Availability: The R package tspair is freely available from Bioconductor: http://www.bioconductor.org

Top scoring pairs (TSPs) are pairs of genes that accurately classify patients into clinically relevant groups based on their ranks (Geman et al., 2004; Tan et al., 2005; Xu et al., 2005). The basic idea is to search among all pairs of genes, and look for genes whose ranking most consistently switches between two groups. To understand how the classification scheme works, consider the simulated gene expression data in Figure 1. In this figure there are two groups of arrays, separated by the black line. These groups could represent healthy patients versus cancer patients, or two distinct subtypes of cancer. For all but one array in Group 1, Gene 1 has higher expression than Gene 2, and the reverse is true in Group 2. In this case, Genes 1 and 2 form a classifier based on their relative levels of expression. A new sample where the gene expression for Gene 1 was higher than the gene expression for Gene 2 would be classified as Group 1.

The function tspsig() can be used to calculate the significance of a TSP classifier by permutation as described in Geman et al. (2004). The class labels are permuted, a new TSP is calculated for each permutation, and the null scores are compared with the observed TSP score to calculate a P-value. Since the maximum score is calculated for each null permutation, tspsig() performs a test of the null hypothesis that no TSP classifier is better than random chance.

Once a TSP has been calculated, the tspplot() function can be used to visualize the classifier. The resulting TSP figure (Fig. 2) plots the expression for the first gene in the pair versus the expression for the second gene in the pair. The true group difference is indicated by the color of the points, and the score for the TSP classifier is shown in the title of the plot. The black 45^○ line indicates the classification from the TSP; the better the black line separates the colors the better the accuracy of the TSP.

A major advantage of the TSP approach is that predictions are very simple and can be easily calculated either by hand or using the built-in functionality of the tspair package. In this example, the expression value for ‘Gene5’ is greater than the expression value for ‘Gene338’ much more often for the diseased patients. In a new dataset, when the expression for ‘Gene5’ is greater than the expression for ‘Gene338’ I predict that the patient will be diseased. The tspair package can be used to predict the outcomes of new samples based on new expression data. The new data can take the form of a new expression matrix, or an expression set object. The R function predict() searches for the TSP gene names from the original tspcalc() function call, and based on the row names or featureNames of the new dataset identifies the genes to use for prediction. If multiple TSPs are reported, the default is to predict with the TSP achieving the top tie-breaking score (Tan et al., 2005), but the user may also elect to use a different TSP for prediction.