Introduction to Experimental Design - University of...
Transcript of Introduction to Experimental Design - University of...
1
Introduction to Experimental DesignIntroduction to Experimental DesignApplication to Gene Expression MicroarraysApplication to Gene Expression Microarrays
Kathleen KerrKathleen KerrBiostatistics 577 / Statistics 577Biostatistics 577 / Statistics 577
University of WashingtonUniversity of WashingtonSummer 2007Summer 2007
• Principles of statistical experimental
design.
• Design and analysis of agricultural
experiments.
• Introduction to gene expression
microarrays.
• Experimental design for gene expression
microarrays.
2
R.A. Fisher1890-1962
Statistician, experimentalist,geneticist
“To call in the statistician after the experiment is done may be no more than asking him to perform a postmortem examination: he may be able to say what the experiment died of.''
R.A. Fisher, Indian Statistical Congress, Sankhya, ca 1938
3
“If I had to replicate my experiments, I could only do half as much.''
-Well-known Stanford molecular biologist, ca. 2000
Introduction to Experimental Design
4
Agricultural experiment to compare the yields of varieties of a crop.
Multiple blocks of land to use in the study.
Blocks of land vary in fertility, sunlight, rainfall, etc.
Block A Block B
Variety 1 Variety 2
Experiments in BlocksCase 1
5
Block A Block B
Variety 1 Variety 2
Block effects and variety effects are confounded.
Experiments in BlocksCase 1
1 212 1221
Block A Block B
Case 2
6
1 212 1221
Block effects and variety effects are orthogonal(unconfounded) because blocks and varieties are balanced.
Block A Block B
Case 2
Block A Block B
1 12 122
Case 3
7
Block A Block B
“In between” case:Block effects and variety effects are partially confounded.
1 12 122
Case 3
The data contain information about the relativevariety yields because varieties are grown on the same block of land.
There is a duality that can be overlooked:
There is information about the blocks of land because they have varieties in common.
8
One can simultaneously estimate the relativeyields of the varieties and the relative effectsof the blocks of land.
The tool to do this is a simple linear model.
The precision of estimates of variety differencesdepends on the experimental design.
Let yik be the yield from variety k on block i
Modelyik = μ + Bi + Vk + εik
μ is the mean yieldBi is the effect of block iVk is the effect of variety kεik is error
9
Block 1 Block 2 Block 3
ExampleGoal: Compare three varieties. Resources: Three blocks of size two. How should you put the varieties onto the blocks?
1 2 3 2 1 3
Block 1 Block 2 Block 3
R 1 R 2 R 3
Block 1 Block 2 Block 3
Design 1: “Balanced design”
Design 2: “Reference Design”
10
1 2 3 2 1 3
Design 1: “Balanced design”
y=X β+ ε
-1-1-1-1101-1-1110101
-1-11011001101011
µ B1 B2 V1 V2
X=
Error df=1
yik = μ + Bi + Vk + εik
Identifiability constraints: B1+B2+B3=0, V1+V2+V3=0
1 2 3 2 1 3
Design 1: “Balanced design”
Var( β) = σ2(X’X)-1
Var(V1-V2)=Var(V1-V3)=Var(V2-V3)= 4σ2/3
11
R 1 R 2 R 3
Design 2: “Reference Design”
B1+B2+B3=0 V1+V2+V3+3VR=0
100-1-11-1/3-1/3-1/3-1-11
010101-1/3-1/3-1/3101
001011-1/3-1/3-1/3011
X=
µ B1 B2 V1 V2 V3
Error df=0
R 1 R 2 R 3
Design 2: “Reference Design”
Var( β) = σ2(X’X)-1
Var(V1-V2)=Var(V1-V3)=Var(V2-V3)= 4σ2
4σ2 compared to 4σ2/3 for the balanced design
12
1 2 3 2 1 3
R 1 R 2 R 3
Design 1: “Balanced design”
Design 2: “Reference Design”
What is the balanced design more efficient?
Intuition: In the reference design, varieties 1,2,3 are compared only via the reference. In the balanced designed, each variety is compared directly as well as indirectly.
Introduction to Gene Expression Microarrays
13
The nucleus of each cell of our body has chromosomes. These contains our genes.
Genes are made up of DNA ( deoxyribonucleic acid): linear sequences of four letters or bases: A,G,C, and T.
