Limits...
Testing significance relative to a fold-change threshold is a TREAT.

McCarthy DJ, Smyth GK - Bioinformatics (2009)

Bottom Line: The most widely used methods successfully test whether the true differential expression is different from zero, but give no assurance that the differences found are large enough to be biologically meaningful.We present a method, t-tests relative to a threshold (TREAT), that allows researchers to test formally the hypothesis (with associated p-values) that the differential expression in a microarray experiment is greater than a given (biologically meaningful) threshold.When the magnitude of differential expression is taken into account, TREAT improves upon the false discovery rate of existing methods and identifies more biologically relevant genes.

View Article: PubMed Central - PubMed

Affiliation: The Walter and Eliza Hall Institute of Medical Research, 1G Royal Parade, Parkville, Victoria 3050, Australia.

ABSTRACT

Motivation: Statistical methods are used to test for the differential expression of genes in microarray experiments. The most widely used methods successfully test whether the true differential expression is different from zero, but give no assurance that the differences found are large enough to be biologically meaningful.

Results: We present a method, t-tests relative to a threshold (TREAT), that allows researchers to test formally the hypothesis (with associated p-values) that the differential expression in a microarray experiment is greater than a given (biologically meaningful) threshold. We have evaluated the method using simulated data, a dataset from a quality control experiment for microarrays and data from a biological experiment investigating histone deacetylase inhibitors. When the magnitude of differential expression is taken into account, TREAT improves upon the false discovery rate of existing methods and identifies more biologically relevant genes.

Availability: R code implementing our methods is contributed to the software package limma available at http://www.bioconductor.org.

Show MeSH
FDRs for six different gene selection statistics from the analysis of simulated data. The rates are the means of actual FDRs for 1000 simulated datasets.
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
getmorefigures.php?uid=PMC2654802&req=5

Figure 1: FDRs for six different gene selection statistics from the analysis of simulated data. The rates are the means of actual FDRs for 1000 simulated datasets.

Mentions: Figure 1 shows that, when used to analyse the simulated data, TREAT has the lowest FDR overall. In general, the statistics based on moderated t do best for small numbers of genes, whereas methods based on fold-change do best for large numbers of genes. TREAT successfully combines the advantages of both types of statistic, matching the best statistics at the two extremes, and having clearly lowest FDR of all the methods for the intermediate range of 250–500 genes selected. Above about 600 genes selected, all the methods which use fold-change are similar. Table 1 shows that TREAT has the highest area under the ROC curve, confirming it is the best overall method for these data. Ordinary t is by far the worst-performing method.Fig. 1.


Testing significance relative to a fold-change threshold is a TREAT.

McCarthy DJ, Smyth GK - Bioinformatics (2009)

FDRs for six different gene selection statistics from the analysis of simulated data. The rates are the means of actual FDRs for 1000 simulated datasets.
© Copyright Policy - creative-commons
Related In: Results  -  Collection

License
Show All Figures
getmorefigures.php?uid=PMC2654802&req=5

Figure 1: FDRs for six different gene selection statistics from the analysis of simulated data. The rates are the means of actual FDRs for 1000 simulated datasets.
Mentions: Figure 1 shows that, when used to analyse the simulated data, TREAT has the lowest FDR overall. In general, the statistics based on moderated t do best for small numbers of genes, whereas methods based on fold-change do best for large numbers of genes. TREAT successfully combines the advantages of both types of statistic, matching the best statistics at the two extremes, and having clearly lowest FDR of all the methods for the intermediate range of 250–500 genes selected. Above about 600 genes selected, all the methods which use fold-change are similar. Table 1 shows that TREAT has the highest area under the ROC curve, confirming it is the best overall method for these data. Ordinary t is by far the worst-performing method.Fig. 1.

Bottom Line: The most widely used methods successfully test whether the true differential expression is different from zero, but give no assurance that the differences found are large enough to be biologically meaningful.We present a method, t-tests relative to a threshold (TREAT), that allows researchers to test formally the hypothesis (with associated p-values) that the differential expression in a microarray experiment is greater than a given (biologically meaningful) threshold.When the magnitude of differential expression is taken into account, TREAT improves upon the false discovery rate of existing methods and identifies more biologically relevant genes.

View Article: PubMed Central - PubMed

Affiliation: The Walter and Eliza Hall Institute of Medical Research, 1G Royal Parade, Parkville, Victoria 3050, Australia.

ABSTRACT

Motivation: Statistical methods are used to test for the differential expression of genes in microarray experiments. The most widely used methods successfully test whether the true differential expression is different from zero, but give no assurance that the differences found are large enough to be biologically meaningful.

Results: We present a method, t-tests relative to a threshold (TREAT), that allows researchers to test formally the hypothesis (with associated p-values) that the differential expression in a microarray experiment is greater than a given (biologically meaningful) threshold. We have evaluated the method using simulated data, a dataset from a quality control experiment for microarrays and data from a biological experiment investigating histone deacetylase inhibitors. When the magnitude of differential expression is taken into account, TREAT improves upon the false discovery rate of existing methods and identifies more biologically relevant genes.

Availability: R code implementing our methods is contributed to the software package limma available at http://www.bioconductor.org.

Show MeSH