Construction And Application Of Machine Learning Prediction Method For Combinatorial Drugs

Drug combination is a promising and important research field in the current pharmaceutical industry.Traditional pharmacology follows the concept of "one drug and one target",but drug resistance and side effects often appear when using a single drug treatment,and the discovery of new targets is relatively slow,making new drug development lags behind market demand.Due to the complexity of the organism itself,complex diseases have little effect under the action of a single drug.The combined use of drugs is a common therapy approved by the FDA.There have been cases that the combined drugs have good curative effect on complex diseases and can play the role of "increasing efficacy and reducing toxicity".The current combination drug verification is based on the verification of biological tests,which is not only time-consuming and labor-intensive,but also a great waste of funds.In recent years,the accumulation of high-throughput sequencing data and the advancement of machine learning algorithms have also laid the data and theoretical foundation for the application of computational drug models in drug research and development.However,the existing calculation methods at this stage only use part of the data and do not make full use of the accumulated multi-omics data,which also makes many calculations results less reliable.This study collected drug and target and indication data in Drug Bank,Therapeutic Target Database(TTD)and The Drug Gene Interaction Database(DGIdb),and also collected gene expression profile data that are influenced by 1309 small drug molecules in Connectivity Map(CMap).The combination drug data comes from the Drug Combination Database(DCDB)and Pre DC database.We first used a dual clustering algorithm to cluster 1309 small drug molecules in CMap,and then intersected CMap drugs with drugs in the three databases of Drug Bank,TTD,and DGIdb;afterwards,the chemical structure similarity of drugs was calculated by Pub Chem and calculated the Tanimoto coefficient between the two drugs through the collected drug data;then used the Tanimoto coefficient as a parameter to use support vector machine,naive Bayes and logistic regression to construct a combined drug prediction model with different positive and negative sample ratios and filter out support vector machine is used as the best model.Then,using the neighbor recommender method to construct five similarity models and using logistic regression as the ensemble learning algorithm to construct the ensemble model and perform feature filtering.This study finally found that the ensemble model constructed using drug target similarity,drug indication similarity,drug chemical structure similarity,and drug expression profile similarity AUROC value was 0.89 and AURP value was 0.383;after the screening of the ensemble model and support vector machine models,the former had better prediction effect.In addition,this study applied the ensemble method to the published data sets and achieved better results than the original articles.Finally,the optimal model is applied to the combination drug prediction of paclitaxel.The predicted drug combination has been proved to have a better combination synergistic effect by preliminary experiments.In summary,our ensemble model method can not only get robust prediction model from our own data,but also predict the combination of drugs with biological activity,which has potential application value for the research and development of combination drugs in the pharmaceutical industry.