Drug Response Prediction Method Based On Collaborative Filtering

Cancer is the number one disease threatening human health.At present,both experimental and computational methods are widely used in the screening of anticancer drugs.Among them,the method based on drug sensitivity test to find suitable anticancer drugs has high accuracy,but it is limited by the experimental equipment and experimental environment,and can not meet the needs of large-scale drug screening.Computational methods can achieve large-scale drug screening at low cost,but most of the existing models only aim at the prediction design of a single drug,and lack the comprehensive consideration of the structure,target and drug reaction information between different drugs.In this paper,the drug response data of 23 drugs to 491 cell lines in CCLE database are selected to construct the drug interaction matrix of cell lines,and the drug response prediction is concretely transformed into the problem of filling the blank value in the drug interaction matrix of cell lines.Firstly,the correlation between cell lines and drugs was mined by matrix decomposition and multi-layer perceptron technology,respectively,to realize the prediction of anticancer drug sensitivity of cell lines,and the prediction correlation was0.7172 and 0.7385,respectively.In order to synthesize the linear advantage of matrix factorization and the nonlinear advantage of multi-layer perceptron,the matrix factorization and multi-layer perceptron are spliced together,and the prediction correlation of 0.7430 is obtained.Therefore,this paper proposes a deep matrix decomposition model,which further integrates the advantages of the two technologies.The prediction correlation of the model is 0.7511.Secondly,based on the drug response data in CCLE,the deep learning model of drug response prediction was established by fusing the gene expression data of cell lines and the one-dimensional and two-dimensional chemical structure data of drugs in Pub Chem database as edge information.The experimental results show that the predictive correlation of the model is 0.7526.Finally,the sensitivity of cell lines with similar gene expression to the same drug and the effect of drugs with similar chemical structure on the same cell line were studied respectively.A linear method was proposed to fuse cell line similarity network and drug similarity network,and a double-layer similarity network model was established to predict drug response.The experimental results show that the prediction correlation of the model is 0.7638.The drug response prediction model based on collaborative filtering proposed in this paper has good accuracy and generalization ability for drug sensitivity prediction,which provides a more reliable reference for drug screening,and also provides a new idea for the application of collaborative filtering in other fields.This paper contains 51 pictures,4 tables and 112 references.