Etkileşim tahmini için protein-protein etkileşim kümelerinin ve niteliklerin detaylı karşılaştırılması

Protein-protein interactions (PPI) are of crucial importance at all levels of biological processes. The experimentally identified PPI are deposited in several databases. These databases contain diverse information about PPI; but their coverage is low when we consider full processes in cells. Thus, reliable, accurate computational methods are needed to improve the coverage. Many research groups have developed PPI prediction algorithms with varying accuracies based on different data and methods. However, to develop a new PPI prediction method with high accuracy is challenging.This study aims to assess existing sequence based PPI prediction methods and to propose a new algorithm with improved accuracies. The predictions are made via Support Vector Machines (SVM), which is a machine learning algorithm. SVM creates models based on training sets and predicts interactions via those models. In this study, positive training sets contain experimental PPI and negative training sets contain computational non-interacting proteins. In order to represent interaction data in SVM, n-gram frequencies of proteins are calculated according to their amino acid sequences. It is shown that SVM performance is strongly affected by interactions in training datasets, amino acid categorization techniques, n-gram frequencies, and ? values used. SVM models are created for eight datasets and the critical assessment of those datasets is made via their SVM scores. Based on those scores, combined training datasets are created that make accurate prediction of interactions in every dataset. Then, the best feature set that leads to the highest SVM scores is found. Finally, the best SVM models are utilized to eliminate false positives in putative protein interactions predicted by PRISM (Protein Interactions by Structural Matching) algorithm.