Gene marker selection from gene expression profiles has been extensively investigated in the Bioinformatics literature. The aim is usually to find a compact set of genes potentially correlated to a particular disease, which can then be candidate targets for new drugs and treatments. Available gene expression data sets are often noisy and sparse, having a low number of patient samples, for which a high number of expressed genes is recorded. These characteristics may pose challenges in finding proper gene markers. Using some available gene expression data sets for cancer diagnosis, we experimentally try to understand the influence of their sparsity in the performance of two popular gene marker selection methods.