Characterization and prediction of residues forming functional protein-protein interfaces using network analysis
Characterization of protein-protein interfaces can provide understanding of the fundamental properties of their functionality. Additionally, such understanding aids functional annotation of the interfaces and prediction to help in drug design. In this work, we classified protein interfaces broadly into two functional categories: Functionally Linked Interfaces of Proteins (FLIPs) and Functionally uncorrelated Contacts (FunCs). We used a method, based on network analysis followed by statistical analysis, to classify protein interfaces based on their functionality. Additionally, the method used for classification of protein interfaces characterizes each subunit in the complex individually allowing to expand the use of our method towards prediction of functionally linked interfaces. We assessed two different types of network constructs, as well as a method combining the best features from both the network constructs, for their classification accuracy and ability to predict functional interfaces.
In the first network construct, Residue Interaction Networks (RINs), amino acid interactions are represented by spatial distance cut-offs of 8 Å between two Cα atoms of the residues. This simple approach classified FLIPs from FunCs with a consistent accuracy of 72%. It also identified dissimilarities among FLIP and FunC organizations. The positive correlation of FLIPs and network organization features suggest that FLIPs have a surface rich in concavities, while FunCs have smooth surface. The second network construct, Protein Energy Network, identified interactions among the amino acid residues using PyRosetta generated interaction energy scores. This construct was designed to assimilate chemical information in addition to geometrical information based RIN construct to study energy influenced organization of interfaces. A 70% classification accuracy was achieved with this method; however, the compromise in overall accuracy is balanced by accuracies of categorization of a few functional subcategories such as structural interfaces were classified with 100% accuracy. Finally we combined the identified spatial and energetic organization features from both the constructs and analyzed the classification potential of the combined model. It achieved an accuracy of 73% in classifying FLIPs from FunCs. Analysis of organization features from both the constructs, and also, the combination using Receiver Operating Characteristic (ROC) Curves identified that the classification abilities of combined organization features can be translated to a better prediction method among the three. Predictions of functionally linked interfaces of protein showed that 60% of the observed interfaces were correctly predicted, along with identification at least one connected triplet in 87% of the FLIPs. However, a 75% over prediction rate was observed. Analysis of FunCs on the other hand showed that we identified a connected triplet in 70% of the proteins, but rarely predicted functionally uncorrelated interface residues. Overall, our work suggests that network analysis can successfully characterize, classify and predict functionally linked interfaces of proteins.