Structural Characterization and Function Prediction of Immunoglobulin-like Fold in Cell Adhesion and Cell Signaling

Domains that belong to an immunoglobulin (Ig) fold are extremely abundant in cell surface receptors, which play significant roles in cell-cell adhesion and signaling. Although the structures of domains in an Ig fold share common topology of β-barrels, functions of receptors in adhesion and signaling are regulated by the very heterogeneous binding between these domains. Additionally, only a small number of domains are directly involved in the binding between two multidomain receptors. It is challenging and time consuming to experimentally detect the binding partners of a given receptor and further determine which specific domains in this receptor are responsible for binding. Therefore, current knowledge in the binding mechanism of Ig-fold domains and their impacts on cell adhesion and signaling is very limited. A bioinformatics study can shed light on this topic from a systematic point of view. However, there is so far no computational analysis on the structural and functional characteristics of the entire Ig fold. We constructed nonredundant structural data sets for all domains in Ig fold, depending on their functions in cell adhesion and signaling. We found that data sets of domains in adhesion receptors show different binding preference from domains in signaling receptors. Using structural alignment, we further built a common structural template for each group of a domain data set. By mapping the protein-protein binding interface of each domain in a group onto the surface of its structural template, we found binding interfaces are highly overlapped within each specific group. These overlapped interfaces, we called consensus binding interfaces, are distinguishable among different data sets of domains. Finally, the residue compositions on the consensus interfaces were used as indicators for multiple machine learning algorithms to predict if they can form homotypic interactions with each other. The overall performance of the cross-validation tests shows that our prediction accuracies ranged between 0.6 and 0.8.