This pipeline calculates clusters based on a consensus nonnegative matrix factorization (NMF) clustering method [1][2]. This pipeline has the following features:

Convert input data set to nonnegativity matrix by column rank normalization.

Classify samples into consensus clusters.

Determine differentially expressed marker miRs for each subtype.
We filtered the data to 150 most variable miRs. Consensus NMF clustering of 96 samples and 150 miRs identified 3 subtypes with the stability of the clustering increasing for k = 2 to k = 8 and the average silhouette width calculation for selecting the robust clusters.
SampleName  cluster  silhouetteValue 

TCGABC407201  1  0.016 
TCGABCA10Q01  1  0.19 
TCGABCA10R01  1  0.021 
TCGABCA10T01  1  0.034 
TCGABCA10X01  1  0.3 
TCGABCA11001  1  0.2 
TCGABCA11201  1  0.14 
TCGABDA3ER01  1  0.2 
TCGACC525801  1  0.16 
TCGACC526001  1  0.076 
SampleName  K=2  K=3  K=4  K=5  K=6  K=7  K=8 

TCGABC407201  1  1  1  1  1  1  1 
TCGABCA10Q01  1  1  1  3  3  2  2 
TCGABCA10R01  1  1  3  3  3  2  2 
TCGABCA10S01  1  3  3  4  4  2  3 
TCGABCA10T01  1  1  3  4  4  3  4 
TCGABCA10X01  1  1  1  1  1  4  5 
TCGABCA11001  1  1  1  3  6  5  5 
TCGABCA11201  1  1  1  3  3  2  2 
TCGABDA3ER01  1  1  1  3  6  5  7 
TCGACC525801  1  1  1  1  1  4  5 
Samples most representative of the clusters, hereby called core samples were identified based on positive silhouette width, indicating higher similarity to their own class than to any other class member. Core samples were used to select differentially expressed marker miRs for each subtype by comparing the subclass versus the other subclasses, using Student's ttest.
Composite.Element.REF  p  difference  q  subclass 

HSAMIR1247  1.1e06  2.7  0.000022  1 
HSAMIR483  0.0019  2.7  0.011  1 
HSAMIR200B  8.6e06  2.4  0.00011  1 
HSAMIR200A  8.8e06  2.4  0.00011  1 
HSAMIR134  6.7e09  2.1  8.8e07  1 
HSAMIR323  9e06  2.1  0.00011  1 
HSAMIR429  0.00034  2  0.0024  1 
HSAMIR493  6.6e08  2  3.5e06  1 
HSAMIR758  7.6e08  2  3.6e06  1 
HSAMIR203  0.000052  1.9  0.00045  1 
miRseq (MIMATs) of RPM value (reads per million reads aligned to miRBase mature) with log2 transformed was as the input data for the clustering
Nonnegative matrix factorization (NMF) is an unsupervised learning algorithm that has been shown to identify molecular patterns when applied to gene expression data [1],[2]. Rather than separating gene clusters based on distance computation, NMF detects contextdependent patterns of gene expression in complex biological systems.
We use the cophenetic correlation coefficient [1] to determine the cluster that yields the most robust clustering. The cophenetic correlation coefficient is computed based on the consensus matrix of the CNMF clustering, and measures how reliably the same samples are assigned to the same cluster across many iterations of the clustering lgorithm with random initializations. The cophenetic correlation coefficient lies between 0 and 1, with higher values indicating more stable cluster assignments. We select the number of clusters k based on the largest observed correlation coefficient for all tested values of k.
Silhouette width is defined as the ratio of average distance of each sample to samples in the same cluster to the smallest distance to samples not in the same cluster. If silhouette width is close to 1, it means that sample is well clustered. If silhouette width is close to 1, it means that sample is misclassified [3][4].
This is an experimental feature. The full results of the analysis summarized in this report can be downloaded from the TCGA Data Coordination Center.