Research On Cancer Subtypeing And Drug Response Prediction Based On Multi Omics Data

The continuous progress of high throughput screening technology and single-cell multi-omics technology has brought various omics data related to cancer.The advancement of machine learning and deep learning technology provides us with tools for processing multi omics data.This dissertation studies two key issues in the current cancer research,which are identify the molecular subtype of tumor according to the multi omics data and use multi omics data to predict drug response for cancer patients respectively and proposed deep learning models for these issues.This dissertation first studies the biological characteristics,integration methods,and pre-processing methods of multi omics data,and proposed evaluation indicators for cancer subtypeing and drug response prediction,which pave the way for the next study.The main research work in this dissertation includes:(1)This dissertation proposes a autoencoder model based on multi omics data for cancer subtyping.This model gets the hidden variable representation of multi omics data through encoding and decoding process and multi-scale featrue fusion method,and enter them into the K-means algorithm for clustering.The experimental results show that the model can better handle high dimensional but few samples problems of multi omics data,which is better than the existing algorithm in important indicators.(2)This dissertation proposes a multi-modal neural network model based on multi-omics data to predict the cancer drug response,which uses a multi-modal structure to establish a sub-network for each omics data to extract the features of them,during the data pre-processing stage the model uses a data set selection rule to maximize the use of data and use Morgan fingerprint to represent the drug to capture the three-dimensional structure information of the drug.The experiment on test set and CCLE dataset proves the effectiveness of the model proposed in this dissertation.(3)This dissertation analyzes the impact of omics data selection method on data integration.The preliminary experimental results show that more types of omics data not always help produce better result,it depends on the specific problem,but in this case,the indicators of multi omics data are closer to the best value,and the results obtained are more stable,considering the time factor,we should choose to use multi omics data under normal circumstances.