| Protein is the most basic material constituting organism and the carrier of life activity,all life activities are performed through the function of proteins.For example,extracellular matrix proteins(ECMP)actively promote basic cellular processes such as differentiation,proliferation,adhesion,migration and apoptosis.Research clearly shows that extracellular matrix proteins play a major role in cell adhesion,proliferation and morphogenesis.Fertility-related proteins regulate the developmental events(spermogenesis,oogenesis)and differentiation process(embryogenesis,organogenesis)of germ cells.The phage virion proteins(PVP)are the basic components of infectious virus particles,and play an important biological role in the interaction between phage and host cells.Therefore,the study of protein functions and its related issues is of great significance.In recent years,the rapid development of biological science and technology has led to the explosive growth of biological information.At present,methods and tools for protein function analysis need to be updated urgently.Rapid and accurate prediction of protein function is one of the most challenging tasks at the moment.In order to further enhance and improve the accuracy and performance of protein function prediction methods,based on multi-information fusion and machine learning,the main research and achievement of this thesis are as follows:1.This thesis proposes a method based on multi-information fusion,elastic net and random forest to predict extracellular matrix proteins,which is called ECMP-RF.Firstly,the method uses encoding based on grouped weight(EBGW),pseudo amino-acid composition(Pse AAC),pseudo position-specific score matrix(Pse PSSM),composition,transformation and distribution(CTD),and autocorrelation descriptors(AD)to extract protein sequence features,then fuses five types of feature coding information to construct an initial feature space.Secondly,this method uses the synthetic minority oversampling technique algorithm(SMOTE)to balance sample data and uses elastic net to filter the optimal feature subset.Finally,ECMP-RF constructs the extracellular matrix protein prediction model through random forest.Strict leave-one-out cross-validation shows that the balanced accuracy on training dataset and independent test dataset reach 97.3% and 97.9%,respectively,which are superior to other extracellular matrix protein prediction methods.2.This thesis proposes a fertility-related proteins prediction method based on multi-information fusion and Light GBM,which is called Fertility-Light GBM.Firstly,in order to fully express protein sequence information,the method selects six feature coding methods(Pse AAC,amino acid composition(AAC),dipeptide composition(DC),CTD,AD and EBGW)to extract amino acid residues information,the encoded feature vector is fused to obtain the initial feature space.Then,for selecting effective features and improve computing efficiency,least absolute shrinkage and selection operator(LASSO)is used to select the optimal feature subset.Finally,the optimal feature subset is input into the Light GBM classifier for prediction.The 5-fold cross-validation shows that the prediction accuracy on training dataset and independent test dataset are 88.45% and 91.48%,respectively,and the performance of Fertility-Light GBM outperforms the state-of-the-art prediction methods.3.This thesis proposes a phage virion proteins prediction method based on gradient boosting decision tree-recursive feature elimination(GBDT-RFE)and Cat Boost,which is called PVP-Cat Boost.Firstly,the method fuses the evolutionary information,sequence information and physicochemical property information to construct an initial feature space.Secondly,GBDT-RFE is used to reduce the high-dimension of data.It is the first time that uses synthetic minimum oversampling technology edited nearest neighbors(SMOTE-ENN)to reduce the impact of bias caused by imbalance of data classes and to improve the learning ability of the model for minority samples.Finally,PVP-Cat Boost uses Cat Boost to classify the samples.The results show that the accuracy reach 97.3% on training dataset and 97.9% on independent test dataset through leave-one-out cross-validation. |
