| In drug design,many existing target spaces are close to saturation,and the drug design of emerging or potential targets has become a current research central issue.However,the information of known active ligand samples possessed by these emerging or potential drug targets is often insufficient.With the rapid development of deep transfer learning,it provides us with good solutions to these challenges in the ligand-based virtual screening with insufficient samples.In order to solve the problem of insufficient samples in virtual screening,in this paper we proposed a deep transfer learning algorithm for virtual screening of ligands,which can be used to predict the biological activity values of drug targets and ligands.The steps of the method are as follows.First,a target dataset with insufficient sample size was obtained,and similar datasets with sufficient sample size as source domain datasets was found;Then,a learning model was built for the WDL2 algorithm by the usage of the source domain datasets,and the parameter model could be obtained after training.Next,based on the parameter transfer,the parameter model obtained from the source domain was migrated to the target domain,which helped the target domain to similarly obtain the learning model.Finally,the predicted bioactivity values of the target domain were yielded through the random forest algorithm.We tested our algorithm on 54 data sets.The regression prediction results are measured by two commonly used evaluation indicators(r~2 and RMSE).The comparison algorithm selected in the experiment is the weighted deep learning algorithm and the weighted deep transfer learning algorithm.The final experiment shows that the r~2 of the regression prediction model is increased by an average of 45%over the weighted deep learning algorithm and 24%over the weighted deep transfer learning algorithm.The results shows that our algorithm is effectiveness in the modelling of bioactivities of ligands with insufficient samples. |
PDF Full Text Request