Modellenmiş protein-protein etkileşim arayüzlerinin 3 boyutlu evrişimli sinir ağları ile değerlendirilmesi ve sıralanması

Biological processes depend on protein-protein interactions that occur through protein-protein interfaces. Identification of protein-protein interface regions is crucial to understand mechanisms of protein binding and predict new interactions. Therefore, it is critical to be able to determine protein-protein interfaces easily and reliably. To extract a protein interface, protein-protein complex structure is needed. Since even determining a single protein complex structure experimentally takes time and effort; it is promising to see that computational docking tools can provide thousands of possible complex structures, called decoys, in a short time. A challenge that comes with computational methods is to be able to discriminate near-native decoys among thousands proposed. Multiple methods have been developed for scoring and ranking the decoys to identify the biologically relevant ones that can be used in further biological research. In this thesis, we improved a three-dimensional convolutional neural network-based decoys` scoring and ranking approach called DeepInterface. We compared multiple convolutional neural network architectures and searched the hyperparameters` space to find the best performing DeepInterface model that resulted in having the VGG16 resembling architecture. We built positive datasets from the complexes stored in the Protein Data Bank, and negative datasets from the incorrect decoys stored in the PPI4DOCK and DOCKGROUND docking databases. We showed that the model we suggested can discriminate positive interfaces, similar to ones stored in the PDB, from the incorrect ones, according to the CAPRI criteria, with approximately 81% accuracy. We also compared our models with other decoys` scoring/ranking tools including IRAD and DOVE using ZDOCK docking benchmark 4.0 decoys and showed that our models are competitive. This method could be used to reduce the computational cost of the protein-protein interactions` predictions.