NCIS: a network-assisted co-clustering algorithm to discover cancer subtypes based on gene expression

Cancer subtype information is critically important for designing more effective treatments. In this thesis, we introduce a new co-clustering algorithm for cancer subtype identification, which combines the information of gene networks to simultaneously group samples and genes into biologically meaningful clusters. We call our method network-assisted co-clustering for the identification of cancer subtypes (NCIS). Prior to clustering, we assign weights to genes: those that play key roles in the network and/or show significant variations among samples would be prioritized. This new approach allows us to rely more on genes that are informative and representative by including the weights as an importance indicator in the clustering step. Here we introduce a new weighted co-clustering method based on semi-nonnegative matrix tri-factorization. We evaluated the effectiveness of the algorithm on large-scale Glioblastoma multiforme (GBM) and breast cancer (BRCA) datasets from TCGA and on simulated datasets. We found that our NCIS method can achieve more reliable results with respect to the clinical features compared to conventional semi-nonnegative matrix tri-factorization methods and consensus clustering. We also train two classifiers for GBM and BRCA subtypes identification based on NCIS's results. This new method will be very useful to comprehensively detect subtypes that are otherwise obscured by cancer heterogeneity, from various types of cancers based on high-throughput and high-dimensional gene expression data