Survival-Related Clustering of Cancer Patients by Integrating Clinical and Biological Datasets
Date of Award
Master of Science (MS)
Mathematical and Statistical Sciences
Subtype-based treatments and drug therapies are essential aspects to be considered in cancer patients' clinical trials to provide appropriate personalized therapies. With the advancement of the next-generation sequencing technology, several computational models, integrating genomic and transcriptomic datasets (i.e., multi-omics) in the prediction of subtype-based classification in cancer patients, were emerged. However, integration of the prognostic features from the clinical data, related to survival risks with the multi-omics datasets in the prediction of different subtypes, is limited and an important research area to be explored. In this study, we proposed a data integration pipeline with the prognostic features from the clinical data and multi-omics datasets to predict the survival-risk-based subtypes in Kidney Renal Clear Cell Carcinoma (KIRC) patients from The Cancer Genome Atlas (TCGA) database. Firstly, we applied an unsupervised clustering algorithm on KIRC patients and clustered them into two survival-risk-based subgroups, i.e., subtypes. Then, using the clustering-based subtype labels as class labels for cancer patients, we trained a supervised classification model to determine the class label of un-labeled patients.In our clustering step, we applied multivariate Cox Proportional Hazard (Cox-PH) model to select the survival-related prognostically significant features (p-value < 0.05) from the patients’ multivariate clinical data. Then, we used the Silhouette Coefficient to determine the optimal number (k) of the clusters. In our classification step, we integrated high dimensional multi-omics datasets with three different data modalities (such as gene expression, microRNA expression, and DNA methylation). We utilized a dimension-reduction approach, followed by a univariate Cox-PH for each reduced data modality with patients’ survival status. Then, we selected the survival-related reduced-omics-features in our classification model. In this step, we applied a supervised classification method with 10-fold cross-validation to check our survival-based subtype prediction accuracy. We tested multiple machine learning and deep learning algorithms in different steps of the pipeline for clustering (K-means, K-modes and, Gaussian mixture model), dimension-reduction (Denoising Autoencoder and Principal Component Analysis) and classification (Support Vector Machine and Random Forest) purposes. We proposed an optimized model with the highest survival-specific-subtype classification accuracy as the final model.