ABSTRACT
Specific protein interactions are responsible for most biological functions. Distinguishing Functionally Linked Interfaces of Proteins (FLIPs), from Functionally uncorrelated Contacts (FunCs), is therefore important to characterizing these interactions. To achieve this goal, we have created a database of protein structures called FLIPdb, containing proteins belonging to various functional sub-categories. Here, we use geometric features coupled with Kortemme and Baker's computational alanine scanning method to calculate the energetic sensitivity of each amino acid at the interface to substitution, identify hotspots, and identify other factors that may contribute towards an interface being FLIP or FunC. Using Principal Component Analysis and K-means clustering on a training set of 160 interfaces, we could distinguish FLIPs from FunCs with an accuracy of 76%. When these methods were applied to two test sets of 18 and 170 interfaces, we achieved similar accuracies of 78% and 80%. We have identified that FLIP interfaces have a stronger central organizing tendency than FunCs, due, we suggest, to greater specificity. We also observe that certain functional sub-categories, such as enzymes, antibody-heavy-light, antibody-antigen, and enzyme-inhibitors form distinct sub-clusters. The antibody-antigen and enzyme-inhibitors interfaces have patterns of physical characteristics similar to those of FunCs, which is in agreement with the fact that the selection pressures of these interfaces is differently evolutionarily driven. As such, our ECR model also successfully describes the impact of evolution and natural selection on protein-protein interfaces. Finally, we indicate how our ECR method may be of use in reducing the false positive rate of docking calculations.
