Evaluation of Affymetrix array normalization procedures based on spiked cRNAs

1
Evaluation of Affymetrix array normalization
procedures based on spiked cRNAs

• Andrew Hill
• Expression Profiling Informatics
• Genetics Institute/Wyeth-Ayerst Research

2
Outline

• The GI/Harvard C. elegans array dataset as a
  normalization testbed
• Some general challenges of array data reduction
• GeneChip Scaled Average Difference (ADs)
• the constant mean assumption
• A purely spike-based normalization strategy
  (Frequency)
• A hybrid normalization (Scaled Frequency)
• Conclusions

3
GI/Harvard C. elegans dataset

• This data set used to evaluate several
  normalization procedures
• Experiments
• 8 developmental stages of the worm C. elegans
  were profiled, ranging from egg to adult worm
• n2-4 replicate hybridizations for most array
  designs at most stages
• 52 total arrays
• Arrays
• Three custom worm GeneChip designs (A, B, and C)
• Each array monitors between 5700-6700 ORFs, in
  aggregate 98 of the worm genome
• Chip A ORFs with cDNA/EST matches in AceDB
• Chips B/C other ORFs
• Several worm ORFs tiled on all 3 arrays for
  across-array-design comparisons
• Science 290 809-812 Genome Biology (in the
  press)

4
Some challenges of Affymetrix GeneChip data
reduction

• Array data from Affymetrix GeneChip sofware
  (pre-MAS 5.0)
• negative low intensity signals
• lack of across-design normalization standard
• limited QC information
• Spike-based normalization methods can help to
  address each of these challenges
• Normalization array scaling of average
  difference data from multiple arrays/designs to
  minimize technical noise among arrays
• Current standard normalization procedure is a
  global scaling procedure the GeneChip scaled
  average difference (ADs)

5
GeneChip Scaled Average Difference (ADs)

• The trimmed (2) mean intensity of all probesets
  on all arrays is scaled to a constant target
  level.
• Works well in many cases (e.g. replicates)
• Some obvious situations where the constant mean
  assumption may not be well supported.

6
Constant mean assumption problematic cases

• Chips monitoring a small fraction of
  transcriptome
• Non-random gene selection on arrays (e.g. C.
  elegans A vs. B/C)
• Large biological variation in expression

7
A cRNA spike-based normalization procedure
(Frequency)

• Add 11 biotin-labeled cRNA spikes to each
  hybridization cocktail
• Construct a calibration curve
• Use the Absent/Present calls for the spikes to
  estimate array sensitivity
• Dampen AD signals below the sensitivity level to
  eliminate negative AD values.

8
Eleven spiked cRNAs
9
Response to spikes over 2.5 log range
Figure 2

• Fit response with S-plus GLM, gamma error model,
  zero intercept.
• Power law fit ADkFn yields n0.93
• cRNA mass, scanner PMT gain are important
  determinants of response

10
Chip sensitivity calculation

• Consider A/P calls as binary response against
  log(known frequency)
• Compute sensitivity as 70 likelihood level by
  either interpolation or logistic regression
• Dampen computed frequencies below sensitivity
• F lt 0 F avg(0,S)
• 0ltFltS Favg(F,S)

11
How well does it work?
12
Reproducibility of F metric (A array)
13
Example of spike-skewed hybridization (36 hr
sample)

• cRNA spikes are well normalized at the expense of
  worm genes
• Suggests inconsistency between ratio of spikes
  to worm cRNA across samples spike skew

14
Sources of spike skew

• Actual concentration of spikes may not be nominal
  due to variation in cRNA purity
• Causes liquid handling of small microlitre
  volumes, side reactions in cDNA/IVT process
  produce UV-absorbing, non-hybridizable
  contaminants
• Result random per-hybe noise term introduced
  into normalized frequencies

15
An alternative hybrid normalization Scaled
frequency (Fs)

• Need to reduce or eliminate spike skew as a
  source of experimental variation in normalized
  frequencies
• Average the globally scaled spike response over a
  complete set of arrays

16
Scaled frequency description

• Define a set of arrays
• Compute ADs for all arrays
• Pool spike responses and fit single model to
  pooled response
• Calibrate all arrays with single calibration
  factor
• Compute array sensitivity and dampen frequencies
  as in the frequency approach.

17
A pooled, scaled spike response

• Fit response with S-plus GLM, gamma error model,
  zero intercept.

18
Reproducibility of Fs metric (A array)
19
Scaled frequency cross design reproducibility
(A,B,C arrays)
Three messages tiled on all array designs and
called Present on all 0h arrays
20
Conclusions

• Array response to spiked cRNAs can be close to
  linear over 2.5 logs of concentration.
• A chip sensitivity metric can be computed from
  Absolute Decisions associated with spikes a very
  useful QC metric.
• Normalization based only on spikes performs
  inconsistently in some cases due to
  ill-quantitation of cRNAs, but can still be
  valuable when constant-mean assumption is
  violated. Better cRNA quantitation and process
  control will help.
• A hybrid approach based on global scaling and
  spikes performs the same as global AD scaling for
  single designs, and also allows cross-design
  comparisons

21
Acknowledgements

• Donna Slonim
• Maryann Whitley
• Yizheng Li
• Bill Mounts
• Scott Jelinsky
• Gene Brown

• Harvard University
• Craig Hunter
• Ryan Baugh

22
Extra slides follow ( not part of presentation)
23
Simulations (description)

• Simulations were performed
• Governing equation

24
Figure 4
CV characteristics of simulated data
25
Simulations spike skew degrades reproducibility
of frequency (A array)
26
Figure 7
Simulations spike skew degrades accuracy of
frequency