SRDFM: Simaese Response Deep Factorization Machines Network For Personalized Anti-cancer Drug Recommendation

Predicting the response of a particular cancer to treatment is not only the main goal of the cross research of modern oncology and computer science,but also an important step toward personalized treatment.Although a number of computational methods have been applied to accurately predict drug responses to specific cancers and patients,accu-racy is still far from perfect.In this study,we focused on drug recommendation which is more practical and effective.Instead of predicting the exact value of drug response,we proposed a deep learning-based method for personalized drug recommendation,named Siamese Response Deep Factorization Machines(SRDFM)Network,that ranks drugs and directly provides the top effective drugs.SRDFM was compared with other ma-chine learning methods,such as the widely used elastic net,random forest,kernel ridge regression and a newly published approach called kernelized rank learning on two public benchmark datasets,the genomics of drug sensitivity in cancer(GDSC)and the cancer cell line encyclopedia(CCLE).Our approach is significantly superior to other methods for drug recommendation on both dataset.And with the response unit,our approach further explained the biological interaction between drugs and the corresponding gene expressions.We found that some drugs with the same target had more similar gene weight matrices,which agrees with the reported biological mechanism related to drug targeting.SRDFM learns cell line gene information and drug chemical information at the same time,so that the trained model can be directly used in the field of drug design,and researchers can test new drugs according to the patient’s gene expression profile,but the predictive effect of new drugs needs to be improved.We also made an in-depth analysis of the internal relationship between the drug targeting and the recommended order.We found that the drugs with the same target showed aggregation in the order.In addition,we also applied our method to drug combination recommendation,which shows the potential for expansion of our approach.