A comparison of two classes of methods for estimating false discovery rates in microarray studies

Area of Science:

• Genomics
• Bioinformatics
• Statistical Genetics

Background:

• Microarray studies aim to identify differentially expressed genes between populations.
• Estimating the false discovery rate (FDR) is crucial for interpreting these gene lists.
• Accurate FDR estimation relies on estimating π 1, the proportion of truly differentially expressed genes.

Purpose of the Study:

• To evaluate methods for estimating π 1 directly from test statistics, bypassing P-value computation.
• To compare the performance of these novel methods against established P-value-based approaches.
• To identify the most reliable methods for estimating π 1 in microarray data analysis.

Main Methods:

• Adapted existing methods to estimate π 1 from t- and z-statistics.
• Developed methods for direct π 1 estimation from various test statistics.
• Compared these new methods against two established π 1 estimation techniques using P-values.

Main Results:

• Methods for estimating π 1 varied significantly in bias and variability.
• Estimates derived directly from test statistics did not perform reliably.
• The 'convest' mixture model applied to P-values from a pooled permutation null distribution yielded the least biased and most variable estimates of π 1.

Conclusions:

• Computing P-values, while sometimes viewed as problematic, remains a necessary step for reliable π 1 estimation.
• The 'convest' method with permutation-based P-values offers superior accuracy for controlling FDR in gene expression studies.
• Direct estimation of π 1 from test statistics is not recommended based on current findings.