Some aspects of the design and analysis of gene expression

advertisement
Statistical Issues in the Design of Microarray
Experiment
Jean Yee Hwa Yang
University of California, San Francisco
http://www.biostat.ucsf.edu/jean/
Stat 246 Lecture 18, Mar 30, 2004
SAGE
** ***
Nylon membrane
Illumina
Bead Array
Different
Technologies
GeneChip Affymetrix
cDNA microarray
Agilent: Long oligo Ink Jet
CGH
From W. Gibbs, Scientific American, 2001
Life
Cycle
Biological question
Experimental design
Failed
Microarray experiment
Quality
Measurement
Image analysis
Preprocessing
Normalization
Pass
Analysis
Estimation
Testing
Clustering
Biological verification
and interpretation
Discrimination
Experimental design
Proper experimental design is needed to
ensure that questions of interest can be
answered and that this can be done
accurately, given experimental constraints,
such as cost of reagents and availability of
mRNA.
Definition of probe and target
cDNA “A”
Cy5 labeled
TARGET
PROBE
cDNA “B”
Cy3 labeled
Two main aspects of array design
Design of the array
Allocation of mRNA
samples to the slides
cDNA “A”
Cy5 labeled
cDNA “B”
Cy3 labeled
Arrayed Library
(96 or 384-well plates of
bacterial glycerol stocks)
Spot as microarray
on glass slides
Hybridization
Two main aspects of array design
Design of the array
Allocation of mRNA
samples to the slides
WT
MT
Hybridization
Some aspects of design
1. Layout of the array
Which sequence to print?
- Affymetrix – selecting the PM and MM’s.
- cDNA Library – Riken, NIA etc
- Selecting oligo probes:
-
Operon (commercial).
Aglient (commercial).
Illumina (commercial).
OligoPicker (Wang & Seed, Harvard)
ArrayOligoSelector (Bozdech etal, DeRisi Lab, UCSF)
OligoArray 1.0 (Rouillard, Herbert and Zuker)
Controls:
Normalization and quality checks.
Number:
Duplicate or replicate spots within a slide position.
Commonly asked questions
• Should we put duplicates on a slides.
[Discussion in Smyth et al (2003)]
• What should be the percentage of control spots?
• Where should the control spots be placed? [These
relates to preprocessing such as quality assessment and
normalization].
External
Controls
External
controls
From Van de Peppel et al, 2003
From Van de Peppel et al, 2003
References
• Microarray sample pool.
- Yang et al (2002). Normalization for cDNA microarray data: a
robust composite method addressing single and multiple slide
systematic variation. Nucleic Acids Research, Vol. 30, No. 4,
e15.
• Titrated external controls.
- J. van de Peppel, et al (2003). Monitoring global mRNA changes
with externally controlled microarray experiments. Embo Reports
4: 387-393.
• An example of doping controls for microarray.
- http://genome-www.stanford.edu/turnover/supplement.shtml
• Commercial
- Lucidea Universal ScoreCard by Amersham.
http://www1.amershambiosciences.com/
Some aspects of design
2. Allocation of samples to the slides
A Types of Samples
-
Replication – technical, biological.
Pooled vs individual samples.
Pooled vs amplification samples.
This relates to both
Affymetrix and
two color spotted array.
B Different design layout
-
Scientific aim of the experiment.
Robustness.
Extensibility.
Efficiency.
Taking physical limitations or cost into consideration:
-
the number of slides.
the amount of material.
Avoidance of bias
• Conditions of an experiment; mRNA extraction
and processing, the reagents, the operators, the
scanners and so on can leave a “global
signature” in the resulting expression data.
• Randomization.
• Local control is the general term used for
arranging experimental material.
Preparing mRNA samples:
Mouse model
Dissection of
tissue
RNA
Isolation
Amplification
Probe
labelling
Hybridization
Preparing mRNA samples:
Mouse model
Dissection of
tissue
RNA
Isolation
Amplification
Probe
labelling
Hybridization
Biological Replicates
Preparing mRNA samples:
Mouse model
Dissection of
tissue
RNA
Isolation
Amplification
Probe
labelling
Hybridization
Technical replicates
Preparing mRNA samples:
Mouse model
Dissection of
tissue
RNA
Isolation
Amplification
Probe
labelling
Hybridization
Technical replicates
Technical replication - amplification
