Biclustering is a very interesting technique for unsupervised analysis of gene expression data. It aims to discover subsets of genes having similar behavior on a subset of conditions; such biclusters are related to close biological functions. The majority of existing biclustering algorithms are based on statistical criteria (e.g. size, coherence and structure...) to define the bicluster quality. However these measures are not directly related to biological knowledge and they may produce results that are difficult to interpret by a biologist. In fact, it is recognized that the integration of some biological information to guide the extraction of the biclusters ensures their relevance and their non-triviality. Therefore this work proposes a local search method that relies on ontological knowledge to cluster the genes while a correlation measure is used to cluster the conditions. An experimental study is performed using real microarray datasets. The results demonstrate the importance of the integration of the biological knowledge in the search process, to promote the discovery of non-trivial and biologically relevant biclusters.