Biomedical Engineering Reference
In-Depth Information
a biomaterial (Hanagata
et al.,
2010). All of these will require a different
combination of processes for data analysis from the experimental design to
the meta-analysis but the overall concepts will be the same: replication, nor-
malization, analysis method, multiple testing correction and interpretation.
8.2.1 Replicates
At the heart of any experimental design is the need for replicates. As we are
dealing with a large number of assays on a chip, and with the random varia-
tion due to experimental conditions, there is going to be a high rate of false
positive results. One way to help control this is to replicate the experiment
multiple times so the analysis can better estimate the error. If you perform
assays on samples which are very diverse (e.g. humans), you will need even
more replications in order to estimate the error due to the differences in
patients versus difference in treatments. The intrinsic variability of biologi-
cal samples depends on their source: for example, classifi ed from least het-
erogeneous to most heterogeneous are cell lines, mice (from the same litter)
and human. As a rough estimate, the desired sample size for each would be a
minimum of three for cell lines (though six is preferred) to a minimum of six
for mice (12 is preferred) to hundreds for humans. To get a closer estimate
on the number of replicates needed, a power analysis, which depends on the
platform, should be performed. If you are using an experimental platform,
you may need to set up a preliminary experiment to be able to estimate the
variability and power. The power is going to depend on the amplitude of
change you are looking for (the smaller the amplitude, the greater the num-
ber of replicates you will need), the sample size (the greater the sample size,
the more power), the number of false positive you are willing to accept, and
the standard deviation for each test on the array. For more information on
power analysis, and online tools to estimate it, please refer to the following
references: M. D. Anderson Cancer Center (2006), Gadbury
et al.
(2004),
Page
et al.
(2006) and Seo
et al.
(2006).
8.2.2 Experimental design and variability
Once you have the number of replicates required, you need to examine the
probable sources of variability in the experiment: experimenters, days, cul-
ture days, treatment days, etc., setting replicates so they are spread out over
the variables. For example, do not create all the replicates and experiments
of treatment A on day 1 and B on day 2. Instead, you should perform half of
A and B on day 1 and the other half on day 2. This is called blocking; with
statistical analysis, you will be able to account for the day-to-day variability
in your model and obtain more accurate results.
