Identification of Genes Affected during Myeloid Differentiation Induced by Valproic Acid and Retinoic Acid: The Choice of Array Pre-Processing and Normalization Algorithm Has a Critical Effect on Results.

Valproic acid is an anti-epileptic drug with histone deacetylase inhibitor activity currently in clinical trial for the treatment of leukemia. In order to understand the effects of this agent in myeloid differentiation and leukemia therapy we determined the changes in gene expression that occurred when the myeloid cell line ML-1 was treated with valproic acid (VPA), all-trans retinoic acid (ATRA) or the combination for 24 or 96 hours. Each condition was compared to no drug treatment at each time point, as full biological triplicates. RNA samples were labeled and hybridized to Affymetrix HGU133A genechips. The data was processed using the Affymetrix MAS5.0 calculated intensity signals or was pre-processed using three different algorithms to normalize data: GC-RMA (GeneChip-Robust Microarray Analysis), RMA (Robust Microarray Analysis) or MBEI (Model-Based Expression Index), using the original Affymetrix CEL files. From the resulting normalized data sets, we then identified genes that were up or down regulated at least two-fold by the combination VPA + ATRA treatment when compared to the no drug condition or either drug alone. If any of these genes were affected by either ATRA or VPA, we selected only those with an additional change of 50% or more with the combination therapy. There were striking differences in the gene sets obtained from each normalization algorithm with respect to both list size and content. GC-RMA identified 100 genes significantly differentially expressed according to the above criteria at the 24 hour time point, MBEI found 19 and MAS5.0 found 39. Using the RMA algorithm, 33 genes were determined to be significantly affected by the combination treatment and all of these were also found with GC-RMA suggesting that GC-RMA has a greater sensitivity than RMA. This may be because GC-RMA, a refinement of RMA, takes into account mismatched probe set signals and probe set sequence content. Only 12 genes were common to all lists, including endothelin receptor B, myosin binding protein H, IL8, and G0S2. We are currently using real-time PCR to validate the extent of differential regulation of the commonly identified genes in reponse to VPA+ATRA. This will be contrasted with the levels of regulation of gene sets found by only one or two algorithms in order to determine the sensitivity and specificity of these methods. These data indicate the potential power of new bioinformatic tools for microarray analysis and the need for systematic validation to ensure biological relevance.