An immediate challenge in integrated genomic analysis involving several types of genomic factors all measured genome-wide is the ultra-high dimensionality. Screening all possible relationships among the genomic factors is an NP-hard problem; therefore in practice proper dimension reduction is necessary. In this paper we develop the Phenotype-Driven Dimension Reduction (PhDDR) approach to the analysis of gene co-expressions, and discuss its extensions to integration of other genetic factors. This approach is then illustrated by an application to gene co-expression analysis of treatment response of childhood leukemia.