Prediction Of Gram-negative Bacteria Type Ⅲ And Type Ⅳ Secreted Effectors Based On Deep Learning

Gram-negative bacteria type III and type IV secreted effectors(T3SEs,T4SEs)play a very important role in the interaction between various bacteria and the host.Accurately predicting these two secreted effectors is very important for us to understand the pathogenic principles of pathogens.Because traditional biological experiments and mathematical methods are difficult to effectively predict these two effectors,how to develop an efficient and accurate prediction method has become a key challenge.In this paper,two methods of attention mechanism combined with convolutional neural network(CNN)and convolutional block attention module(CBAM)combined with HSRes Net are used to predict the type III and type IV secreted effectors of gram-negative bacteria.The specific research work is as follows:(1)Prediction of gram-negative bacteria type III and type IV secreted effectors based on attention mechanism and CNN.The deep learning framework of this method is mainly composed of CNN and the attention layer.First,the one-hot encoding which is an extrinsic sequence feature and the position-specific scoring matrix(PSSM)which is an intrinsic evolution information feature are extracted from the cropped secreted effectors sequence,and then merge these two features to get the input feature.Subsequently,CNN is used for feature extraction,and the attention layer is used to strengthen the network model’s attention to the connection of the information before and after the sequence and the focus of the key information.Finally,two lightweight models ACNNT3 and ACNNT4 are obtained after training,which are used to predict T3 SEs and T4 SEs,respectively.This method achieved accuracy rates of 96.7%,88.7%,and 95.0% on two independent test sets of T3 SEs and one independent test set of T4 SEs,respectively.(2)Prediction of gram-negative bacteria type III and type IV secreted effectors based on CBAM and HS-Res Net.The deep learning framework of this method is mainly composed of CBAM,the hierarchical-split block(HSB)and Res Net.First,input the tailored effector protein sequence into the HHsuite program to obtain a multiple sequence alignment(MSA),and then calculate the one-hot,PSSM,position-specific frequency matrix(PSFM)and precision matrix from MSA,after improving the PSSM and PSFM,then perform feature fusion to obtain the input feature.This method puts CBAM and the hierarchical-split block in the basic framework of Res Net,which can not only use the identity mapping and residual mapping of Res Net to effectively process the sequence features,but also give full play to the advantages of CBAM and the hierarchical-split block in extracting feature information.After training,two models with better prediction effects,CHRT3 and CHRT4,can be obtained,which are used to predict T3 SEs and T4 SEs,respectively.This method achieved accuracy rates of 96.7%,89.9%,and 96.1% on two independent test sets of T3 SEs and one independent test set of T4 SEs,respectively.The experimental results show that the two methods proposed in this paper have good experimental results for the prediction of T3 SEs and T4 SEs,and are superior to other existing methods in most indicators.Finally,this article summarizes the prediction research work of gram-negative bacteria type III and type IV secreted effectors,and looks forward to the focus of future work.