Transcript Exploratory analysis of Affymetrix Data

Exploratory Data Analysis of High
Density Oligonucleotide Array
Rafael A. Irizarry, Bridget Hobbs,
Terry Speed
http://biosun01.biostat.jhsph.edu/~ririzarr/Raffy
Outline
•
•
•
•
•
Review of technology
Form of Data
Description of Data
Normalization
Future/current work:
Defining expression
Probe Arrays
GeneChip Probe Array
Hybridized Probe Cell
Single stranded,
labeled RNA target
*
*
*
*
*
Oligonucleotide probe
24µm
1.28cm
Millions of copies of a specific
oligonucleotide probe
>200,000 different
complementary probes
Image of Hybridized Probe Array
Compliments of D. Gerhold
Image analysis
• About 100 pixels per
probe cell
• These intensities are
combined to form one
number representing
expression for the
probe cell oligo
• What about genes?
PM MM
Data and notation
PMijn , MMijn = Intensity for perfect/mismatch
probe cell j, in chip i, in gene n
i = 1,…, I (ranging from 1 to hundreds)
j=1,…, J (usually 16 or 20)
n = 1,…, N (between 8,000 and 12,000)
$64K Question
• How do we define expression?
or
• What is the one number summary of the 20
PMs and 20 MMs that best quantifies
expression?
• How about differential expression?
Current default
• GeneChip® software uses Avg.diff
1
Avg.diff 

 ( PM
j
j
 MM j )
with A a set of “suitable” pairs chosen by software.
• Log ratio version is also used.
• For differential expression Avg.diffs are compared
between chips.
What is the evidence?
Lockhart et. al. Nature Biotechnology 14 (1996)
• Chips used in Lockhart et. al. contained
around 1000 probes per gene
• Current chips contain
20 probes per gene
• These are different situations
• We haven’t seen a plot like the previous
one, for current chips
Possible problems
What if
• a small number of the probe pairs hybridize much
better than the rest?
• removing the middle base does not make a
difference for some probes?
• some MM are PM for some other gene?
• there is need for normalization?
We explore these possibilities using data from 3
experiments
Experiment 1
• 8 Rats, under 4 experimental conditions
–
–
–
–
Control NV21
Ventilation V21
Oxygen NV100
Oxygen and Ventilation V100
• 2 rats in each condition
• RNA is pooled and divided to form 2
technical replicates for each condition
Notice
• Experimental condition is confounded with
couples: we can’t distinguish between
biological variability and variability due to
experimental condition
• NV21, V21 and NV100,V100 processed in
different scanners/fluidic stations: Oxygen
effect confounded with scanner/fluidic
station effect
Experiment 2
• 6 Rats, under 3 experimental conditions
– Control
– ENOS
– NNOS
• 2 rats in each condition
• RNA is pooled and divided to form 2
technical replicates
Notice
• One of the chips for NNOS did not “work”
• Biological variability confounded with
variability due to experimental condition
• About 1/5 of the probes on chips used
where defective.
Experiment 3
• Five mice with different characteristics:
–
–
–
–
4 week old female NOD (J4FD, R4FD)
4 week old female NOD (J4FD)
4 week old male NOD (J4MD)
4 week old female homozygous transgenic
mouse which can't get diabetes (R4FN)
Notice
• Each of the 5 chips were scanned twice
• Two separate stains are used
• This gives us 10 sets of results
Properties of Data that make
defining expression hard
•
•
•
•
There can be saturation
log2(PM / MM) and PM-MM are noisy
MM >> PM for many probes
PMs of the same probe vary about 5 times
less from chip to chip than from probe to
probe within the same probe set.
Saturation problem
Probes reaching maximum in experiment 1
Scanner
Chip
PM MM Value
2
NV21a
354 25
46140
2
NV21b
564 57
46144
2
V21a
1004 83
46141
2
V21b
665 51
46139
1
NV100a
1917 328 46154
1
NV100b
1265 168 46160
1
V100a
3399 1085 46155
1
V100b
2267 446 46149
log2(PM/MM) for defective and normal probe sets in
a chip from experiment 2
The Good News
Section of Data
Bad Probes
All Data
20%
Top 1% of PM / MM
11%
Top 1% of PM
14%
Bottom 1% of PM / MM
29%
Top 0.1% of PM / MM
7%
Top 0.1% of PM
10%
Bottom 0.1% of PM / MM
29%
Histograms of log2(PM/MM) stratifies by log2(PMxMM)/2
for one of the chips in experiment 1
Histograms of log2(PM/MM) stratifies by log2(PMxMM)/2
for chip in experiment 2 for defective and normal probe
Histograms of log2(PM/MM) stratifies by log2(PMxMM)/2
for one of the chips in experiment 3
ANOVA
Normalization
• There are many sources of experimental variation:
– During preparation: e.g. mRNA extraction, introduction of labeling
– During manufacture of array: e.g. amount of oligos on cells
– During hybridization: e.g. amount of sample applied, amount of
target hybridized
– After hybridization: e.g. optical measurements, label intensity,
scanner
• Proper normalization is need before intensities
from different chips are compared
Log ratio vs. average log intensity (MVA) plots of PM,MM
Log ratio vs. avg log intensity (MVA) plots for PM / MM
Normalization
• Pair-wise normalization?
• Which chips do we compare?
• The following three plots show the 3
pairwise comparisons of chips
Control A, ENOB, and NNOA
Normalization based on combined PMs and MMs
Cyclic algorithm (version 0.1)
• For chip j, with entries X1 define the functions
f1,…,fj-1,fj+1,…,fJ
to be the results of smoothing the scatter plot
{Xj-Xk , (Xj+Xk)/2}
• Define the normalized chip as
Xj’= Xj- (f1+…+fj-1+fj+1+…+fJ)/J
• Chips X1,…,XJ are normalized in the same way
• We iterate until Xi’, Xi are very similar for all i.
Before and after normalization
Experiment 1
Experiment 2
Experiment 2
Combined PM and MM
Experiment 2
PM / MM
Experiment 2
PM – MM in a hybrid log scale
Experiment 3
Combined PM and MM
Competing definitions of expression
• Li and Wong fit a model
PM ij  MM ij   i j   ij ,  ij  N (0,  )
2
Consider  i expression in chip i
• Efron et. al. consider log PM – 0.5 log MM
• Another is second largest PM
How do we compare?
• We want small variance, small bias.
• Up to now we don’t know truth in any of
our data sets so hard to assess bias.
• One possibility is to assume some gene is
differentially expressed in the experiments
we study, find it, and look at its probe
profile.
Conclusion
• Features of data suggest that avg.diff may
be improved as a definition of expression
• It seems that normalization is needed to
remove experimental variation and make
meaningful comparison of data from
different chips fair
Acknowledgements
• JHU: Leslie Cope, Tom Coppola, Shwu-Fan Ma,
Skip Garcia
• CNMC: Rehannah Borup, Josephine Chen, Eric
Hoffman
• UC Berkeley: Ben Bolstad
• WEHI: Runa Daniel, Len Harrison