Mutation-dependent Drug Repositioning For Personalized Cancer Treatment

Objectives:Cancer is a major complex disease in humans with high mortality,and the incidence of cancer is increasing in recent years.Most of the drugs that currently used to treat cancer are developed for specific patient groups,but individuals within the group may have different sensitivities to the same drug.Following the idea of precision medicine,current cancer therapy is suggested to treat the same disease with different drugs and homotherapy for heteropathy,such as drug repurposing.Compared with traditional drug development process,drug repurposing has the advantages of shorter time,lower cost and higher success rate.There are many successful cases to apply drugs designed for other diseases to the treatment of cancer.Therefore,cancer drug repurposing of non-oncological drugs has important scientific and clinical significance.It has been reported that the integration of genetic evidence during drug target selection can significantly increase the success rate of drug development.Currently,most existing genetics-based drug repurposing methods are designed to make drug repurposing for population based on single gene,and the drug's mechanism of action is not considered.This study aims to develop a method for predicting drug repurposing at individual level based on the target information of the drug,the mutation information of the cancer as well as the gene and the molecular network.This work can calculate probability of potential repurposing of drugs based on patient mutation information,provide evidence for the selection of population for clinical trials of drug repurposing and prioritize drug candidates for individualized cancer treatment,which has promising prospects in clinical application.Methods and materials:We collected drug target information from The Drug–gene Interaction Database(DGIdb)and Drug Repurposing Hub,and collected drug indication information from DrugBank and other drug-related databases.The cancer mutations information including single-nucleotide variants(SNVs),short insertions and deletions(Indels),copy-number variation(CNVs),chromosomal translocations,deletions,duplications and inversions were collected from International Cancer Genome Consortium(ICGC)covering 51 types of cancers.Then,we used the Combined Annotation Dependent Depletion(CADD)score to predict and screen for pathogenic mutations and mapped the pathogenic mutations to target genes at different levels.We integrated reported cancer driver genes from Catalogue Of Somatic Mutation In Cancer(COSMIC)and OncoKB,and made predictions by OncodriveROLE to generate a more comprehensive cancer driver gene set.We linked drugs with cancers by two strategies.If the drug targets were the same as the cancer genes,the relationship between the drug targets and the cancer genes was defined as a direct link.If the drug targets were not the same as the cancer driver genes,the drug targets and cancers driver genes would be associated by random walk with restart algorithm based on the Reactome Pathway network.Based on the drug-cancer matching information obtained in the above two ways,we extracted and compiled the input features at individual patient level,and trained the logistic regression model on the known drug-cancer pair set to predict the probability of drug repurposing of all the drug-cancer pairs.Using cross-validation,patients drug sensitivity data and gene expression-based drug sensitivity prediction scores,we comprehensively evaluated the predictive model and applied the model to validate clinical actionable mutations.Results:In this work,we got 5,948 drugs with known indications.The cancer samples in ICGC were mapped to 51 types of cancers covering 22 tissues.The pathogenicity of mutations including SNVs/Indels,CNVs,deletions,duplications,inversions and translocations were predicted.We also integrated 2,270 cancer genes with known functions.256,889 drug-cancer pairs were identified by gene-based and network-based approaches.We have built a logistic regression model that can predict the probability of drug repurposing for a given drug-cancer pair.The training dataset of the model includes positive samples which were drug-cancer pairs labeled with sensitivity or response in CIViC and negative samples which were drug-cancer pairs labeled with terminated or withdrawn clinical trials in repoDB.And the training set of the model was used to estimate the model's performance through cross-validation and The Area Under Curve(AUC)was 0.794.We found that the most predicted values of drug repurposing probability distributed between 0-0.75.There were 416 drug-cancer pairs with predicted probabilities are close to 1.They were considered to be potential drug repurposing opportunities with high success rates.To construct an independent test dataset,we collected drug-cancer pairs which had been proved in clinical trials from other resources as positive test samples.According to the number of the positive test samples,nine times the number of negative test samples were randomly drawn from the all drug-cancer pairs,excluding the drug-cancer pairs in the positive samples or in the training dataset of the model,obtained by our methods.We used these samples to assess our model,and obtained an AUC of 0.838.We also used patient drug sensitivity data from The Cancer Genome Atlas(TCGA)to evaluate the prediction performance of our method.The drug-patient pairs of which drug sensitivity was complete response were defined as positive samples while pairs of which drug sensitivity was stable disease and clinical progressive disease were defined as negative samples.When testing our model with TCGA data,we got an AUC of 0.614.In addition,the prediction of our method was consistent with that of another transcriptome-based method.The above evaluations indicate that the study has good prediction accuracy and clinical application value.