The Research Of Predicting Hot Spots At Protein-Protein Interface Based On ELM

As the material basis of all lives,protein plays an significant role in life science research.Among them,its structure and function are two vital-focused aspects.The process of protein interaction is closely related to these two aspects.Therefore,it is necessary to study this process.There are a large number of residues on Protein-protein interaction interface,and a small part are very important to the function of the interface,which called Hot spots.Predicting hot spots can further reveal the nature of the interaction between proteins.In addition,it also helps us to understand the structure and function of protein.In this paper,we obtain the training data set and independent test data set from the Protein Data Bank according to some previous relative papers.Themainly traditional biological method of predicting hot spots is alanine scanning mutagenesis technique,this method is not applicable on a large scale since it is quite complex,costly and time-consuming.Recently,researchers have introduced several methods to predict hot spots at protein-protein interfaces based on some efficient data mining and learning algorithms,which called computational method.How to select the effective characteristics of residues is the most important research object in the residue prediction.In this paper,we first select 62 kinds of structural information features of protein residues based on the training data.And on this basis,we design a multiple steps features processing method to choose important features,including the ReliefF feature selection algorithm,the strategy of removing redundant features and choose important features.We obtain 15 kinds of features from the all features above as important features set finally.Then,based on the Extreme Learning Machine and different subsets of the 15 kinds of features,we can construct different prediction models.According to the experiment results,we can assure the combination of finally features subset containing 4 features and the best parameter of ELM.Finally,according the parameter and features obtained,we can establish a prediction model that can predict the hot residues correctly,and introduce a voting strategy to optimize the model in the process of classification.In order to verify the validity of our prediction model,we conduct a prediction experiment based the independent test data set and other case data sets.The experimental results show that our prediction model is effective to classify the residues,and has better performance than other prediction models.