[Analysis of gene expression data regulated by clock-genes: methodological approach and optimization]. / Analyse de données d'expression transcriptomiques rythmées par des gènes-horloge: approche méthodologique et optimisation.

ABSTRACT

BACKGROUND: In microarray data, wide-scale correlations are numerous and increase the number of genes correlated to a test condition (phenotype, mutation status, etc.) either positively or negatively. Several methods have been developed to limit the effect of such correlations on the false discovery rate, but these may reject too many genes that have a mild or indirect impact on the studied condition. We propose here a simple methodology to correct this spurious effect without eliminating weak but true correlations. RESULTS: This methodology was applied to a microarray dataset designed to distinguish heterozygous BRCA1 mutation carriers from non-carriers. As our samples were collected at different times in the morning, we evaluated the effect of correlations due to circadian rhythm. The circadian system is a well-known correlation network, regulated by a small number of period genes whose expression varies throughout the day in predictable ways. The downstream effects of this variation on the expression of other genes, however, are incompletely characterized. We used two different strategies to correct this correlation bias, by either dividing or multiplying the expression of correlated genes by the expression of the considered period gene according to the sign of the correlation between the period gene and correlated gene (respectively positive or negative). CONCLUSIONS: We observed a linear relationship between the number of false-positive/negative genes and the strength of the correlation of the candidate gene to the test condition. BRCA1 was highly correlated to the period gene Per1; our correction methodology enabled us to recover genes coding for BRCA1-interacting proteins which were not selected in the initial direct analysis. This methodology may be valuable for other studies and can be applied very easily in case of well-known correlation networks.