Assessing Signal-to-Noise in Quantitative Proteomics: Multivariate Statistical Analysis in DIGE Experiments - Difference Gel Electrophoresis (DIGE): Methods and Protocols - page 32

noise is low, the biological replicates can control for both technical
and biological noise. Due to the low amount of technical noise
demonstrated for DIGE ( 1 ), this technology platform enables statis-
tically powered experiments (see Note 1) using a relatively low
number of biological replicates without the need for additional tech-
nical replicates of each sample (see Note 2).
In the simplest type of quantitative proteomics experiment,
there are two conditions being measured, typically some experi-
mental condition and a control. Even here the need for indepen-
dent biological repetition remains paramount; it is insuffi cient to
simply multiplex a single experimental and control sample into the
same analytical run (e.g., DIGE gel) and be able to determine if
the observed change represents a relevant biological signal or tech-
nical/biological noise. Even pooling independent samples does
not alleviate this issue, as once samples are pooled the ability to
distinguish signal from noise is lost (see Note 3).
Using the requisite number of biological replicates, the most
commonly used statistical test for quantitative proteomic changes
is the univariate Student's t -test, whereby the distribution about
two means is compared with the magnitude of difference between
these means, and the resulting p -value refl ects the likelihood that
the measurements are derived from the same distribution (the null
hypothesis). When experimental conditions become greater than
two, then the univariate analysis of variance (ANOVA) test is com-
monly invoked (the t -test is a special case of the ANOVA test).
Despite their commonplace usage in quantitative proteomics, these
univariate tests only assess changes on a feature-by-feature basis; a
single species relative to itself across all samples and conditions.
Although multiple testing correction algorithms are available to
compensate for univariate tests performed within large datasets
(e.g., Bonferroni and Benjamini-Hochberg methods), univariate
tests do not take into account the variation present in the global
experimental system. Especially in the small-sample-size regime,
the likelihood of measuring a change by chance increases regard-
less of what the univariate p -value suggests. Thus, knowledge of
the global variation is essential and in some cases can infl uence the
likelihood that univariate changes are biologically signifi cant. The
univariate p -value is only a guide, a means to an end but not an end
unto itself.
In contrast, multivariate tests enable the visualization of the
experimental variation on a global scale, analyzing all of the vari-
ables simultaneously. Technical noise (poor sample prep, run-to-run
variation) and biological noise (normal differences between samples,
especially present in clinical samples) are almost always associated
with analytical datasets in quantitative proteomics and may well
override any variation that arises due to actual differences related
to the biological questions being tested. Multivariate tests can
highlight major sources of variation within a dataset, and when
performed in an unsupervised fashion, can test if this variation is