Motivation: Modern methods such as microarrays, proteomics and metabolomics often produce datasets where there are many more predictor variables than observations. Research in these areas is often exploratory; even so, there is interest in statistical methods that accurately point to effects that are likely to replicate. Correlations among predictors are used to improve the statistical analysis. We exploit two ideas: non-negative matrix factorization methods that create ordered sets of predictors; and statistical testing within ordered sets which is done sequentially, removing the need for correction for multiple testing within the set. Results: Simulations and theory point to increased statistical power. Computational algorithms are described in detail. The analysis and biological interpretation of a real dataset are given. In addition to the increased power, the benefit of our method is that the organized gene lists are likely to lead better understanding of the biology.