Interpreting Principal Components

It is our belief that PCA is one of a set of tools that can help investigators better understand the sources of variation present in a microarray dataset. Each principal component measures, and to some degree models, some source of variance observed in the dataset. At the simplest level one observes the variance explained by each principal component, and perhaps takes note of that principal component's eigen vector. By analyzing each principal component a bit further we can better appreciate what is driving that component's variation. Our strategy is to identify the data points (genes) at the extremes of each principal component axis, and then determine which conditions are driving those outlier genes to be significantly differentially expressed. Beyond that, the covariate factors that correlate well with the significant conditions can be identified, so we may hypothesize that they are substantial sources of gene expression variation.