Biology Reference
In-Depth Information
In cases such as this where the biological signal of interest is
not readily evident but the technical signal is low, additional data
reduction may be helpful. In this case, low signals of interest may
be revealed if the background noise (whether it be technical or
biological in nature) can be removed from the PCA, and this can
effectively be done by applying a statistical fi lter. Analysis of vari-
ants (ANOVA) is an appropriate univariate statistical test to use for
this purpose when multiple variables are under consideration
because it identifi es features that are changing in one group relative
to any of the others without specifying any pairwise comparisons
(see Note 11).
In the example from Loh et al. (
8
), applying an ANOVA fi lter
(
p
< 0.05) to the dataset defi ned a subset of 168 features that were
of biological signifi cance for the experiment. Although this imposes
the bias of the biological experimental design onto the PCA, it
only selects those features that were changing signifi cantly in one
of the four classifi cation groups relative to the other three, and for
each feature this could be an independent classifi cation. By remo-
ving those variations that contributed to the background noise in
the original set of 639 features, PC1 (genotype) increased to 81%
of the remaining variation, and now PC2 was found to organize
the samples based on pH treatment, accounting for 6.4% of the
variation or about ten features (see Fig.
2c
). Thus, the experimental
conditions appeared to induce changes of interest, but they were
too subtle in magnitude and/or number to be visualized in the
context of greater sources of variation in the experiment.
In this third example (also from
H. pylori
, now cultured under
different medium conditions), the signal is so low from any of the
biological sources of variation that now dye-labeling bias appears to
be ordered by the fi rst principal component in an unfi ltered analysis
of 977 matched features across a six-gel DIGE experiment (see
Fig.
3
). Dye-labeling bias is known to exist and is typically con-
trolled for using a dye-swap labeling scheme in the experimental
design (see Note 12). That Cy3/Cy5 labeling bias appears here as
the driving force behind PC1 (see Fig.
3b
) indicates not only very
low biological signal but also very low noise because this anticipated
bias typically contributes very little to the overall variation. Thus,
the fi nding of organization by dye labeling in PC1 is expected in the
absence of signifi cant signal or other sources of technical noise.
In this example, using a univariate ANOVA fi lter (imposing
bias from the experimental design) prevents these changes from
infl uencing the PCA when a dye swap is used in the experimental
design (see Fig.
3c
, Note 13). As was found in Example 2, applying
the ANOVA fi lter now enables PCA to organize the samples based
on the biological signal of different culture conditions. The weak
biological signal represented in these 68 features is qualitatively
evident by the loose nature of the sample clustering but nevertheless
3.6. Example 3: Very
Low Signal, Low Noise
Search WWH ::
Custom Search