Prediction Of Protein-protein Interaction Sites Based On MRMR And IFS

Byinteracting with other proteins,compounds,RNA and DNA,proteins play critical roles in nearly all biological events.Understanding of protein-protein interaction sites is the basis of the understanding of molecular recognition process.Protein rarely act alone,in many cases is through a part of the biological network function.Protein interactions are very important toalmost all aspects of cell functions,including regulation ofsignaling and metabolic pathways,protein synthesis,DNAreplication and gene translation,as well as immunologicalrecognition.Identifying the protein interaction sites would provide valuable clues tounderstand and determine the functions and structures ofprotein complexes,could contribute to the logic of pharmacological target identification and ultimately to provide support for drug design.Therefore,the prediction of protein-proteininteraction sites is of great significance.Two kinds of existing method was divided into method of predicting protein interaction site biological experiment method and calculation methods,the biological experiment method is costly and time-consuming,and first using calculation method to predict again with biological experiment method validation can greatly reduce the cost.In this work,according to a Random forest(RF)algorithm,which have a Minimum-Redundancy Maximal Redundancy(mRMR)method,which followed by incremental feature selection(IFS),a new method to predict protein-protein interaction sites has been discovered.Sequence features,structure features and 3D structural features are used and three different kinds of feature spaces are built by us.From the analysis,we can find that 3D structural characters such as ASA,SC are conducing most to the prediction.Compared with three models of feature space,effect of sliding window is relatively better.During the experiment,our own predictor performed better than any other current predictors.The accuracy in doing one prediction was 81.44%,sensitivity was 82.17%,and specificity was 80.71%.