Abstract: Antibodies have been very extensively investigated for decades and therefore much is known about their sequence-structure-function relationship. The ability of accurately modeling the structure of antibodies stems from the recognition that the hypervariable loops only exhibit a limited number of main-chain conformations called “canonical structures”. Most sequence variations in five of these loops only modify the surface generated by the side chains on a canonical main chain structure. The third loop of the heavy chain (H3) has a different behavior, and has revealed to be very difficult to model given its high variability in both length and structure. We applied a machine-learning approach combining sequence and structural related features to identify candidate loops as templates to build the structure of a target H3 loop. Models are subsequently ranked based on a score reflecting the likelihood of the presence/absence of specific interactions between the H3 residues and its structural environment. The method has led to a significant improvement in the prediction of the H3 region and the overall antigen-binding site.

We next analyzed how differences between antigen-binding sites might be linked to their specificity. To this purpose, we developed a superposition free method for comparing the surfaces of antibody binding sites based on shape descriptors. We showed that similar antigen-binding sites could be better detected based on shape descriptors than using traditional structure similarity metrics. Finally, we showed that a classification procedure based on this approach could be applied to derive information about the recognized antigen, representing a step towards the very elusive goal of predicting antibody specificity.

Venue: Aula de Teleensenyament (UPC)

Click here to see the full presentation