Research On Cancer Classification Methoes Based On Multi-Omics Data

Cancer is currently the second leading cause of death worldwide,with an average of one in six people dying from cancer in the world.In order to reduce its impact on human health,a lot of research work has been devoted to cancer diagnosis and treatment technology.The use of human omics data to classify and predict cancer types and clinical events is an important topic,and it is of great research significance and practical value for the early accurate diagnosis of cancer patients.The comprehensive analysis of multi-omics data provides a comprehensive perspective of the patient,which may make clinical decisions more accurate.However,the effective use of multi-omics data for the classification and analysis of cancer is still a huge challenge.The expression profile data set has high feature dimensions and few samples.Due to the high cost of obtaining omics data,even unlabeled data samples are rare.If you do not consider the association between different groups,using multidata as the input set will make the feature dimensions higher and more likely to cause overfitting.In view of the above problems,this paper proposes a method for classifying cancers using various types of omics data of cancer patients.First,in order to tap the potential connections between molecules and between molecules and patients,this paper proposes and builds a heterogeneous network that integrates multiple biomolecular interaction networks,which can embed different types of biomolecules and patients on the interaction network into the same space.Because the network is weighted,this paper proposes a weighted heterogeneous representation learning algorithm Attr Metagraph2 path,which is used to learn representations of the network to obtain node representations with associated information.Next,this paper proposes a multiview autoencoder with interactive knowledge constraints,and uses the previously learned representations of molecular interaction networks for downstream classification tasks.This method integrates the knowledge of molecular interaction network into the learning of multi-view factorization and self-encoding,and is used to guide the patient’s representation of a variety of cancer expression profile data,thereby using biological knowledge to obtain the regularization effect.In the task of PFI classification of cancers using four kinds of expression profile data,it is shown that the method proposed in this paper improves the AUC.The experimental results on the TCGA dataset using the method of multi-omics data classification proposed in this paper show that this method will integrate the knowledge of molecular interaction network into the classifier,which significantly improves the level of classification.This paper also verifies that the potential relationship of network representation learning capture is helpful for canonical representation of cancer expression profile features in the classifier.This also makes the representation learning of features from different perspectives more relevant.In order to further explore multiple omics data The links provide help and guidance.In addition,the Attr Metagraph2 vec algorithm proposed in this paper can use other weighted heterogeneous representation network learning areas.