Assessing Signal-to-Noise in Quantitative Proteomics: Multivariate Statistical Analysis in DIGE Experiments - Difference Gel Electrophoresis (DIGE): Methods and Protocols - page 43

But without the utility of visualizing this very complex dataset
with PCA, this high degree of potentially confounding background
variation would likely have remained undetected.
4. Notes
1. Statistical power is the ability to visualize an X -fold magnitude
change (effect size) at a Y -% confi dence interval (e.g., p < 0.05)
and be “correct” (usually expressed as power of 0.8, or 80% of
the time it is “correct”). Statistical power depends heavily on
the analytical variation of the instrumentation as well as the
number of independent (biological) replicates. For example,
the experimental noise of DIGE is extremely low owing to the
internal standard experimental design, enabling statistically
powered experiments with very few biological replicates ( 1 ).
2. Technical and biological replicates are important for ensuring
the accuracy and biological signifi cance of quantitative measure-
ments. Technical replicates are necessary to control for variation
in the analytical measurement. However, biological replicates
are vital to assess whether or not changes in protein abundance/
modifi cation are descriptive of the biology rather than arising
from unanticipated sources of experimental variation (e.g.,
inter-subject variation, sample preparation variation, analytical
variation of the instrument). Technical replicates provide con-
fi dence in the result from the tested samples but do not provide
any confi dence to the biological relevance. Determining
whether any observed changes come from the biology rather
than technical variation can only be assessed with independent
biological replicates. In cases where technical noise has been
demonstrated to be suffi ciently low, then biological replicates
can also serve as technical replicates, which is what is typically
done with the DIGE platform.
3. Pooling independent (biological) replicates should be done
with extreme caution with respect to the statistical power of
the resulting data and is not advised. If it is known a priori that
technical variation is low between samples, then pooling can be
effective, but if it is high, then a pooling strategy can be disas-
trous. Even with the low analytical noise of DIGE, the pooled
N = 1 comparison on a single gel assumes that the averaging of
populations is refl ective of biological signal. In some cases it
may be valid to create subpools from a larger experiment, to
either produce suffi cient material or to minimize costs (the
“economics of proteomics”), but in these cases it is still essen-
tial to maintain some degree of individualization of samples to
retain statistical power ( 8 ).