Implications

1. Clustering is sensitive and may reveal associations that are otherwise obscured.

2. The validity of clustering results can be difficult to assess.

3. Clustering techniques can be improved if used in conjunction with statistical techniques. For example, use a statistical filter before clustering to remove transcripts that have not significantly changed.

4. If transcripts are classed together as signatures, that knowledge can be built into a statistical model. In this way we test whether each biologically relevant signature is influenced by a drug instead of testing each transcript. The advantage of this classification is that we test a modest number of hypotheses instead of a hypothesis for each transcript. As a result, our multiple testing problem is reduced and the Bonferroni multiple comparison correction method is more appropriate. Signature classes are best defined through empirical testing. However, clustering methods can be used to define signatures, especially if the number of clusters is rigorously defined by a GAP statistic (Tibshirani et al. 2000; Kerr and Churchill 2001; Dudoit and Fridlyand 2002, 2003)

0 0

Post a comment