Olfactory bulb experiment:
• 3 sets of Anterior vs Dorsal performed on different days
• #10 and #12 were from the same RNA isolation and
amplification
• #12 and #18 were from different dissections and amplifications
• All 3 data sets were labeled separately before hybridization
Data provided by
Dave Lin (Cornell)
Pooling: looking at very small amount of tissues
Mouse model
Dissection of
tissue
RNA
Isolation
Pooling
Probe
labelling
Hybridization
Design 1
Pooled vs Individual
samples
Design 2
Taken from
Kendziorski etl al (2003)
Pooled vs Individual samples
• Pooling is seen as “biological averaging”.
• Trade off between
- Cost of performing a hybridization.
- Cost of the mRNA samples.
Cost or mRNA samples << Cost per hybridization
Pooling can assists reducing the number of
hybridization.
Pooled vs
amplified samples
amplification
amplification
Design A
amplification
amplification
pooling
Design B
pooling
Original samples
Amplified samples
Pooled vs Amplified samples
• In the cases where we do not have enough material from
one biological sample to perform one array (chip)
hybridizations. Pooling or Amplification are necessary.
• Amplification
- Introduces more noise.
- Non-linear amplification (??), different genes amplified at
different rate.
- Able to perform more hybridizations.
• Pooling
- Less replicates hybridizations.
References
• Pooling vs Non-Pooling
- Han, E.-S., Wu, Y., Bolstad, B., and Speed, T. P. (2003). A study
of the effects of pooling on gene expression estimates using high
density oligonucleotide array data. Department of Biological
Science, University of Tulsa, February 2003.
- Kendziorski, C.M., Y. Zhang, H. Lan, and A.D. Attie. (2003). The
efficiency of mRNA pooling in microarray experiments.
Biostatistics 4, 465-477. 7/2003
- Xuejun Peng, Constance L Wood, Eric M Blalock, Kuey Chu
Chen, Philip W Landfield, Arnold J Stromberg (2003). Statistical
implications of pooling RNA samples for microarray experiments.
BMC Bioinformatics 4:26. 6/2003
Some aspects of design
2. Allocation of samples to the slides
A Types of Samples
-
Replication – technical, biological.
Pooled vs individual samples.
Pooled vs amplification samples.
B Different design layout
-
Scientific aim of the experiment.
Robustness.
Extensibility.
Efficiency.
Taking physical limitation or cost into consideration:
-
the number of slides.
the amount of material.
Graphical representation
Vertices: mRNA samples;
Edges: hybridization;
Direction: dye assignment.
Cy3
sample
Cy5
sample
Graphical representation
• The structure of the graph determines which
effects can be estimated and the precision of the
estimates.
- Two mRNA samples can be compared only if there is
a path joining the corresponding two vertices.
- The precision of the estimated contrast then depends
on the number of paths joining the two vertices and is
inversely related to the length of the paths.
• Direct comparisons within slides yield more
precise estimates than indirect ones between
slides.
Common reference design
T1
T2
Tn-1
Tn
Ref
•
Experiment for which the common reference design is appropriate
Meaningful biological control (C) Identify genes that responded differently /
similarly across two or more treatments relative to control.
Large scale comparison. To discover tumor subtypes when you have many
different tumor samples.
•
Advantages:
Ease of interpretation.
Extensibility - extend current study or to compare the results from current
study to other array projects.
Experiment for which a number of designs are
suitable for use
Time Series
T1
T2
T3
T4
Ref
T1
T2
T3
T4
T1
T2
T3
T4
T1
T2
T3
T4
Experiment for which a number of designs are suitable
for use
4 samples
A
B
C
A.B
C
A
C
A
C
A
B
A.B
B
A.B
B
A.B
Comparing 2 classes of estimates
direct vs indirect estimates
The simplest design question:
Direct versus indirect comparisons
Two samples (A vs B)
e.g. KO vs. WT or mutant vs. WT
Indirect
A
Direct
A
B
average (log (A/B))
2 /2
R
B
log (A / R) – log (B / R )
22
These calculations assume independence of replicates: the reality is not so simple.
In practice: data
Wt
Wt
Mutant
Mutant
Method A
Method B
