Research On Cancer Subtypes Prediction Based On Subspace Clustering

Cancer is characterized by strong aggressive and high mortality rate.In recent years,cancer prediction research has become a key research direction in bioinformatics.Generally,a type of cancer is composed of multiple subtypes.Effective prediction of cancer subtypes is of great significance to improve the overall diagnosis and cure effect of cancer and provide targeted precision medicine.High-throughput technologies have generated unprecedented amount and types of omics data,which is conducive to investigate the pathogenesis and evolution of cancer from the biological molecular level,and discover its crucial disease-causing genes,and provides the possibility of bridging the distance between cancer genome to cancer phenotypes.The traditional researches on cancer subtypes only utilize single-level omics data for prediction,which has some limitations.The prediction effect can be effectively improved if an overall analysis of the correlation between multiple omics data can be carried out.The prerequisite for the integrated analysis of multi-level omics data is that each omics data is complete and available,which limits its applicability in clinical practice.In this thesis,the research of cancer subtypes is based on subspace learning.In this thesis,a novel computational method is proposed,i.e.Deep Subspace Mutual Learning for cancer subtypes prediction(DSML).The algorithm combines sparse representation with mutual learning theory,and adds a network self-expression layer to the auto-encoder structure to obtain the discriminative feature representation of the data.The network is composed of several branches and a concentrated main-stem.The branch part learns the feature representation and self-expression relationship of single-level omics data.The concentrated main-stem part integrates multiple omics features and learns the similarity relationship of the overall data.Finally,the similarity matrix and spectral clustering algorithm are used to predict cancer subtypes.In this thesis,we proposed a low-rank representation algorithm for cancer subtypes prediction.The algorithm adds non-negative constraints and rank constraints to the low-rank representation model and iteratively solves it through the alternating direction method,so as to obtain the lowest-rank feature representation of the sample.The similarity matrix can be constructed by the lowest rank feature representation of sample data.Firstly,the corresponding low-rank model is established for each single-level omics data and optimized iteratively.Then,the obtained low-rank representation matrices are fused and the similarity matrix is constructed.Finally,the result of predicting cancer subtypes can be obtained by normalized cut on the similarity matrix.The performance of the algorithms is verified by using three-level omics data from five publicly available cancer datasets.Clustering evaluations and survival analysis both demonstrate that the proposed algorithms deliver comparable or even better results than many other similar methods,which provides reference value for designing effective cancer treatment.