Study On Prediction For Anticancer Drug Sensitivity Based On Matrix Completion

Cancer is the leading cause of death worldwide.One of the key goals of precision medicine is to formulate personalized treatment plans based on individual differences while minimizing adverse side effects.Anti-cancer drug sensitivity prediction is an important research direction in precision medicine.Using computational models to predict the unknown response of different drugs when acting on cancer cell lines is very valuable for the development of personalized therapies for patients.Anti-cancer drug sensitivity prediction is not only computationally challenging,but also requires further exploration and research in the use of side information.The data used in this article includes two cell line-drug response datasets GDSC(Genomics of Drug Sensitivity in Cancer)and CCLE(Cancer Cell Line Encyclopedia)and the cell line-cell line and drug-drug similarity of the two datasets.In this thesis,three matrix completion algorithms are used to predict the susceptibility of anticancer drugs.They are matrix completion algorithm based on nuclear norm regularization,matrix completion algorithm based on self-expression,and matrix completion algorithm based on similarity heterogeneous network.In the experimental analysis part,10-fold cross-validation was used to evaluate the performance of the model.Parameter optimization experiments were performed on several models to find the optimal parameter combination.Prediction experiments with different degrees of missing were set up to test the robustness of the model.Case studies of typical drugs,case studies of targeted gene drugs in the PI3 K pathway,and case studies of the consistency of gene mutation drug sensitivity were carried out.In the matrix completion algorithm model based on nuclear norm regularization(Nuclear Norm Regularization Matrix Completion Model,NNRMCM),the objective function is to minimize the nuclear norm of the matrix to be filled and the relaxed form of the constraints that can tolerate noise.The NNRMCM model needs to set the rank of the input matrix in advance,and solve it through the alternating direction multiplier method.The experimental results show that the prediction correlation of NNRMCM model can reach 0.73 on the GDSC dataset and 0.77 on the CCLE dataset.In the self-expressive matrix completion algorithm model(Self-Expressive Matrix Completion Model,SEMCM),the objective function is to minimize the nuclear norm of the matrix to be filled and the respective-norms of the coefficient matrix and the error matrix in the self-expressive formula.The SEMCM model does not need to pre-set the rank of the input matrix,and is also solved by the alternating direction multiplier method.Experimental results show that the prediction correlation of the SEMCM model can reach 0.84 on the GDSC dataset and 0.80 on the CCLE dataset.In the matrix completion algorithm based on the Similarity Heterogeneous Network(SHN),combined with the similarity heterogeneous network,the NNRMCM and SEMCM models are improved and the SHN-NNRMCM and SHN-SEMCM models are proposed.The experimental results show that the SHN-SEMCM model has a better prediction effect than other models.The prediction correlation can reach0.85 on the GDSC dataset and 0.81 on the CCLE dataset.It can be proved that the similarity as auxiliary information helps to improve the prediction performance of the model.In summary,we can see that the three algorithms proposed in this article are effective and reliable.There are 26 figures,14 tables and 92 references in this thesis.