The analyses can be done in two ways:
A. Exclude the 2 wt vs. wt arrays (one-sample)
B. Include the 2 wt vs. wt arrays (two-samples)
Data provided by
Grant Hartzog (UCSC)
Experimental results
• 5 sets of experiments with
similar structure.
• Compare (Y axis)
• Theoretical ratio of (A / B) is
1.6
• Experimental observation is
1.1 to 1.4.
SE
A) SE for aveMmt
B) SE for aveMmt – aveMwt
Sample size
Data provided by Matt Callow
Caveat
• The advantage of direct over indirect comparisons was
first pointed out by Churchill & Kerr, and in general, we
agree with the conclusion. However, you can see in the
last MA-plot that the difference is not a factor of 2, as
theory predicts.
• Why? Possibly because mRNA from the same
extractions - and pools of controls or reference material
are the norm - give correlated expression levels. In other
words, the assumption of independence between
log(T/Ref) and log(C/Ref) is not valid.
Preparing mRNA samples:
Mouse model
Dissection of
tissue
RNA
Isolation
Amplification
Probe
labelling
Hybridization
Extreme technical replication - labeling
• 3 sets of self – self hybridization: (cerebellum vs
cerebellum)
• Data 1 and Data 2 were labeled together and hybridized
on two slides separately.
• Data 3 was labeled separately.
• Comparing log-ratios between the 3 experiments
Data 1
Data 1
Data provided by
Elva Diaz (UC Davis)
Pairs plot of log-intensity
Looking only at constantly
expressed genes only
A
A’
B
B’
a= log2A and b = log2B
Technical
replicates
Data provided by Grant Hartzog and
Todd Burcin from UCSC
A
B
a = log2A and b = log2B
a’ = log2A’ and b’ = log2B’
t2 = var(a) ; variance of log signal
g1 = cov(a, b); covariance between measurements on
samples on the same slide.
g2 = cov(a, a’); covariance between measurements on
technical replicates from different slides.
g3 = cov(a, b’); covariance between measurements on
samples which are not technical replicates and not on
the same slide.
A
B
B
A
A
C
D
C
Implication for design
2 = var(a-b) = 2(t2 – g1)
c0 = cov(a – b, c – d) = 0
c1 = cov(a – b, a’ – b) = g2 – g3
c2 = cov(a – b, a’ – b’) = 2(g2 – g3) = 2c1
B
Direct vs Indirect - revisited
Two samples (A vs B)
e.g. KO vs. WT or mutant vs. WT
Indirect
A
Direct
A
B
R
B
y = (a – b) + (a’ – b’)
y = (a – r) - (b – r’)
Var(y/2) = 2 /2 + c1
Var(y) = 22 - 2c1
2 = 2c1
c1 = 0
efficiency ratio (Indirect / Direct) = 1
efficiency ratio (Indirect / Direct) = 4
Summary
• Create highly correlated reference
samples to overcome inefficiency in
common reference design.
• Not advocating the use of technical
replicates in place of biological replicates
for samples of interest.
Linear Models
Gene expression data
Gene expression data on G genes for n hybridization.
Gene by array data matrix
mRNA samples
sample1 sample2 sample3 sample4 sample5 …
Genes
1
2
3
4
0.46
-0.10
0.15
-0.45
0.30
0.49
0.74
-1.03
0.80
0.24
0.04
-0.79
1.51
0.06
0.10
-0.56
0.90
0.46
0.20
-0.32
...
...
...
...
5
-0.06
1.06
1.35
1.09
-1.09
...
Gene expression level of gene i in mRNA sample j
M=
Log( Red intensity / Green intensity)
Function (PM, MM) using MAS or dchip or RMA
Combining data across arrays
How can we design experiments and combine data
across slides to provide accurate estimates of the
effects of interest?
Linear models and extensions thereof can be used to
effectively combine data across arrays for complex
experimental designs.
C
A
B
AB
Experimental design
Regression analysis
Comparing K treatments
(i) Indirect design
A
B
R
(ii) Direct design
C
C
A
B
Question: Which design gives the most precise
estimates of the contrasts A-B, A-C, and B-C?
Common reference design
A
B
C
Log ratios: y
Model: E(y1) = A – R
E(y2) = B – R
E(y3) = C – R
R
Question 1: Estimate the log-ratios between sample A
and B (the parameter is: a-b)
Calculate y1-y2 for every gene
Question 2: Which genes are differentially expressed
in at least one of the samples relative to R.
Calculate the F-statistic for every gene
Linear model analysis
Gene by Gene analysis
C
Log ratios: y
Parameters: b = ( a-p, l-p ), where
a = log2A, p = log2P and l = log2L
Model: E(y1) = c – b
E(y2) = b – a
E(y3) = a – c
 y1    1 0 
  
