Biomedical Engineering Reference
In-Depth Information
average log ratios and get the correct average log ratio of 0 for no change.
The most common log transformations used are the log base 2 and the
log base 10.
8.3.5 Summary-level normalization
Summary-level normalization will help standardize results and reduce fur-
ther variability arising from differences in the quality and quantity of sam-
ples. For DNA microarrays, it is often assumed that the amount of DNA
or RNA for each sample/treatment does not vary with the treatment and
therefore the amounts are quantifi ed with the same amount applied to each
slide. Under this scenario, the same overall signal should be detected in all
the arrays. Some normalization approaches of this type are global as we
look at the signal on the slide overall, while others can be uniform (using
a scaling factor) or they can take into account any signal intensity bias (in
the case of a lowess normalization or quantile normalization (Berger
et al.,
2004; Durbin and Rocke, 2004)).
Other normalization methods are local, for example, print-tip lowess
where a lowess normalization per array pin is performed (Smyth and Speed,
2003; Yang
et al.,
2002) because the morphology of printed spots is not uni-
form. This method corrects a problem that originates from using many differ-
ent pins to print each array with each pin varying in the amount of material
deposited. It has the advantage of correcting both the intensity-dependent
bias and spatial effects that are associated with print-tip groups.
In other cases, it cannot be assumed that the same level of signal intensity
should be observed for all the treatments. For example, you may have treat-
ments that destroy or inhibit the molecule you are measuring, or you simply
do not know and cannot assume an overall homeostasis in the cell or that
every cell would contain the same amount. Internal controls can be used if
present or alternatively these can be spiked in. The latter solution can be
delicate, as it will not normalize for quality issues in your samples, which are
frequent when you are dealing with different sampling locations, personnel
and sample storage and transport, but will work very well for lab experi-
ments where the environmental conditions are well controlled.
Many arrays are distributed with a user manual containing predefi ned
normalization methods that work well with their specifi c platform. The user
should look at these fi rst as they have been tailored to the array and ensure
that the assumptions made by the normalization process are consistent with
the experiment.
Finally, normalization should be accompanied with before and after plots
of the data to verify that it performs as it is intended to. The most common
