While it is important to use gene expression clustering algorithms to explore the similarity among the available experiments, it is, sometimes, more important to study the disimilarity between samples, particularily for comparison experiments in cancer research. Given the tissue categories derived either from tissue origins or their pathologies, or given the clusters obtained from clustering algorithms, we naturally would like to know which gene mostly associates with the given classification, or the weight for each gene based on its discriminative ability for the sample classification. Here we introduce some of most common methods that we've employed in some of studies.

1. Weighted Method. Assume we have K categories (or clusters) for a set of samples, a discriminative weight for each gene can be evaluated by,
w = Average(BD) / (Average(WD) + a)
where Average(BD) is the average of the between cluster Euclidean distance for all pairs of clusters (total of (K*K-1)/2 pairs), and Average(WD) is the weighted average of the within-cluster distance (weighted by the number of samples in the cluster). The within-cluster distance is the average distances of all pairs samples in the cluster. a is a small constant to prevent zero denominator case. Genes may then be ranked on the basis of w. The equation for weight w is not only designed to evaluate discriminative ability for single gene, but also capable of evaluate discriminative ability for 2 or more genes together. Theoretically, weight w < 1 indicates the gene of no discriminative ability. Statistically, the significance of the weighted value can be assessed by randomly permute the sample labeling, and the critical value can thus be established.

2. t-statistic or F-statistic method.

3. TNoM score.