A systematic comparison of genome-scale clustering algorithms

BMC Bioinformatics

Table 1 Overview of algorithms tested

			Allows Overlapping Clusters
				Pre-specified Number of Clusters ( k )
					Thresholded Correlations
Method	Type	Result Range				Parameters Tested
Ward	Hierarchical	k		Y		Average cluster size
Average	Hierarchical	k		Y		Average cluster size
McQuitty	Hierarchical	k		Y		Average cluster size
Complete	Hierarchical	k		Y		Average cluster size
k-Means	Partitioning	k		Y		Number of clusters
SOM	Neural network	k		Y		Grid size/type^a
QT Clust	Partitioning	24-385				Maximum cluster diameters
CAST	Graph-based	1-6162			Y	Threshold
CLICK	Graph-based	4-32				Cluster homogeneity
WGCNA	Graph-based	4-160				Power, Module detection method
NNN	Graph-based	23-52	Y^b			Minimum neighborhood size
k-Cliques Communities	Graph-based	1-68	Y		Y	Threshold, Clique size
Maximal Clique	Graph-based	1,000-64,000	Y		Y	Threshold
Paraclique	Graph-based	8-615	Y^c		Y	Threshold, Glom factor

Clustering methods are listed by name, along with the type of algorithm, and a general listing of parameters tested. Number of clusters in the result, given the parameters and data set tested, is only provided here as an approximate figure. Empty results are obviously not included. ^aGrid type can be either rectangular or hexagonal, in an m x n layout. We tested both types, but used an m x m layout for simplicity (k = m²). ^bRarely occurs in practice. On this data set we observed no overlap with NNN. ^cOptional, not used in this analysis.

ISSN: 1471-2105