tanghaibao / treecut

Find nodes in hierarchical clustering that are statistically significant

treecut's Issues

Study whether the hierarchical clustering method is suited for grouping genotypes with markers

Explore the alternative method, for example using STRUCTURE, this is explored in the following paper

http://www.ncbi.nlm.nih.gov/pubmed/21472410
Determination of genetic structure of germplasm collections: are traditional hierarchical clustering methods appropriate for molecular marker data?

In population genetics, STRUCTURE is much more useful than hierarchical clustering. Correlating phenotype information with STRUCTURE results will become emergent problem to solve.

Study this and discuss possible implementations.

Study the statistical properties when performing tests on internal node

Currently only two tests are performed on internal nodes, and quite simplistic: t-test (for continuous data), and Fisher's exact test (for categorical data). Study the theory behind this package:

http://www.is.titech.ac.jp/~shimo/prog/pvclust/

and possibly porting some of the functions over.

Incorporate clustering uncertainties

Based on Jingping's idea below:

Since the program takes an existed tree and that tree construction itself can be fishy sometimes. I wonder if the algo might also try to consider alternative tree shapes to improve cutting. Search of tree shape itself is NP-hard problem, but some simplified thing might be doable. For example, the algo could take not only the tree shape but also node support for the input tree. Then for nodes whose support is low the algo could try to switch around its "children". For example in the figure attached, say if the support for node N4 is low, the algo could try the alternative clustering shown in the right part of the figure, which in this case might narrow done the defining of the group (from 130, 150, 90 to 130, 150). I assume in some other cases this could result in expansion of group definition too.