Research On Matrix Completion With Side Information For Better Modeling Bioactivates Of Drug Leads

For drug discovery,the implementation of virtual screening is to predict bioactivities of lead compounds and accurate modeling of their bioactivities is the most critical.Clinical trials for a drug candidate usually are required for a long R&D cycle,a large number of test compounds,huge demand for equipments and massive financial support.For each drug-target task,the number of its associated lead compounds entries via biological assays usually are insufficient.The inclusion of multiple drug target tasks in learning bioactivities of drug leads through matrix completion potentially enhances the model performance due to the utilization of correlation information among drug target tasks.Therefore,in this paper we propose a noval method for predicting bioactivities of drug leads based on matrix completion,where our method overcomes the problem of existing a large number of missing bioactivity values in the association target-lead matrices.To improve the model performance,our method also couples some side information about the extended-connectivity fingerprints(ECFPs)of drug leads.The pipeline of our method is as follows.First,it is to obtain the target-lead association matrices.Then,it is to extract the ECFPs information of leads with 512 dimensionals.Next,it is to train the predictors using three-fold cross validation after obtaining the optimal parameters.Finally,it is to achieve the predicted target-lead association matrices and to return the model performance.We tested our method on a representative 72 G protein-coupled receptors(GPCRs)drug targets,which cover 24 subfamilies of GPCRs.The results show that our method is overall superior to classic single-task learning methods and matrix completion methods.In addition,our method achieves better performance than state-of-the-art methods based on deep multi-task learning on most datasets,where it obtained an average improvement of 18%on the correlation coefficient(r~2)and 12%on the root mean square error(RMSE)over the Deep Neural Net-QSAR predictors.