Predicting Protein-protein Interactions Based On Domain Combination Information

Protein domains are the structural or functional units that participate in intermolecular interactions. Recently, domain-based protein-protein interaction prediction has been studied by several research groups. Most of the methods used in those researches infer domain-domain or domain combination pairs interacting information from statistics which usually suffer from lack of training sets to evaluate the interacting probability of each domain or domain combination pair. Thus, not all of the protein pairs can be judged to interact or not.In this paper, a new domain-combination based method is proposed to predict protein-protein interactions. After feature vector of each domain pair based on its amino acid sequence information, SVMs is used to construct classify model to predict interactions of domain combination pairs, and statistical calculating method is used to analysis their appearance frequencies. Then interaction possibilities of domain combinations are calculated by integrating the two features and protein-protein interactions are predicted based on them.In the prediction of yeast protein pairs, sensitivity (70%) and specificity (62.8%) are achieved in our model. Especially, for protein pairs with high-frequency domain combination pairs, high sensitivity (96.77%) and specificity (78%) are achieved, which is better than PreSPI system proposed by Han et al. Results show that the method performs well in the prediction of domain combination interactions and protein-protein interactions.A computational tool, PPIPred, is developed based on this strategy, and it can be accessed via http://infosci.hust.edu.cn.