John Quackenbush, Ph.D.

1
Presented by John Quackenbush, Ph.D. at the
June 10, 2003 meeting of the Pharmacology
Toxicology Subcommittee of the Advisory
Committee for Pharmaceutical Science
2
The Experimental Design

• The Experimental Design dictates a good deal of
  what you can do with the data
• Good normalization and processing reflects the
  experimental design
• The design also facilitates certain comparisons
  between samples and provides the statistical
  power you need for assigning confidence limits to
  individual measurements
• The design must reflect experimental reality
• The most straight-forward designs compare
  expression in two classes of samples to look for
  patterns that distinguish them.

3
Sample Pairing for Co-Hybridization Experiments
Direct Comparison with Dye Swap
A1
B1
A2
B2
A3
B3
A4
B4
A1
B1
A2
A3
B2
B3
A4
B4

• RNA sample is not limiting (e.g. plenty of
  sample)
• Flip dyes account for any gene-dye effects

Balanced Block Design
A1
B1
A2
B2
A4
B4
A3
B3

• RNA sample is limiting
• Balanced blocking accounts for any gene-dye
  effects

4
Multiple Sample Pairings
Reference Design (Indirect Comparison)

• More than two samples are compared
• (e.g. tumor classification, time course)
• Flip dyes are not necessary but can be done to
  increase precision
• Ratio values are inferred (indirect)
• Suited for cluster analysis need common
  reference

Loop Design
5
Loops and Reference Designs
23 Hybs
10 hybs
Standard flip-dye expt
S. Wang , K. Kerr, J. Quackenbush, G. Churchill
6
Loops and Reference Designs
Both approaches can give equivalent results
S. Wang , K. Kerr, J. Quackenbush, G. Churchill
7
Loop vs. Reference Designs

• Loop design
• Can provide direct measurements
• Give more data on each experimental sample with
  the same number of hybs
• Require more RNA per sample
• Can unwind with a bad sample or for a gene
  with bad data
• Reference design
• Easily extensible
• Simple interpretation of all results
• Requires less RNA per sample
• Less sensitive to bad RNA samples and bad
  array elements

8
One Possible Experimental Paradigm
Examining Genotype, Phenotype, and Environment
Parental - stressed
Derived - stressed
Parental - unstressed
Derived - unstressed
9
Basic Design Principles

• Biological replicas are more informative than
  correlated replicas (independent RNA, independent
  slides)
• More replicas are better higher statistical
  power
• For loops, hybridizations of individual samples
  should be balanced (as many Cy3 as Cy5
  labelings)
• Self-self hybs add data on reproducibility and
  can be used to produce error models
• At a minimum, should use dye swap replicates to
  compensate for any dye biases in labeling or
  detection

10
How Many Replicates?
(Simon et al., Genetic Epidemiology 23 21-36,
2002)
n 4(za/2 zb)2 / (d/1.4s)2
Where za/2 and zb are normal percentile values at
significance level a and false negative rate b
parameter d represents the minimum detectable
log2 ratio and s represents the SD of log ratio
values. For a 0.001 and b 0.05, then za/2
-3.29 and zb -1.65. Assume d 1.0 (2-fold
change) and s 0.25, Therefore n 12 samples
(6 query and 6 control).