DNA has the famous double-helix structure discovered by Watson and Crick. The double-helix is formed by complementary base-pairing. That is, A and T always pairtogether, and G and C always pair together.
...ATGATGATCCAGC...
...TACTACTAGGTCG...
Illustration:Post-genome Informatics
Minoru Kanehisa,Oxford University Press, 2000
14
The four base alphabet of DNA is really a recipe to make things.
More specifically, via the genetic code DNA contains the instructions for making proteins.
1. DNA is transcribed into messenger RNA.2. mRNA is translated into protein.
15
Post-genome InformaticsMinoru Kanehisa,
Oxford University Press, 2000
All of the cells in our body have the same DNA.
Why are they different from each other?
Each type of cell expresses a different set of genes. Two cells that express the same gene may express them at different levels.
16
Some Goals of Gene Expression Studies
• Discover gene function – “guilt by association” Genes with similar patterns of expression may be in similar pathways
• Distinguish cancer cells from normal cells on the molecular level (for further understanding of cancer, or to identify diagnostic or prognostic markers, or to identify drug targets)
• Characterize complex diseases at the molecular level
Gene Expression Microarray
Idea: instead of directly measuring the amount ofprotein corresponding to each gene, measure the amount of message to make each protein, the mRNA.
17
Each spot on the arraycontains single-strandedDNA representing oneparticular gene.
cDNA Microarray
mRNA
cDNA
DNA microarray Re-created from Brown and Botstein, Nature Genetics Supplement, 1999
18
Quantitative problems in microarrays
One of the goals of a microarray study is to estimate the relative expression of the genes in the different RNAs.
19
• There is variation in the amount of DNA from spot to spot on the arrays.
• The relative “red” and “green” signal from a spot is interesting because the sample that contained more transcript should produce higher signal.
• By taking ratios (log ratios) of “red” and “green” for a spot we are taking “within spot” estimates of expression differences (just like block design!)
A Arrays i = 1,...,ID Dyes j = 1,2V Varieties k = 1,...,K G Genes g = 1,...,N
Experimental Factors in a Microarray Study
20
Example: “Dye Swap” Experiment
Two “varieties.”
Experimental Design: “Dye Swap”
ControlTreatmentRedTreatmentControlGreen
21Array
R
G R
GControlTreatment
A Arrays i = 1,...,ID Dyes j = 1,2V Varieties k = 1,...,K G Genes g = 1,...,N
Microarray Main Effectsand interpretation
21
Effects Involving Gene
• main effects G
• array-by-gene interaction (spots) AG
• dye-by-gene interaction DG
• variety-by-gene interaction VG
VGkg - VGk’g = change in gene expression
Dye-swap: Pretend we have just two spots (two genes).
Gene 1
Gene 2
red signalgreen signal
red signalgreen signal
Array 1
Gene 1
Gene 2
red signalgreen signal
red signalgreen signal
Array 2
22
-11-1-11-1-11-1111-1-111-11-11
-1-1-111-111111-1-11111111GVDAµ
Dye-swap: Pretend we have just two spots (two genes).Then the design matrix for the main effects looks like:
Note that every pair of columns is orthogonal.
Array 1
Gene 1
Gene 1
Gene 2
Gene 2
-11-1-11-1-11-1111-1-111-11-11
-1-1-111-111111-1-11111111GVDAµ
Dye-swap: Pretend we have just two spots (two genes).Then the design matrix for the main effects looks like:
Every column has four 1’s and four -1’s.
Array 1
Gene 1
Gene 1
Gene 2
Gene 2
23
-111-11-1-111-11-1-11-111-1-111-1-11
-11-11-11-1111-1-1-1-111
-1-1-1-11111-1-111-1-11111111111
VGDGAGGVDAµ
Add some two-way interactions:
Every column remains orthogonal to every other column.
1-11
-1-11
-11
AD
-111-11-1-111-11-1-11-111-1-111-1-11
-11-11-11-1111-1-1-1-111
-1-1-1-11111-1-111-1-11111111111
VGDGAGGVDAµ
What if we tried to add other interactions?
