Prediction Of Antimicrobial Peptides And Their Functional Types Based On Multi-Label Transductive Learning

Antimicrobial peptides,a type of polypeptides with broad-spectrum antimicrobial activity,are widely found in organisms and are important components of their innate immune systems.Because of a slim chance of bacterial resistance,antimicrobial peptides have become a preferred option for the pharmaceutical industry to develop new antibacterial preparations.In this sense,it is of great significance to identify more antimicrobial peptides and make clear their antibacterial functional types.Among all ways of predicting antimicrobial peptides and their functional types,the calculation method which is based on machine learning is increasingly popular because of its high-accuracy,low-cost,high-feasibility and high-reliability.However,the existing methods cannot identify the antimicrobial peptides and predict their antimicrobial functional types at the same time,and there is still room for improvement in terms of the calculation accuracy.In view of this,this paper proposes a new method based on machine learning to predict antimicrobial peptides and their functional types.The main contents are as follows:1.Proposes a multi-label classification method based on a single optimization problem,which can not only predict whether a polypeptide is an antibacterial peptide,but also predict what antimicrobial functional type(s)would be,single type or multiple types including antimicrobial function,antifungal function,anti-tumor function,anti-virus function,anti-HIV function.2.In view of the limited amount of labeled samples and abundant information contained in the massive unlabeled samples,proposes a graph-based transductive prediction model,aiming at improving the prediction performance through the study of labeled training data and unlabeled data to be tested.In addition,different weighting factors are given to the categories of antimicrobial functions when constructing the nearest neighbor graph using transductive learning,differentiating the contributions to the prediction method of different categories.3.In order to further testify the generalization of the proposed method,including taking advantage of the existing datasets,this paper proposes a new test set that has lower homology with the training set by using newly-published antimicrobial peptides of APD3(Antimicrobial Peptide Database 3)and non-antimicrobial peptides of UniProt(Universal Protein).This method extracts the polypeptide characteristics by Composition of K-Spaced Amino Acid Pairs(CKSAAP).Experiments have shown that the proposed method is more accurate,compared with iAMP-2L method,in terms of overall prediction and multi-label prediction.In order to better carry out the communication,the online prediction platform based on the proposed method has also been available,able to provide download and prediction services for researchers who take interests in it.