Approaches to reduce false positives and false negatives in the analysis of microarray data: applications in type 1...

As studies of molecular biology system attempt to achieve a comprehensive understanding of a particular system, Type 1 errors may be a significant problem. However, few investigators are inclined to accept the increase in Type 2 errors (false positives) that may result when less stringent statistical cut-off values are used. To address this dilemma, we developed an analysis strategy that used a stringent statistical analysis to create a list of differentially expressed genes that served as "bait" to "fish out" other genes with similar patterns of expression.

Comparing two strains of mice (NOD and C57Bl/6), we identified 93 genes with statistically significant differences in their patterns of expression. Hierarchical clustering identified an additional 39 genes with similar patterns of expression differences between the two strains. Pathway analysis was then employed: 1) identify the central genes and define biological processes that may be regulated by the genes identified, and 2) identify genes on the lists that could not be connected to each other in pathways (potential false positives). For networks created by both gene lists, the most connected (central) genes were interferon gamma (IFN-gamma) and tumor necrosis factor alpha (TNF-alpha). These two cytokines are relevant to the biological differences between the two strains of mice. Furthermore, the network created by the list of 39 genes also suggested other biological differences between the strains.

Taken together, these data demonstrate how stringent statistical analysis, combined with hierarchical clustering and pathway analysis may offer deeper insight into the biological processes reflected from a set of expression array data. This approach allows us to 'recapture" false negative genes that otherwise would have been missed by the statistical analysis.