Study Design in DIGE-Based Biomarker Discovery - Difference Gel Electrophoresis (DIGE): Methods and Protocols - page 198

2. Control of the false discovery rate (FDR), i.e., the expected
proportion of the identifi ed proteins not related with the clinical
outcome among all identifi ed proteins (see ref. ( 6 )). If an FDR
of 0.05 is controlled, we can expect in the long run a propor-
tion of 5% false-positives among the signifi cant proteins.
In this chapter, a protein is called “identifi ed” if the adjusted p
value is smaller than the prespecifi ed signifi cance level for the type 1
error. Controlling the FDR generally relaxes the multiple testing
criteria compared with controlling the FWER and may increase the
number of truly identifi ed proteins (methods to control the FDR
and further investigations are, e.g., given in refs. ( 6-9 )). The FDR
algorithm has been implemented in the DeCyder software for anal-
ysis of DIGE experiments as described below.
Irrespective of the method used to deal with the problem of
multiple testing, mostly a large number of parallel analyses (sample
size) per group are needed to get statistically interpretable results.
When testing only a single protein, the sample size problem in case
of a two-sample t test is formulated as the number of gels needed
to ensure a specifi c power (probability of identifying a protein that
is, in truth, related with the clinical outcome, called “effective”
protein in this chapter) to detect a standardized effect size q
(difference of group means in terms of the common standard
deviation) with a prespecifi ed type 1 error. In clinical studies, the
power is generally set to 0.80. For the example of a two-sided two-
sample t test with a-level 0.05, we would need a sample size of
about 17 gels/group to detect a difference between group means
of size of 1 standard deviation with 80% power. Increasing the
effect size to 1.5 or 2 would reduce the sample size of each group
to 9 or 6, respectively. However, an effect size of 1.5 or more is
regarded as a very large effect in clinical studies. The main problem
for the sample size calculation is that the true effect size is unknown.
Sample size calculations are therefore often based on the minimum
clinical relevant effect size an experimenter wants to detect with
their study. On the other hand, before starting a clinical trial,
frequently pilot studies are performed to get an intention of the
effect size. In the proteomic setting, where thousands of proteins
are tested simultaneously, things again get more complicated. The
fi rst problem is that, as for the type 1 error rate, also the concept of
power can be generalized in various ways in the context of multiple
hypothesis testing. Usually, the power is defi ned as the proportion
of truly identifi ed effective proteins among all investigated effec-
tive proteins. There have been several publications on sample size
estimation in microarray studies. For example, Pawitan et al. ( 10 )
investigated the relationships between FDR, power, and sample
size. Although their approach may be useful in deciding on the
sample size, no direct algorithm was provided. Jung ( 11 ) gives an
algorithm to calculate the sample size for a prespecifi c power based
on Storey's asymptotic results (see also refs. ( 8, 9 )). The power was
defi ned as discussed above.