in different secondary structures

As shown in Fig. 3b, we found that residues that are more than 20% exposed, interact with drugs having significantly lower binding affinity than those exposed less than 20% (Fig. 3b). Example conformations can be seen in Additional file 1: Figure S3A and S3B.

Now, we can ask if these properties have some relation with binding affinity to drugs.

Interestingly, as shown in Fig. 3c, we found that the secondary structure presents great differences. Residues in helix conformation (Additional file 1: Figure S3D) have weaker binding affinity than residues in sheets (Additional file 1: Figure S3C) and loops (Additional file 1: Figure S3E), and surprisingly residues in sheets have significantly better binding affinity than the other two cases. This difference contrasts heavily with the fact mentioned before that drugs have a negative preference for contacting residues in sheet conformation.