Robust biomarker identification in a two-class problem based on pairwise log-ratios

A new method, robust Pair-wise Log-Ratios (rPLR), is proposed for the identification of biomarkers, distinguishing between two groups of observations. The method can cope with the size effect problem, since it is based on log-ratios between the values of all pairs of variables. rPLR makes use of the variance of pairwise log-ratios, computed for the single groups and for all data jointly. When using a robust estimator of variance (or scale), the method is highly robust against data outliers. The robustness weights are aggregated and displayed in a diagnostics plot, which allows to reveal outlying cells in the data matrix.