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

Protein-protein interactions play a critical role in nearly all aspects of cellular functions such as metabolism and signal transduction. Studies of molecular mechanisms for protein-protein interactions revealed that only a small subset of binding interface named hot spots account for the majority of binding free energy and are critical for stability and function of protein association. The identification of hot spots, a small subset of protein interfaces that accounts for the majority of binding free energy, is becoming more important for the research of drug design and cancer development.At present,identifying hot spots by biology experiments cannot be applied in a large scale due to its cost. So, it is in emergent need to predict hot spots by the method of machine learning.Here we proposed a novel hot spot prediction method. For each hot spot residue, we firstly constructed a wide variety of 108 sequence, structural, and neighborhood features to characterize potential hot spot residues, including conventional ones and new one (pseudo hydrophobicity) exploited in this thesis. We then selected 3 top-ranking features that contribute the most in the classification by a two-step feature selection process consisting of minimal redundancy maximal relevance algorithm and an exhaustive search method.We used support vector machines to build our final prediction-model. When testing our model on an independent test set, our method showed the highest Fl-score of 0.70 and MCC of 0.46 comparing with the existing state-of-the-art hot spot prediction methods. Our results indicate that these features are more effective than the conventional features considered previously, and that the structural and physicochemical properties are important determinants of hot spots.