Study On Protein - Vitamin Binding Site Prediction Based On Unbalanced Learning

Vitamins are important cofactors in various enzymatic-reactions, and they are essential organic compounds for human metabolic activity. Enzymes are a kind of biological macromolecules which have biological catalytic function, they are mainly composed of proteins, and they play important roles in catalytic action of biochemical changes.In the body of healthy people, vitamins attend to the biochemical reactions through the binding interaction between vitamins and enzyme molecules. But for patients, the interaction of enzymes and vitamins can not proceed smoothly. These enzymes has become drug targets for many diseases, so prediction of protein and vitamin binding sites has very important significance of biological pharmaceutical industry. The traditional laboratory determination of protein and vitamin binding sites has many shortcomings of time-consuming, laborious and costly. Thus, the traditional laboratory determination methods cannot meet the urgent needs of the present industry development. Using intelligent calculation methods to predict protein and vitamin binding sites can improve the efficiency of the experiment, so this kind of research has important significance now.In this paper, we study the protein and vitamin binding sites prediction problem, because of the huge difference between the number of binding sites and non-binding sites, this problem become an imbalanced learning problem. Combining the characteristics of imbalanced learning problems and the particularity of protein and vitamin binding prediction. This paper proposed an improved under sampling method which bases on K-means algorithm, it is called MUS method, and we combine MUS method and classifier ensemble technology to set up a model for protein and vitamin binding sites prediction, called: MUS Vita AdaBoost. Because of the difference of the interaction between different kinds of vitamins and proteins, here we do experiments for Vitamin A, Vitamin B, Vitamin B6 and vitamins which have many kinds respectively. And during our experiments we used the combined features of LogisticPSSM and PSS. Through those experiments, MUS_Vita_AdaBoost method achieved good results in our experiments. And, preparing with other models, the model we proposed achieved a better generalization performance.