Transcript Biometrics 59, 822-828 December 2003 M. Kathleen Kerr

M. Kathleen Kerr
“Design Considerations for Efficient
and Effective Microarray Studies”
Biometrics 59, 822-828; December 2003
Biostatistics Article
Oncology Journal Club
May 28, 2004
A couple introductory points


Different kinds of microarrays
Two main distinctions




One-color (e.g. Affymetrix, long oligo)
Two-color (e.g. spotted cDNA)
Some of the statistical tools are the same and
some are different
Using two color arrays is slightly more
complicated in terms of design
Statistics and Microarrays


Statistical Principles certainly apply to microarray
analyses
We should be considering some of the same basic
tenets when performing microarray studies




Randomization
Sample size/Replication issues
Experimental design
Good design is critical to making efficient and valid
inferences.
Randomization



Might not sound applicable
But…
 If you have a ‘treatment’ you are giving, samples should be
randomly assigned to treatment groups
 Randomize order in which samples are processed
 Randomize order in which hybridizations are performed
 Randomize the order in which arrays are chosen from array
batch.
Example: Dosing study
 Looking for genetic changes in cells as a function of dose
 Perform all dose=0 experiments first, then dose=1, then dose=2,
etc….
 But, as you proceed, you learn more, get better at processing
samples, hybridizations, using scanner….
 Your results be associated with dose even if dose has no affect
on genetic changes: CONFOUNDING!
Sample Size and Replication

Three types of ‘replication’ in microarrays
A. Spotting genes multiple times on same array
B. Hybridizing multiple arrays to the same RNA
samples
C. Using multiple individuals of a certain type



A and B are considered ‘technical’ replicates
C describes ‘random sampling’ from the
population
THESE ARE CRITICALLY DIFFERENT!
Sample Size and Replication

Technical replication:





DOES NOT address biological variability
DOES address measurement error of assay
Usually, interested how a condition affects
individuals in general
NOT usually interested in how a condition affects
any given individual
Example: AML


Do we want to make inferences about differences in gene
expression across AML subpopulations?
Or, do we want to make inferences about differences in
gene expression in two particular AML patients, each of
whom has a different type of AML?
Sample Size and Replication

Why/When would we be interested in
technical replication?



Medical diagnosis
Need to know how precise the measures are
Sensitivity and specificity of the assay depend on
that
Sample Size and Replication

Biological replicates



Tell us about the variability across samples of the same
type.
Biological variability is critical for
 finding differences in gene expressions across populations
 Classification procedures which try to use gene expression
patterns that differentiate individuals of different types
If you use just one sample or cell line to make
inferences about the population of interest


You are making a BIG assumption: “Population is relatively
homogeneous”
Cannot evaluate your assumption based on the data from
the study.
Sample Size and Replication

For a fixed sample size:




But, if it is expensive to sample new individuals



It is preferable to sample NEW individuals rather perform
technical replicates
Why? It is more efficient in terms of variance, power, etc.
You gain much less by replicates than new samples
Examples: samples are very rare, recruitment is difficult,
procedure for acquiring samples is risky or expensive
In this case, it might be worthwhile to perform some
technical replicates due to “cost-benefit” analysis
GENERAL RULE: TRUE REPLICATION BEATS
TECHNICAL REPLICATION FOR GAINS IN
PRECISION WHEN ESTIMATING PARAMETERS
Pooling of Samples



Often motivated by insufficient quantity of RNA, which is
reasonable.
Sometimes, proposed to ‘control’ for biological variability
 Bad idea!
 We need to understand, not eliminate biological variability
To understand the differences in mean expressions across two
populations (e.g. Normal karyotype and t(15:17)), we need to be
able to estimate the populations means
 We cannot do that if we have pooled RNA
 We can estimate mean difference in two groups based on pooled
samples
 But, we cannot make inferences about whether of not there is a
difference in mean expression.
Pooling of Samples

Pooling is ALWAYS bad if your goal is


Finding classification scheme
Discovering unknown subtypes

‘In between’ strategy for pooling when we
are interested in determining if average
expression is different in two phenotypes
(Kendziorski et al (2003)).

Pooling RNA for use as a ‘reference’ is OK
(more in a minute).
Experimental Layout

Discussion specific to two-color arrays




Complicated due to pairing of samples on arrays
One-color array design considerations usually more
straightforward
Critical determinant of design efficiency.
Three main types of designs in two-color arrays:



Reference
Loop
Dye swap
Reference Design






Each arrow represents an array
Lets say that origin of arrow is
green and head of arrow is red
Each sample of interest is
paired with the same
“reference” sample
AML example: reference was
11 pooled cell lines
Here, each sample is labeled
with red (Cy5) and reference is
labeled with green (Cy3)
Each sample is only hybridized
to ONE array (each reference)
Type 1
Type 2
Reference sample
Loop Design






Each sample is paired with a
sample of the other type (no
reference!)
Each sample is hybridized to TWO
arrays and is both red and green
Can compare any two arrays by
comparing arrays between them in
loop.
Relative efficiency is 4 to 1
comparing loop to reference
Downside: what if just ONE array
goes bad? Loop is not a loop
anymore!
Good design for small number of
samples: uses information very
effectively
Type 1
Type 2
Dye Swap Design







Each sample is paired with the
same sample of the other type
TWICE
Each sample is hybridized to TWO
arrays
Dyes are swapped
Relative efficiency is 4 to 1
comparing loop to reference
More robust than loop
Less complicated than loop
Direct comparisons are not as
easy because samples are not
linked through other samples as
in other two designs
Type 1
Type 2
Why reference so often?

As population variance increase, loop and dye
swaps have less advantage.





Sample comparisons must go ‘through’ loop
Direct comparisons not easy in dye swap if samples are
not on same chip.
If you have large number of samples, loop is risky
due to ‘bad chips’
Logically, however, by using reference on every
chip, we are ‘wasting’ a resource.
But, less efficiency advantage in complex designs
as number of RNAs increases
Robustness

Two robust alternatives: require 2x as many arrays
“Double reference”
“Double Loop”
Practical Considerations

Simplicity


Extendability




Large study with many technicians
Open-ended
Can add additional samples at a later time depending on
what early results suggest
Reference and “symmetric” reference designs
Useful subdesigns


“subgroup analyses”
Example: all AMLs vs. normal karyotype