24
1-11
-1-11
-11
AD
-111-11-1-111-11-1-11-111-1-111-1-11
-11-11-11-1111-1-1-1-111
-1-1-1-11111-1-111-1-11111111111
VGDGAGGVDAµ
AD looks exactly like V – design matrix is no longer full rank
11
-1-1-1-111
DVG
-111-11-1-111-11-1-11-111-1-111-1-11
-11-11-11-1111-1-1-1-111
-1-1-1-11111-1-111-1-11111111111
VGDGAGGVDAµ
25
11
-1-1-1-111
DVG
-111-11-1-111-11-1-11-111-1-111-1-11
-11-11-11-1111-1-1-1-111
-1-1-1-11111-1-111-1-11111111111
VGDGAGGVDAµ
Confounding Structureof a Dye-Swap Experiment
Each experimental effect is confounded with one other effect, its alias. Non-aliased effects are orthogonal.
AVG~DGDVG~AGADG~VG
ADVG~GAD~VAV~DDV~A
ADV~µ
26
• A and D are orthogonal
• D and V can be confounded, partially confounded, or orthogonal ─This depends on the chosen experimental design
• A and V will be partially confounded if there are more than 2 varieties
Structure of the Design Space:Arrays, Dyes, and Varieties
Structure of the Design Space:Gene-specific Effects
• G orthogonal to A, D, and V
• AG orthogonal to DG
• AG and VG are partially confounded if there are more than 2 varieties
• DG and VG can be confounded, partially confounded, or orthogonal ─This depends on the chosen experimental design
27
Dye and Variety
D and V … and so also DG and VG … will be orthogonal if the design is “even.” That is, each RNA is labeled with both dyes and the design uses each labeled sample equally often.
This is important because the variety-by-gene interactions (VG) are the effects we care about.
“Dye Swap” DataData are from a drug-treated variety and a control.
ControlTreatmentRedTreatmentControlGreen
21Array
R
G R
GControlTreatment
Mouse liver RNA
28
•Only 78 genes spotted on each array.
•Each gene spotted 4 times per array.
•(Atypically small experiment)
Dye-Swap Data
Array 1
29
Array 2
Analysis of Variance
3.3513.35Dye45.69145.69Variety11.9277917.53Gene0.46289145.19Spot
13431219.35Adjusted Total0.024182019.72Residual
0.287721.30Dye*Gene0.867766.46Variety*Gene
0.1110.11ArrayMSdfSSSource
SS=Sum of Squares, df=degrees of freedom, MS=Mean Square=SS/df
30
We assume that there is independent, random error εijkg with mean 0.
yijkgr = µ + Ai + Dj + Vk + Gg + (VG)kg + (AG)igr + (DG)kg + εijkgr
Statistical model underlying the analysis of variance:Let yijkgr be the signal* from rth spot for Gene g fromArray i, Dye j, and Variety k
*On log scale or similar
DG1g─DG2g forg=1,…,78
31
Q: What would we have concluded if we had run 1 array instead of 2 arrays?
The “design” of an experiment has many different facets
1. The samples selected for comparison
2. The specification of the units to which the samples will be applied
3. The way the samples are allocated to the experimental units
4. The specifications of the measurements to be made
32
The “design” of an experiment is
1. The samples selected for comparison
2. The specification of the units to which the samples will be applied
3. The way the samples are allocated to the experimental units
4. The specifications of the measurements to be made
Representation of Microarray Designs
Mouse 1RNA
Mouse 3RNA
Mouse 2RNA
Mouse 4RNA
RNA samples are representedas rectangles or circles
Microarrays are representedby arrows, where one end is the “red” channel and the opposite end is the “green” channel
33
Example: “Dye Swap”
ControlTreatmentRedTreatmentControlGreen
21Array
R
G R
GControlTreatment
R 1 R 2 R 3
Block 1 Block 2 Block 3
321
R represents:
34
“Reference” Design
Reference sample
Samples of interest
1 2 3 2 1 3
Block 1 Block 2 Block 3
Becomes…
1
23
35
“Loop” Design
“Reference” Design
“Loop” Design
36
Reference Design
• Uses v arrays to study v samples
• V and D confounded;VG and DG confounded
• May lack error degrees of freedom
• Inefficient* estimation
Loop Design
• Uses v arrays to study v samples
• V and D orthogonal;VG and DG orthogonal
• Increased error degrees of freedom
• Efficient* estimation
*Efficiency measured according to A-optimality. The loop design is more efficient for 10 or fewer samples.
A-Optimality
For a given number of varieties and a given number of arrays, the A-optimal design minimizes the average variance over allcomparisons VGkg-VGk’g
(Just one possible criterion, may or may notbe appropriate.)
