hierarc (MatArray Toolbox)

MatArray Toolbox

hierarc

Hierarchical clustering, coded as a mex-file for performance.

Syntax

tree = hierarc(dist, type)
tree = hierarc(dist, type, flag)

Description

The function performs a hierarchical clustering of the distance matrix dist. The clustering can for instance be exported with


clustTV

for visualization with TreeView.

The parameter type determines the type of clustering performed:
0: single linkage
1: average linkage
2: complete linkage
3: Ward's method

The resulting tree is given as a 4-columns matrix, each line of the matrix being a node, as [ num val l r ]
num: number of leaves in the node (1 if the node is a leaf).
val: if the node is a leaf, the index of the leaf. Otherwise, the distance at the junction.
l: Node at the left.
r: Node at the right.
The tree starts on the first line.

If flag is present, tree is given as a UINT16 array. In this case val serves only as indexes for the leaves. This mode is marginally faster, it is essentially useful when the trees are meant to be used by another mex-file.

For the Ward's method, the distance matrix dist should be the squared euclidian distances, like the ones obtained with dl2c . The tree obtained is given with squared distances as well, which may fit or not. There is also a factor 2 between the distances shown and the real distances. To obtain the correct distances, identical to the ones obtained with linkage, simply type

tree = prettytree(tree);
tree(1:(size(tree,1)-1)/2,2) = sqrt(.5*tree(1:(size(tree,1)-1)/2,2));

Examples

Creation of a random ratio matrix with 100 genes and 10 slides, and of a name char array with the slide names.

ratio = randn(100,10);
name = { 'a' 'b' 'c' 'd' 'e' 'f' 'g' 'h' 'i' 'k' };

Say we only want to keep the genes where at least one absolute ratio is over 1.5

w = max(abs(ratio),[],2)>1.5;

Creation of an average linkage hierarchical clustering on the genes where w == 1. The similarity matrix obtained with corrcoef is transformed in a distance matrix

tG = hierarc(1-corrcoef(ratio(w,:)'),1);

Export of the clustering with the name abc for visualization with TreeView

clustTV('abc',ratio,name,w,tG,[]);

The clustering can be viewed by opening abc.cdt in the current directory with TreeView. The result will of course show little order as the ratio matrix was random.