The Research Of The Method To Predict Hot Spots In Protein-protein Interface

The most of the binding free energy of protein interaction is contributed by a few keyresidues, which were called hot spots. Hot spots are crucial for understanding the function ofproteins and studying their interactions. The most popular method of detecting hot spots isalanine scanning mutagenesis experimental, but this method is not applicable on a large scalesince it is time consuming and expensive. Therefore, reliable and efficient computationalmethods for identifying hot spots are greatly desired and urgently required.In this thesis, we introduce a hybrid SVM model which based on features of protein and thealgorithm of support vector machine, that can predict hot spots in protein interface efficiently.Firstly, we obtain the training set from the alanine scanning energetic database (ASEdb) forexperiment, and then we extract60protein interaction features, containing Physicochemicalfeatures, structure features and other relative features. We divide the60features into five classes,using feature selection respectively to obtain five subset features and then build five SVMmodels. By evaluating these five prediction models, we find two of them which can predict hotspots effectively. We build a hybrid SVM model by combining with the two SVM models whichcan further promote the prediction performance. In order to further validate our hybrid SVMmodel, we extract independent test set from the binding interface database (BID) to evaluate ourmodel.We compare our prediction model with previous ones, including Robetta model, FOLDEFmodel, KFC model and MINERVA model, these models have a very important significance inthe field of predicting hot spots. With the same training set and the test set, our model appears tobe higher prediction ability in predicting hot spots, which improve the applicability of prediction.