| As a common pathogen,Gram-positive bacteria secrete proteins that can enter host cells from the bacterial cytoplasm,exert virulence todestroy host cells,and then cause tissue necrosis and lesions.The in-depth study of secretory proteins can help us to further understand the pathogenesis of bacteria and diseases.Among all the secreted proteins,nonclassical secreted proteins are difficult to identify as they lack discernable signal peptide sequences and can make use of diverse secretion pathways.Therefore,how to accurately identify these non-classical secreted proteins has become a challenge task in the field of microbiology.With the rapid development of artificial intelligence,several computational methods have been developed to facilitate the discovery of such proteins.However,the existing computational method have the following shortcomings:(1)the dataset is small or using the simulated data,which may lead to certain deviation and blind area in the prediction model.(2)the extracted protein features are relatively few,resulting in the inability to accurately and efficiently characterize protein information and patterns.(3)using a single machine learning model,which may lead to a bottleneck of the prediction performance.In this study,we proposed a prediction method of non-classical secreted protein of Gram-positive bacteria based on two-layer Light GBM-based ensemble model,called Pe NGa Roo.Firstly,we collected and constructed a high-quality and experimental data set by means of literature review and database retrieval.Then in order to extract effective potential information and correlation patterns from protein sequences,the sequence-derived,evolutionary information-based and physicochemical property-based feature extraction algorithm were systematically explored.From seven different feature extraction methods,1374-dimensional original feature vectors were generated.After that,the Light GBM machine learning method were utilized to train the model,which is the latest algorithm proposed by Microsoft research team.And a new multi-parameter adjustment method based on the particle swarm optimization algorithm was carried out to optimize the parameters.In order to further improve the performance of the model,we proposed a two-layer Light GBMbased ensemble model,in which each feature set was trained separately to build a single feature based model,and the aboved models were integrated into a unified ensemble framework by inter-group integration strategy.Compared to other existing predictors,Pe NGa Roo achieves better performance based on the test set,with the evalution metrics ACC of 0.779,f-value of 0.789,MCC of 0.561 and AUC of 0.962.Finally,in order to make Pe NGa Roo more convenient,efficient and widely used by researchers,we further developed a stable,publicly accessible and load-balanced web-server.It is hoped that the integrated learning model proposed in this study can accelerate the discovery and exploration nonclassical secreted protein in microorganisms,and provide new ideas for the development of next-generation prediction tools. |