 b  c 

E  y2    0 1 
 y   1  1 b  a 
 3 


bˆ  X ' X

1
X y
A
B
 y1 
 
  Cov y2    2 I 3
y 
 3

Cov( bˆ )   2 X ' X

1
I (a) Common
reference
A
P
I (b) Common
reference
L
A
P
2
2
2
w
w
Number of
Slides
N=3
N=6
Ave. variance
2
II Direct
comparison
L
A
2
Units of material
Ave. variance
For k = 3, efficiency ratio (Design I(a) / Design II) = 3
In general, efficiency ratio = 2k / (k-1)
L
P
N=3
0.67
Experimental design
• Efficiency can be measured in terms of
different quantities
- number of slides or hybridizations;
- units of biological material, e.g. amount of
mRNA for one channel.
I (a) Common
reference
A
P
I (b) Common
reference
L
A
P
2
2
2
w
w
Number of Slides
N=3
N=6
Ave. variance
2
Units of material
A=B=C=1
Ave. variance
II Direct
comparison
L
A
2
L
P
N=3
0.67
A=B=C=2
A=B=C=2
1
0.67
For k = 3, efficiency ratio (Design I(b) / Design II) = 1.5
In general, efficiency ratio = k / (k-1)
2 x 2 factorial experiment
two factors, two levels each
Study the joint effect of two treatments (e.g. drugs),
A and B, say, on the gene expression response of
tumor cells.
There are four possible treatment combinations
AB: both treatments are administered;
A : only treatment A is administered;
B : only treatment B is administered;
C : cells are untreated.
C
A
B
AB
n=12
2 x 2 factorial experiment
two factors, two levels each
Example 2
Our interest in comparing two strains of mice (mutant and
wild-type) at two different times, postnatal and adult.
WT.P1

2
WT.P21
 +a
4
1
MT.P1
MT.P21
+b
 +a+b+a.b
4 possible samples:
- WT.P1: WT at postnatal
- WT.P21: WT at adult (effect of time only)
- MT.P1: at postnatal (effect of the mutation only)
- MT.P21: MT at adult (effect of both time and the mutation).
3
2 x 2 factorial experiment
two factors, two levels each
Another possible design:
Use C as a common reference
y1 = log (A / C) = a + error
y2 = log (B / C) = b + error
A
y1
B
y2
y3
y3 = log (AB / C) = a + b + ab + error
C
Estimate (ab) with y3 - y2 - y1.
A.B
2 x 2 factorial experiment
two factors, two levels each
Study the joint effect of two treatments (e.g. drugs),
A and B, say, on the gene expression response of
tumor cells.
There are four possible treatment combinations
AB: both treatments are administered;
A : only treatment A is administered;
B : only treatment B is administered;
C : cells are untreated.
C
A
B
AB
n=12
2 x 2 factorial experiment
For each gene,
consider a linear
model for the joint
effect of treatments A
and B on the
expression response.
 AB   + a + b + g
A   + a
B   + b
C  
: baseline effect;
a: treatment A main effect;
 B main effect;
b: treatment
g: interaction between treatments A and B.
2 x 2 factorial experiment
C
C
B
B

A
A
AB

AB
C  

A   + a
 B  + b
 AB   + a + b + g
Log-ratio M for hybridization
A
AB
estimates
AB  A  b + g
Log-ratio M for hybridization
A
B
estimates
B  A  b  a
Regression analysis
• For parameters q  a, b, g, define a
design matrix X so that E(M)=Xq.
• For each gene, estimate model by robust regression,
least squares or generalized least squares of q.
 M 11   0

