Prediction Of Protein Structural Classes Based On Feature Fusion And 2D-wavelet Denoising

The spatial structure of proteins determine its function,and the determination of its spatial structure is very complex.Previous studies have shown that obtaining the structural classes of proteins can determine the folded form of the peptide chain,thus narrowing the natural search scope of spatial structure,providing direction for its structure determination,saving a lot of time and energy.However,the prediction rate of structural classes has been very low,especially in low similarity sequences.The reason for the low prediction rate is that the feature information extracted by traditional methods contain a lot of redundant information.In this regard,the following work has been done in this thesis:(1)A prediction model based on reduced-dimension of feature fusion and iterative two-dimensional wavelet denoising is proposed,which is called FU-ERD-IWD(Model 1).It is used to solve the problem of increasing computational cost and redundant information after fusion of high-dimensional feature vectors in protein structure class prediction.First,dipeptide and tripeptide composition are used to extract the feature information of protein sequences,and KPCA is used to reduce the dimensions of extracted feature vectors.Secondly,feature vectors after reducing the dimensions are fused.After that,the feature vectors are denoised by two-dimensional wavelet denoising,and the redundant information is removed by iterating the denoising process several times.Although the prediction effect of this model is different from that of other literatures,the strategies adopted are still effective and feasible.(2)A prediction model based on parallel two-dimensional wavelet denoising and feature fusion is proposed,which is called PWD-FU-PseAAC(Model 2).A new fusion strategy is used to enhance the validity of low-dimensional feature vectors and make them easier to predict and recognize.Firstly,the feature vectors of protein sequences are extracted by two types of pseudo-amino acid composition,and then the two feature vectors are denoised by two-dimensional wavelet to remove the redundant information.Secondly,the two feature vectors after denoising are fused and the optimal feature vectors are obtained.Compared with another fusion strategy:fusion before denoising,the prediction effect of the proposed strategy is better.Compared with other prediction methods in literature,the model has higher prediction accuracy.Moreover,this model is expected to be applied in other fields of bioinformatics.