RandomForests (RF) is an ensemble method that has become widely accepted in machine learning and bioinformatics communities during the last few years. Its predictive strength, along with some of the utilities generated by its learning mechanism, has made RF a powerful data mining tool for uncovering complex patterns in high dimensional data. In this presentation we will explore the internal workings of RF learning mechanism when applied to the binary classification problem for high dimensional, low sample size data and will describe its strengths and weaknesses. We will propose a search procedure that explores the high dimensional feature space in a blockwise manner, using RF as the search engine, in order to uncover complex interactions that accurately predict the binary outcome. The proposed method will be illustrated with an application to the detection of informative gene interactions in microarray data.