Multilayer Network-based Method For Drug Response Prediction

Cancer is complex heterogeneous disease.Traditional treatment,which is based on the clinical symptoms and ignores the biomolecular characteristics of cancer patients,cannot meet the requirements of modern medical treatment of cancer.Precision medicine can guide the process of preventing,treating,and alleviating disease by analyzing the biomolecular and pathological information of the patient's proteome,genome,transcriptome,epigenetic group,and metabolome.Data of genomics,transcriptomics,and epigenetics for cancer cell lines provide an opportunity to establish computational models to predict drug response.We proposed a drug response prediction method named MNDRP(Multilayer Network-based Method for Drug Response Prediction),which applied expression,gene mutation and copy number variation for cancer cell line and chemical structure and target for drug.First,a threemultilayer cancer cell line similarity network was constructed based on gene expression,gene mutation and copy number variation of cancer cell lines,and a two-multilayer drug similarity network was constructed based on chemical structural features and target.Secondly,the low-dimensional feature vector of nodes in the three-multilayer similarity network of cancer cell lines and the low-dimensional feature vectors of nodes in the drug two-multilayer similarity network were extracted by diffusion component analysis.Thirdly,the low-dimensional eigenvectors of the cancer cell line and the low-dimensional eigenvectors of the drug were combined as the characteristics for the training data set.The sensitivity relationship of the known drug-cancer cell line was used as the label of the training set.A binomial logistic regression model was obtained based on the training set.Then,we validated performance of MNDRP through cross-validation,comparison with other methods,prediction of cancer cell lines in different tissues,and so on.Finally,the new sensitive associations between cancer cell line and drug were meaningful of literature and enrichment analysis which were predicted by MNDRP.MNDRP obtained the low-dimensional eigenvectors of cancer nodes and drug nodes by using the diffusion component analysis algorithm,and extracted the common parts of the cancer cell line in the three-layer similarity network of cancer cell lines and the common part of the drug in the drug two-layer similarity network.Moreover,the influence of redundant data and noise data on drug reaction prediction was reduced,the accuracy of drug reaction prediction is improved,and drug candidates for biological experiments of anticancer drugs were provided in the future,which reduces the cost of the test.