Application of independent component analysis to microarray data

Independent component analysis (ICA) is introduced from the viewpoint of maximal information transfer for single neurons. This historical motivation for the development of ICA may be interesting from the viewpoint of independent agents because each neuron can be seen as a single agent. We compare the performance of ICA with principal component analysis (PCA) for detecting coregulated gene groups in microarray data measured during different stages of the yeast cell cycle. PCA was shown to find gene groups for which gene expression fluctuates periodically with the cell cycle (N.S. Holter et al., 2001). This result, however, required a very long series of unmotivated preprocessing steps. To compare the performance of both methods, the principal components and the independent components were computed without arbitrary data processing. Only ICA, but not PCA, found coregulated gene groups, suggesting that ICA is more successful in finding coregulated gene groups than PCA.