 
M

 0
12 
M   1
21

 
M

 1
22 
M  
31 

 1
 M 32    1
E

M
41 

 0
 M 42   0

 
 M 51   1
 M   1
52

 
M

 1
61 
M   1
62 


1
1
1
1
0
0
1
1
0
0
1
1
0

0
1

 1
0
 a
0 
 b
1
g
 1

1
 1

0
0






1
ˆ
q  X ' X  X ' M
Experimental design
• In addition to experimental constraints, design
decisions should be guided by the knowledge of
which effects are of greater interest to the
investigator.
E.g. which main effects, which interactions.
• The experimenter should thus decide on the
comparisons for which he wants the most
precision and these should be made within
slides to the extent possible.
2 x 2 factorial
Indirect
A balance of direct and indirect
I)
II)
A
B
C
III)
A.B C
A
IV)
C
A
A.B B
B
# Slides
C
A
A.B B
A.B
N=6
Main
effect A
0.5
0.67
0.5
NA
Main
effect B
0.5
0.43
0.5
0.3
Interactio
n A.B
1.5
0.67
1
0.67
Table entry: variance
Ref: Glonek & Solomon (2002)
2 x 2 factorial
Indirect
A balance of direct and indirect
I)
II)
A
B
C
III)
A.B C
A
IV)
C
A
A.B B
B
# Slides
C
A
A.B B
A.B
N=6
Main
effect A
0.5
0.67
0.5
NA
Main
effect B
0.5
0.43
0.5
0.3
Interactio
n A.B
1.5
0.67
1
0.67
Table entry: variance
Ref: Glonek & Solomon (2002)
References
-
T. P. Speed and Y. H Yang (2002). Direct versus indirect designs for cDNA
microarray experiments. Sankhya : The Indian Journal of Statistics, Vol. 64,
Series A, Pt. 3, pp 706-720
-
Y.H. Yang and T. P. Speed (2003). Design and analysis of comparative
microarray Experiments In T. P Speed (ed) Statistical analysis of gene
expression microarray data, Chapman & Hall.
-
R. Simon, M. D. Radmacher and K. Dobbin (2002). Design of studies using
DNA microarrays. Genetic Epidemiology 23:21-36.
-
F. Bretz, J. Landgrebe and E. Brunner (2003). Efficient design and analysis of
two color factorial microarray experiments. Biostaistics.
-
G. Churchill (2003). Fundamentals of experimental design for cDNA
microarrays. Nature genetics review 32:490-495.
-
G. Smyth, J. Michaud and H. Scott (2003) Use of within-array replicate spots
for assessing differential experssion in microarray experiments. Technical
Report In WEHI.
-
Glonek, G. F. V., and Solomon, P. J. (2002). Factorial and time course designs
for cDNA microarray experiments. Technical Report, Department of Applied
Mathematics, University of Adelaide. 10/2002
Acknowledgments
Statistics
Biology
Gordon Smyth (WEHI)
Natalie Thorne (WEHI)
Terry Speed (Berkeley)
Sandrine Dudoit (Berkeley)
David Erle (UCSF)
Grant Hartzog(UCSC)
Todd Burcin (UCSC)
Dave Lin (Cornell)
John Ngai (Berkeley)
Elva Diaz (UC Davis)
Matt Callow (LBL)
Yuanyuan Xiao (UCSF)
Mark Segal (UCSF)
Ru Fang Yeh (UCSF)
Single factor experiment – time course
T1
T2
T3
T4
T5
T6
T7
Ref
Possible designs:
1) All sample vs common pooled reference
2) All sample vs time 0
3) Direct hybridization between times.
Pooled reference
Compare to T1
t vs t+1
t vs t+2
t vs t+3
Design choices in time series
N=3
t vs t+1
A) T1 as common reference
T1
T2
T3
N=4
T2
T3
C) Common reference
T1
T2
T3
T1T2
T2T3
T3T4
T1T3
T2T4
T1T4
Ave
1
2
2
1
2
1
1.5
1
1
1
2
2
3
1.67
2
2
2
2
2
2
2
.67
.67
1.67
.67
1.67
1
1.06
.75
.75
.75
1
1
.75
.83
1
.75
1
.75
.75
.75
.83
T4
B) Direct Hybridization
T1
t vs t+2
T4
T4
Ref
D) T1 as common ref + more
T1
T2
T3
T4
E) Direct hybridization choice 1
T1
T2
T3
T4
F) Direct Hybridization choice 2
T1
T2
T3
T4
Linear model analysis
A
Log ratios: y
Parameters: b = ( a-p, l-p ), where
a = log2A, p = log2P and l = log2L
Model:
L
P
 y1    1 0 
  
 a  p 

E  y2    0 1 
 y   1  1 l  p 
 3 

2
c 1 c1 
 y1   
1


 

2
2
  Cov y2    c1 
c1     
2
y  c

c

1
 3  1



bˆ  X '  1 X

1
1
X ' y

1




1 

Cov( bˆ )  X '  1 X

1
In a more general framework
A
P
L
w
A
P
2
A
2
L
2
w
A
g2 g3
c1
 2 

2(t 2  g 1 )

L
P
In a more general frame work

Design I, II, III, IV


Download