AEBoost: A Method For Anticancer Drug Sensitivity Prediction

The development of genetics has provided people with a new perspective to understand cancer and anti-cancer drugs,and it has also made it possible to adopt different cancer therapeutics for different people.Therefore,personalized medicine has also attracted wide attention from researchers.Anticancer drug sensitivity prediction is a major challenge for personalized medicine,while the development of high-throughput technology also provides a large amount of genetic information and drug sensitivity data on different cell lines.Some genomic data,such as NCI-60,CCLE and GDSC,has enabled many scholars to carry out a large number of studies on the susceptibility of anticancer drugs based on these data.In this study,we used CCLE as a dataset for anti-cancer drug sensitivity research.Then we selected gene expression data and drug sensitivity data on different cell lines.At the same time,we designed a deep learning and machine learning mixed method to train and predict anticancer drug sensitivity.Specifically,we first constructed a deep autoencoding(AE)to extract important latent variables in gene expression data on different cell lines,reducing the dimension of gene to 500.Then we established a boosting tree model XGBoost(XGB)based on the reduced dimensionality of gene expression values and drug sensitivity.We calculated the feature importance of latent variables and selected the latent variables whose feature importance value is not 0(the effective latent variables for prediction).Finally,we searched for these non-zero latent variables and establish XGBoost models respectively,and calculated the corresponding root mean square error with ten-fold cross-validation.The model established with the optimal number of latent variables is our final model.In order to verify the performance of the AEBoost,this article combined the dimensionality reduction method sure independence screening(SIS),with prediction models random forest(RF)and support vector regression(SVR)for comparison.Specific combinations include: SIS+XGB,AEBoost,AE+SVR,AE+RF.Finally,we also compared the results with the previous method(ISIRS).From the final prediction results,we can see that for the 24 drugs in CCLE,the average RMSE predicted by our method is 0.7564,and the RMSE of 6 drugs is less than 0.5(L-685458,PF2341066,etc.),and 18 drugs have RMSE less than 1.The 10-fold average RMSE of the compared prediction methods are: 0.8023(SIS+XGB),0.8284(AE+SVR),0.8757(AE+RF),ISIRS(0.9258).This shows that our method has stronger generalization ability.