A successful phage infection usually depends on an interaction between a receptor-binding protein (RBP) and a bacterial receptor. In Klebsiella pneumoniae, some phages use depolymerase-containing RBPs to degrade capsule polysaccharides. Previous studies have shown that phage RBPs in Klebsiella are highly diverse and very specific towards the bacterial sugar. However relatively little is known about the diversity of these RBPs at the structural level. Here we hypothesised that structural modelling using AlphaFold2 can improve our understanding of the nature of RBP-sugar specificity by identifying protein folds associated with recognition of specific polysaccharides.

To address it, we constructed a dataset of 36 GenBank Phage RBPs specific to 23 different capsular serotypes in K. pneumoniae (12 serotypes have more than one RBP representative) and whose specificities have been experimentally validated via production of recombinant proteins. We generated homotrimeric structural predictions of all RBPs and, where possible, identified three domains manually using PyMol: N-terminal domain, a central (enzymatic) domain and a C-terminal domain. We then compared all sequences, structures and their domains to each other using BLAST, US-align and Foldseek. We assessed structural similarity between protein pairs based on TM-score (TM-score<0.5 unrelated, TM-score≥0.5 related, TM-score>0.9 structurally identical).

Our results point to several observations. First, RBPs are highly diverse genetically. We found that while only 28% of RBP pairs had detectable similarity at the sequence level, 73% of them were related on the structural level. This suggests that many related RBPs would not be detectable at the sequence level. 

Second, structural modelling can help predict RBP-sugar specificity. Specifically, RBPs specific to the same capsule types were structurally more similar to each other than RBPs without confirmed specificity to the same capsule (Wilcoxon-rank test, average TM-score = 0.77 vs 0.55, p<10^-4). This association was even stronger when the comparison was carried out at the level of central-domains (TM-score 0.89 vs 0.72, p<10^-5) and C-terminal domains (TM-score 0.82 vs 0.34, p<10^-5), both of which are known to confer host-specificity. No significant association was observed at the level of N-terminal domain (p>0.5).

Third, structural modelling can help identify potentially interesting biological and biochemical properties of RBP-capsule interactions. One example were two unrelated structures (TM-score=0.31) of RBPs specific to K47. We found that these structures had been shown to be specific to different subsets of serotype K47 K. pneumoniae strains, suggesting potential differences in capsular sugar compositions. Another example were near-identical structures of RBPs specific to K11 and K13/K2 (TM-score=0.87). These capsule types were found to have α/β-D-Glcp-(1→3) residue in common, suggesting that it might be targeted by the depolymerase.

Finally, we saw that phage morphotype was a significant but an imperfect predictor of N-terminal diversity. While structural similarity of N-terminal domains was greater amongst phages assigned to the same morphotype (average TM-score = 0.58 vs 0.44, p<10^-5), we found multiple examples of structurally identical N-terminals belonging to phages of different morphotypes.

Overall, our results show the importance of applying structural modelling and sequence analysis to study the diversity of RBPs and to understand the evolutionary processes leading to capsular specificity.