37
Relative Efficiency
Without spoteffects
With spoteffects
K
Optimal Designs
For v varieties, these designs are A-optimal among all even designs using v+2 arrays.
“Design A” for HW1
38
Optimal Designs
For v varieties, these designs are A-optimal among all designsusing 2*v arrays.
Example: Study of Aging. Mutant mice live 30% longer than wild-type. The investigators plan to take RNA from young and old mice of each genetic strain. Thus there are four varieties but in a 2-by-2 factorial structure. The most interesting comparison is between the two genetic strains but this is not the only interesting comparison.
Old
Young
MutantWild-type
39
DCOld
BAYoung
MutantWild-type
I.
Original Design
Not possible
Not possible
½1
AGE acrossGEN
½(A+B)-½(C+D)
AGE within GEN
A-C or B-D
GEN across
AGE½(A+C)-½(B+D)
GEN withinAGE
A-B or C-D
Variance of Comparisons*
*The numbers are only relative, to be used to compare designs.
40
Old
YoungMutant
Wild-type
III.
Old
YoungMutant
Wild-type
II.
Old
YoungMutant
Wild-type
IV.
Alternative Designs
1/21.512IV.
12½1.5III.
11.511.5II.
Not possbile
Not possible½1I.
AGE across GEN
AGE within GEN
GEN across AGE
GEN withinAGEDesign
Variance Comparisons
41
Generally, we want efficient designs for an appropriate criterion but designing an experiment requires a lot of practical considerations:
• Robust properties• Extendibility• Simplicity of execution• Useful sub-designs
• Robust propertiesWhat happens to the efficiency if we lose an array? If we lose some spots on every array?
• ExtendibilityWhat if we decide to add more samples to the study later?
• Simplicity of executionWill we be able to keep track of the assays we need to do?
• Useful sub-designsWhat if we want to analyze the data on just a subset of samples?
42
The “design” of an experiment is
1. The samples selected for comparison
2. The specification of the units to which the samples will be applied
3. The way the samples are allocated to the experimental units
4. The specifications of the measurements to be made
“Replication”First level of replication: Genes spotted multiple times per array
Second level of replication: Multiple arrays to study the same samples
contrast with:
Replication in the classical sense: Random sampling of individuals from the populations of interest or randomly assigning individuals to treatment groups. Without this kind of replication, inference is limited to the particular RNAs in the study.
43
Treatments: A B
Replicates:
Dyes:
Arrays:
R G R G R G R GRNA1 RNA2 RNA3 RNA4
1 2 3 4
ReplicationFirst level of replication: Genes spotted multiple times per array
Second level of replication: Multiple arrays to study the same samples
replication repeated measures / subsampling
Replication in the classical sense: Random sampling of individuals from the populations of interest or randomly assigning individuals to treatment groups. Without this kind of replication, inference is limited to the particular RNAs in the study.
MeasurementError
BiologicalVariability
44
Replication: N mice from each of two groups Limited number
of Arrays, say 2N
Example: Comparing 2 “Treatments”
vs.
45
N mice in each group2N arrays
Var(Tmt – Ctl) = 4σ2/N
N mice in each group
2N “looped” assays
2N arrays
Var(Tmt – Ctl) = σ2/N
46
I have presented the design problems in terms of measuring the difference in means between the treated and control mice.
Is Tmt – Ctl really the quantity of interest?
No.
The quantity of interest is really the difference in population means μtmt – μctl. We are only interested in the sample means Tmt – Ctl for the purpose of making inference about μtmt – μctl.
“Reference” Design
“Alternating Loop” Design
47
Let σ2 be error variance as before, but now let τ2 be population variability.
n individuals are sampled from each population. The variance of the estimated difference from the reference design is:
(4σ2 + 2τ2 )/n.
For the alternating loop design the variance is:
(σ2 + 2τ2 )/n.
Conclusion: The loop strategy is always better, but if τ2 >> σ2, not by much.
Summary: Experimental Design• Design determines the information content of data,
the analyses that are possible, and the quality of the results.
– Are effects of interest confounded with other effects?
– How precisely can one estimate the effects of interest?
– Can one estimate error in order to make inference?
• Choosing an experimental design means balancing multiple, competing objectives: precision vs. cost, practical constraints, considerations of robustness, etc.
48
Summary: Experimental Design• Design is the most important part of
conducting a good experiment
– A botched analysis can be re-done, but a botched design can mean things are hopeless.