Synergy Prediction Of Anticancer Drug Combinations Based On Kernel Norm And Ridge Regression

Combination therapy with anticancer drugs is a promising therapeutic strategy.It is important to select highly synergistic drug combinations for specific cancer types to improve cancer efficacy.For specific patients,which drugs should be used in combination therapy and which combination of drugs may be effective are the problems that must be faced and solved in clinical treatment.Identifying synergistic drug combinations is a complex and difficult task.It has become absolutely necessary to use computational tools that can complete high-throughput virtual screening of individual tumor oriented anticancer drug combinations in a reasoning mode,providing the possibility to explore the space of large drug combinations.In this paper,based on the high-throughput screening data set of anticancer drug combinations published by O’Neil’s team,a computational model is constructed to predict the synergistic effect of anticancer drug combinations and promote the application of innovative computational methods in anticancer precision therapy.Based on the idea of matrix filling,a computing model NNRM based on nuclear norm regularization is constructed.Firstly,a symmetric cooperative score observation matrix is constructed for fixed cell lines,which is thinned by folding technique.Then,NNRM is solved by alternating direction multiplier method and soft threshold estimation.Finally,NNRM is applied to the data set published by O’Neil’s team to predict the synergistic score and status of anticancer drug combinations.The prediction performance is better than that of random forest and support vector machine,and some of the prediction indexes are better than that of Deep Synergy.Without the need for any ancillary data,NNRM achieves largescale virtual screening of new drug combinations at low computational cost: for each cell line,simply testing its predicted top 90 drug combinations,the true best drug combination can be selected with 90% probability;the best drug combination predicted by 92% cell lines appears in the top 40 of their observed drug combinations.In order to improve the prediction performance of the model,the integrated prediction of NNRM and RR is realized by means of drug molecular fingerprint and ridge regression(RR).Firstly,the drug molecular fingerprint matrix is processed by variance screening and Z-Score standardization,and then the molecular fingerprints of the two drugs are spliced horizontally to obtain the molecular fingerprint matrix of the drug combinations.Then,the observation matrix is thinned by using the same folding technique as NNRM.In order to make full use of the association information contained in the sparse observation data,the upper triangle of the observation matrix is further transformed into a column vector.The column vector and the drug combination molecular fingerprint matrix are input into ridge function,and the parameters are adjusted by using cross-validation and grid search methods.Finally,the prediction results of NNRM and RR are weighted together to obtain the prediction results of the integrated model,whose overall prediction performance is fully comparable to that of Deep Synergy.In addition,the prediction results of the integrated model for partial synergistic score observations and missing data have been confirmed by the existing literature and clinic,and some highly synergistic drug combinations have been found that have not been studied.These results indicate that the NNRM-RR integrated model can be used as a high-throughput virtual screening tool for anticancer drug combinations,and can provide a theoretical reference for “relocating old drugs to new combinations”.