Many statistical methods have been proposed to identify biomarkers from gene expression profiles. However, from expression data alone, statistical methods often fail to identify biologically meaningful biomarkers related to a specific biological process or disease under study. In this paper, we develop a novel strategy, namely knowledge-driven multi-scale independent component analysis (ICA), to infer regulatory signals and identify biologically relevant biomarkers from microarray data. Specifically, based on partial prior knowledge and clustering results, we apply ICA to find the most knowledge relevant linear regulatory mode in each subset of genes and then extract associated biomarkers according to their weighted loading factors. We have applied our method to a yeast cell cycle microarray dataset to find cell cycle regulated biomarkers. The experimental results indicate that our knowledge-driven multi-scale ICA method outperforms both baseline ICA method and correlation method significantly.