Comparison of Integrative Clustering Methods for Determining Molecular Subtypes

• Activity Number: 418
• Type: Contributed
• Date/Time: Tuesday, August 5, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Section on Teaching of Statistics in the Health Sciences
• Author(s): Milan Bimali*+
• Keywords: Integrative Clustering ; Semi-supervised ; Training/Testing ; genomics ; bioinformatics

Abstract

Traditionally, molecular profiling has been based on a single data type, primarily gene expression data. Recently proposed clustering methods use clinical and 'omic information in a single, comprehensive, clustering algorithm. Two of these methods have been used extensively; iCluster and non-negative matrix factorization (NNMF). However, these algorithms are plagued with the issue of over-fitting. Thus, we propose the following extensions: 1) incorporating a semi-supervised aspect for feature selection; 2) use of cross-validation to prevent over-fitting; and 3) a training/testing framework to assess the performance of the developed signature. We applied our two modified algorithms to The Cancer Genome Atlas (TCGA) ovarian cancer data (499 patients, 27338 DNA methylation probes, 90797 mRNA gene expression probes), where both approaches suggested two subtypes in the training data. However, when the signature was applied to the testing data no significant difference (p > 0.05) in the clinical outcome was observed between the two molecular subtypes. Future work is on-going to develop a powerful, integrative clustering algorithm to determine clinically relevant molecular subtypes.