Dependency and purity in large-scale statistical significance testing : a DNA microarray perspective