Prediction Of Protein Signal Peptide Based On Domain Rules And Hybrid Deep Learning Model

Signal peptide plays an important role in guiding and transporting transmembrane proteins and secretory proteins.In essence,it is a specific amino acid sequence starting at the N-terminal of protein sequence.The study of signal peptide is helpful to explain the mechanism of disease,guide drug design and develop treatment methods.In recent years,with the explosive growth of protein sequences,prediction of protein signal peptide and its cleavage site based on protein sequences has become a research hotspot.At present,there are two challenges in the field of signal peptide prediction.One is how to effectively distinguish three different types of signal peptide,and most of the existing methods are designed for specific type of signal peptide.The functional similarity of these three different types of signal peptide makes it easy to cause misclassification during prediction.The other is how to recognize the cleavage site of signal peptide accurately.Because the length of signal peptide is scattered,it is difficult to predict the cutting site.In this work,we propose an improved prediction method,Signal-3L3.0,which uses a fusion of domain rules and mixed deep model to predict signal peptide and its cleavage site,and this method has the ability to predict three different types of signal peptide.Specifically,in the Signal-3L 3.0prediction method,we designed a self-attention layer on the signal peptide prediction model based on the Long Short-Term Memory network to extract the correlation between amino acid residues.In addition,we also applied the knowledge transfer and model integration method for the bacterial categories which have few training sample.Experiments show that the signal peptide prediction method designed in this paper has a significant effect on improving the prediction effect of signal peptide.In order to further identify the signal peptide cleavage site,we use a conditional random field model based on the mixed Long Short-Term Memory network and one-dimensional convolution network,and use a method based on {-3,-1,+ 1} domain rules to predict signal peptide cleavage site.Then,we use the weighted fusion method to fuse these two methods in the process of 5-fold nested cross validation.Experiments show that the fusion method can improve the prediction of signal peptide cleavage site.Compared with the existing methods,Signal-3L 3.0 has excellent performance.In addition,in order to transform the research results,we build an online prediction platform,Signal-3L 3.0.It's available at:http://www.csbio.sjtu.edu.cn/bioinfo/signal-3l/...