| Protein subcellular localization prediction can help researchers understand the properties and functions of crop proteins,comprehend the reasons for the emergence of crop phenotypic traits,and contribute to the development of new drugs.To solve the subcellular localization problem of crop proteins more accurately and quickly,researchers have gradually expanded their research methods from traditional experimental methods to the field of machine learning.Currently,most research is limited to fixed datasets consisting of bacteria,viruses,apoptotic proteins,and other components,typically consisting of a few hundred data points.There is relatively little research on the subcellular localization prediction of crop proteins,which have large datasets and multiple predicted locations,and there is no suitable method for predicting the subcellular localization of crop proteins at present.In this article,we explore effective methods for predicting the subcellular localization of crop proteins,and improve the prediction accuracy of subcellular localization in soybeans and tomatoes using machine learning algorithms.The following research achievements were obtained:1.We propose a crop protein subcellular localization method based on a Pse PSSM+Top-n-gram multi-feature fusion feature extraction algorithm.By integrating crop protein evolutionary information with amino acid physicochemical properties and residue information,we obtain a multi-feature fusion feature extraction algorithm that can more fully extract sequence information from crop proteins.We then use the SMOTE algorithm to balance the dataset,perform dimensionality reduction using PCA,and finally use the SVM algorithm to predict protein subcellular locations.Using the jackknife method for cross-validation and comparing the prediction results with those of traditional feature extraction methods on two crop datasets,the effectiveness of the new method is demonstrated.2.To explore effective methods for predicting crop protein subcellular localization at different levels of data loss and the performance of different algorithms on different crop datasets,we randomly remove 5%,10%,15%,20%,and 30% of the original data,and predict the subcellular location of the missing sequences using four machine learning algorithms: SVM,Naive Bayes,Random Forest,and Decision Tree.We perform correlation analysis on the predicted location and the original location,and compare the accuracy and performance of different algorithms.The results provide reference for researchers in predicting the subcellular localization of other crop proteins.3.We propose a crop protein subcellular localization prediction method based on multi-feature fusion and ensemble classifiers.On the basis of the Pse PSSM+Top-n-gram feature extraction algorithm,we fuse the pseudo amino acid composition information Pse AAC for feature extraction.The new method contains more abundant amino acid composition information,and the vectors of the two crop proteins are dimensionally reduced to reduce the problem of redundancy in feature vectors when predicting separately.On the basis of using the new method to extract feature information,we select the SVM algorithm with better prediction performance for ensemble prediction.By constructing an ensemble classifier,we achieve ideal prediction results on two crop datasets,demonstrating the effectiveness of ensemble machine learning algorithms,It can provide a new method for the prediction of subcellular localization of other crop proteins. |
