Research Of Cancer Subtyping Based On Tensor Factorization

Classification of cancer through phenotype or tissue organ where cancer sample from is not exactly correct,clinical treatment of cancer will need much more accurate subtypes so that proper medicine could be given.Analysis of microarray data such as m RNA,mi RNA,DNA,protein and other mutation kind of genes could help to reveal accurate cancer subtypes.The integration of multi-source genomic data can not only help to discovery the relationship between tumor and genomic data,but also can help to find the synergy between themselves.How to consider different genetic data in the premise of not losing the information can also analyze the sharing structure is the difficulty of the cancer subtyping.This paper propose a multi-dimensional array tensor to integrate multi-source data,without losing the original expression information but also the premise of data intermediate conversion under the condition of preserving the mining cooperative mode between different pathogenic gene expression data,and also introduces the principle and framework of tensor model.We construct tensor model based on the gene expression profile data and DNA methylation data of breast tumor from TCGA,the method of the construction is to do a differential expression analysis of the pre-processed microarray data,and retain those values whose gene is obvious different with others in the control sample as 1,and make those who are not as 0,so that the gene expression matrix and DNA methylation matrix can be transformed into a three-way tensor.On the basis of the existing CPARP decomposition algorithm,we introduce non-negativity and sparse constraints to optimize the CP decomposition model,according to the difference of gene chip data whose dimension is high but samples are far few.Improved mode uses ALS optimization method based on stochastic gradient descent,and behaves better in calculation when comparing with other method.The effectiveness of the tensor decomposition model in the classification of cancer was demonstrated b comparing the results with the five subtypes of breast cancer which have been verified.Through the analysis of the results of the classification of cancer,it is verified that Her2 has been proved to exist in the clinical subtypes.Also proved the method proposed in this paper can provide some reference for diagnosis and treatment of cancer.