Cong Liu1, Xujun Wang2, Georgi Z Genchev3, Hui Lu4. 1. Department of Bioengineering, University of Illinois at Chicago, Chicago, USA; Center for Biomedical Informatics, Shanghai Children's Hospital, Shanghai, China. Electronic address: cliu55@uic.edu. 2. SJTU-Yale Joint Center for Biostatistics, Shanghai Jiaotong University, Shanghai, China; Department of Bioinformatics and Biostatistics, Shanghai Jiaotong University, Shanghai, China. Electronic address: jameswong.tju@gmail.com. 3. SJTU-Yale Joint Center for Biostatistics, Shanghai Jiaotong University, Shanghai, China; Center for Biomedical Informatics, Shanghai Children's Hospital, Shanghai, China. Electronic address: georgi.z.genchev@gmail.com. 4. Department of Bioengineering, University of Illinois at Chicago, Chicago, USA; SJTU-Yale Joint Center for Biostatistics, Shanghai Jiaotong University, Shanghai, China; Department of Bioinformatics and Biostatistics, Shanghai Jiaotong University, Shanghai, China; Center for Biomedical Informatics, Shanghai Children's Hospital, Shanghai, China. Electronic address: huilu@sjtu.edu.cn.
Abstract
MOTIVATION: New developments in high-throughput genomic technologies have enabled the measurement of diverse types of omics biomarkers in a cost-efficient and clinically-feasible manner. Developing computational methods and tools for analysis and translation of such genomic data into clinically-relevant information is an ongoing and active area of investigation. For example, several studies have utilized an unsupervised learning framework to cluster patients by integrating omics data. Despite such recent advances, predicting cancer prognosis using integrated omics biomarkers remains a challenge. There is also a shortage of computational tools for predicting cancer prognosis by using supervised learning methods. The current standard approach is to fit a Cox regression model by concatenating the different types of omics data in a linear manner, while penalty could be added for feature selection. A more powerful approach, however, would be to incorporate data by considering relationships among omics datatypes. METHODS: Here we developed two methods: a SKI-Cox method and a wLASSO-Cox method to incorporate the association among different types of omics data. Both methods fit the Cox proportional hazards model and predict a risk score based on mRNA expression profiles. SKI-Cox borrows the information generated by these additional types of omics data to guide variable selection, while wLASSO-Cox incorporates this information as a penalty factor during model fitting. RESULTS: We show that SKI-Cox and wLASSO-Cox models select more true variables than a LASSO-Cox model in simulation studies. We assess the performance of SKI-Cox and wLASSO-Cox using TCGA glioblastoma multiforme and lung adenocarcinoma data. In each case, mRNA expression, methylation, and copy number variation data are integrated to predict the overall survival time of cancer patients. Our methods achieve better performance in predicting patients' survival in glioblastoma and lung adenocarcinoma.
MOTIVATION: New developments in high-throughput genomic technologies have enabled the measurement of diverse types of omics biomarkers in a cost-efficient and clinically-feasible manner. Developing computational methods and tools for analysis and translation of such genomic data into clinically-relevant information is an ongoing and active area of investigation. For example, several studies have utilized an unsupervised learning framework to cluster patients by integrating omics data. Despite such recent advances, predicting cancer prognosis using integrated omics biomarkers remains a challenge. There is also a shortage of computational tools for predicting cancer prognosis by using supervised learning methods. The current standard approach is to fit a Cox regression model by concatenating the different types of omics data in a linear manner, while penalty could be added for feature selection. A more powerful approach, however, would be to incorporate data by considering relationships among omics datatypes. METHODS: Here we developed two methods: a SKI-Cox method and a wLASSO-Cox method to incorporate the association among different types of omics data. Both methods fit the Cox proportional hazards model and predict a risk score based on mRNA expression profiles. SKI-Cox borrows the information generated by these additional types of omics data to guide variable selection, while wLASSO-Cox incorporates this information as a penalty factor during model fitting. RESULTS: We show that SKI-Cox and wLASSO-Cox models select more true variables than a LASSO-Cox model in simulation studies. We assess the performance of SKI-Cox and wLASSO-Cox using TCGA glioblastoma multiforme and lung adenocarcinoma data. In each case, mRNA expression, methylation, and copy number variation data are integrated to predict the overall survival time of cancerpatients. Our methods achieve better performance in predicting patients' survival in glioblastoma and lung adenocarcinoma.