Reproducibility of results of Microarray data

1
Reproducibility of results of Microarray
data Preprocessing on Differentially Expressed
Transcript Selection (Affymetrix GeneChips)
Monika Ray Washington University, Saint Louis,
MO September 14, 2006
Johannes Freudenberg, Weixiong Zhang
2
MOTIVATION
Preprocessing
3
MOTIVATION

• Datasets - Analysis of preprocessing algorithms
  were primarily performed only on
• well-controlled, calibration datasets (Affymetrix
  spike-in data).

• Preprocessing stage usually only background
  correction or normalisation

4
Causes of Noise in Gene Expression Data

• dye effects
• scanner effects
• uneven hybridisation
• array design
• experimenter

Remove noise while retaining intrinsic biological
variations
5
Preprocessing stages and Methods
6
Gene selection
NEURAL NETWORKS SUPPORT VECTOR MACHINES DECISION
TREES

• Wrapper methods
• Filter methods

T-STATISTIC LOG RATIO CLUSTERING PCA
SAM and RANKGENE
7
RESULTS
8

• RCRR-MCRR
• RCRR-RQRR
• RCRR-RCMR
• RCRR-RCRM
• MQRR-RQRR
• MQRR- MCRR
• MQRR-MQMR
• MQRR-MQRM
• RQMM-MQMM
• RQMM-RCMM
• RQMM-RQRM
• RQMM-RQMR
• MCMM-RCMM
• MCMM-MQMM
• MCMM-MCRM
• MCMM-MCMR

RESULTS
9
RESULTS
10
RESULTS
11
RESULTS

• Preprocessing methods have greater impact than
  the gene selection tool used.
• Relative difference in bar heights is phenomenal
  for the same gene selection method.
• Dataset quality also affects outcome AD
  (greater similarity) vs. PTC
• (greater discrepancy).
• 4. Clustering does not show any definite
  patterns.
• 5. GCRMA did not perform any differently than
  other algorithms.

12
CONCLUSION

• Different preprocessing methods have varying
  degrees of effect on downstream analysis.
• Choice of gene selection method also affects the
  outcome of analysis.
• Choice of datasets also affect outcome of
  results.
• Discrepancies in results are substantial even on
  a single platform.
• Microarray analysis is subjective.
• No standard tools for evaluating methods.
• Reproducibility of results is difficult without
  any standardisation.
• Emotional dependence on p-values. Statistical
  significance does not always imply biological
• significance

13
What do we do?

• Understand why current methods fail.
• Efforts directed towards increasing consistency
  among results from same platform.
• Along with standardising research protocols and
  statistical analysis tools, attention should
• be paid towards standardising preprocessing
  algorithms.
• Evaluate processing methods on real-world
  datasets as well as well-controlled, calibration
• datasets.
• Microarray chip design is crucial.
• Characteristics of high-throughput data need to
  be well understood and addressed
• accordingly in the design and execution of
  experiments.
• Large sample sizes do not be in a hurry to
  publish results!
• Develop more meaningful tools for significance
  analysis.

14
THANK YOU !!!