Research On Drug-target Interaction Prediction Based On Feature Fusion And Ensemble Learning Method

Drugs prevent and treat various diseases by regulating the human physiological processes.Researches have indicated that drug-target interactions play a crucial role in drug development and reposition.Detecting the potential interacting drug-target pairs helps to find the genes or proteins which cause the diseases.Although biochemistry technology has rapid growth,traditional clinical trials are still time-consuming,laborious,and complex for target specificity and biological network robustness.In the later stages of the clinical trials,researchers have to deal with side effects and the rates of false-positive and false-negative.Hence,the variety of medicines approved for the market does not increase rapidly with the development of proteomics and chemo-genomics.With the continuous research of machine learning,the prediction models provide new solutions for screening reliable drugs.In this paper,two drug-target interaction prediction methods based on feature fusion and ensemble learning methods are constructed.The main contents of the studies are as follows.For the difficulty of obtaining the three-dimensional structure of drug molecules and proteins,studies have proved that molecular fingerprints can effectively characterize the molecular substructure.Meanwhile,many numerical conversion methods can describe the protein sequences.In this paper,Pub Chem molecular fingerprints are utilized to extract the drug structure information.These fingerprints reduce information loss and error accumulation during the description.Meanwhile,the position-specific score matrix(PSSM)was utilized to describe the protein sequence to obtain more evolutionary information.For the PSSMs,the Pyramid Histogram of Oriented Gradients(PHOG)and Fuzzy Local Ternary Pattern(FLTP)were adopted to extract the local features,Zernike Moments(ZMs)was employed to describe the global features.The local and global features were integrated to improve the sample differences.The experiments also show that the fusion features model can extract more hidden information of PSSMs,compared with the single feature.The Random Forest(RF)and Rotation Forest(Ro F)are utilized to identify the interacting drug-target pairs.In four datasets,the accuracies of PHOG-ZMs model were84.43%,81.31%,79.30% and 74.13% respectively;The accuracies of FLTP-ZMs model were 91.05%,88.49%,81.85% and 74.89% respectively.The comparisons of support vector machine and light gradient boosting machine are also built.Meanwhile,the proposed models are compared with other state-of-art models.The results illustrate that the proposed model can effectively predict drug-target interactions.