Exploration, Normalization, Summaries, and Software for Affymetrix Probe Level Data

Transcription

1 Exploration, Normalization, Summaries, and Software for Affymetrix Probe Level Data Rafael A. Irizarry Department of Biostatistics, JHU March 12, 2003

2 Outline Review of technology Why study probe level data? Probe level data expression measures Case study

3 Affymetrix GeneChip 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

4

5 Why Keep Probe Level Data? Quality control Spatial Effects RNA degradation (Leslie Cope) Detection of defective probes Transcript sequence estimates change Ways to reduce to expression measure keep improving

6 Software: Bioconductor R package affy: AffyBatch class exprs Matrix of cel intensities, probes x samples (cel files) se.exprs Matrix of SDs for probe intensities phenodata cdfname annotation description nrow,ncol,notes Sample level covariates, instance of class phenodata Hash table R environment with location (CDF file) info Name of annotation data Object of class MIAME # rows and columns on array and any furthre notes

7 QC

8 QC

9 Case Study Probe level data expression measures Each gene is represented by 20 pairs (PM and MM) of probe intensities Each array has 8K-20K genes Usually there are various arrays Summarize 20 pairs Obtain measure for each gene on each array: Background correction and normalization are issues

10 Default until 2002 (MAS 4.0) GeneChip Avg. diff = software used Avg.diff Α j Α ( PM MM ) j j with A a set of suitable pairs chosen by software. Obvious Problems: Many negative expression values No log transform 1

11 Why use log? Original Scale Log Scale

12 Current default (MAS 5.0) GeneChip new version uses something else signal * = TukeyBiweight{log( PM j MM j )} with MM* a version of MM that is never bigger than PM. Ad-hoc background procedure and scale normalization are used.

13 Expression Data log 2 (expression 1) log2 (expression 2)

14 MA plot A= ½{ log 2 (expression 2) + log2(expression 1) } /2 M=log2 (expression 2 / expression 1)

15 Can this be improved?

16 Use Spike-In Experiment

17 Use Spike-In Experiment

18 Spike-In Data Go back to probe level and find better ways to Background correct Normalize Correct for probe-specific background Summarize Next 4 slides: transcripts spiked-in at increasing concentrations across arrays

19 Why background correct? Concentration of 0 pm Concentration of 0.5 pm Concentration of 1.0 pm 100

20 Why normalize? Compliments of Ben Bolstad

21 Why fit log scale additive model?

22 Statistical Model Instead of subtracting MM Assume PM = B + S To estimate S, use expectation: E[S B+S] After quantile normalization, assume: log 2 S ij = E i + P j + ε ij Estimate E i using robust procedure We call this procedure RMA Does it make a difference?

23 MAS 5.0 Ranks

24 Perfect Ranks

25 MAS 5.0 Ranks

26 RMA Ranks

27 References Irizarry et al: Biostatistics (2003) Bolstad et al: Bioinformatics (2003) Irizarry et al: NAR (2003) affy R package (

28 Current Work Improve Background Model Probe specific background correction Use sequence information to do this MatchAffy Software (Gentleman)

29 GC-content Specific Background Correction

30 Modularity Background correction Normalization Probe Specific BG correction Summary statistic Courtesy of Laurent Gautier

31 Conclusion Working with probe level data can improve bottom line results Important to have flexible software

32 Acknowledgements Ben Bolstadr Francois Collin Leslie Cope Laurent Guatier Robert Gentleman Bridget Hobbs Terry Speed Zhijin Wu Affymetrix Bioconductor Genelogic JHMI Microarray Core Facility R

33 Bottom Line Assessment FC>2, FP FC>2, TP TP, FP=25 AUC, FP<100 MAS % RMA % FC = Fold Change TP = True Positve FP = False Positive AUC = Area under the ROC curve

34 *What makes the difference in AUC? Operation Improvement Preprocessing Log Background (no MM) Normalization Robustness Test-Statistic 60% 78% 343% 5% 20% * * 58% improvement over t-test

35 MAS 5.0

36 RMA

37 What makes the difference (easy)? Operation Log Background (no MM) Normalization Robustness Test-Statistic Improvement 276% 105% 18% 0.3% NA

38 Can this be improved? R= % of data below green line, 50% below red

39 Can this be improved? R= % of data below green line, 50% below red

40 What is the evidence? Lockhart el al: Nature Biotechnology (1996)

41 Affymetrix files Main software from Affymetrix company MicroArray Suite - MAS, now version 5. DAT file: Image file, ~10^7 pixels, ~50 MB. CEL file: Cell intensity file, probe level PM and MM values. CDF file: Chip Description File. Describes which probes go in which probe sets and the location of probe-pair sets (genes, gene fragments, ESTs).