Microarray Data Analysis - PowerPoint PPT Presentation

1 / 7
About This Presentation
Title:

Microarray Data Analysis

Description:

How can we combine the 16 to 20 probe-pair intensities (PMs ... so instead we use quantile normalization. Makes the probe-distributions the same for all arrays. ... – PowerPoint PPT presentation

Number of Views:32
Avg rating:3.0/5.0
Slides: 8
Provided by: math75
Category:

less

Transcript and Presenter's Notes

Title: Microarray Data Analysis


1
Microarray Data Analysis
  • RMA
  • Adele Cutler
  • adele_at_math.usu.edu
  • Credits Many thanks to the the R core team and
    the Bioconductor developers, in particular
  • Sandrine Dudoit
  • Robert Gentleman
  • Rafael Irizzary

2
RMA Irizarry et al. (Biostatistics, 2003)
  • How can we combine the 16 to 20 probe-pair
    intensities (PMs and MMs) into a single
    expression measure?
  • Looks at real data, shows that
  • The majority of probe-sets have at least one
    negative PM-MM. (Bad news for taking logs). Page
    6.
  • MM grows with PM, (Fig 1) so it seems that MM is
    capturing some of the signal. (MM is supposed to
    just capture nonspecific binding).

3
RMA Irizarry et al. (Biostatistics, 2003)
  • For high-abundance probes, the distribution of
    log(PM/MM) is bimodal a positive mode (large
    PM, small MM) and a negative mode (small PM,
    large MM). Fig 1.

4
RMA Irizarry et al. (Biostatistics, 2003)
  • There are clear scanner effects if we look at
    either PM or PM-MM. Fig 2. We need to normalize!
  • If we use 2 reps in place of red/green and do MA
    plots as for 2-color cDNA arrays, we get similar
    pictures to the ones for cDNA. So, normalization
    must depend on abundance.
  • We cant normalize using loess as we did for
    2-color arrays (only one color!) so instead we
    use quantile normalization. Makes the
    probe-distributions the same for all arrays.
  • Normalization helps to identify genes that are
    differentially expressed!! (Fig 3)

5
RMA Irizarry et al. (Biostatistics, 2003)
  • Probe-specific effects are additive in the log
    scale. Fig 4.
  • PM-MM shows probe effects, so subtracting is not
    enough to adjust for probe differences. Fig 4.
  • PM/MM attenuates true differences, because MM
    responds to signal. So, we will lose power if we
    use any measure that depends only on PM/MM. Fig
    4.

6
RMA Irizarry et al. (Biostatistics, 2003)
  • We must adjust for nonspecific binding and noise,
    because if we dont, we will lose power for faint
    signals
  • log(1002s) ? log(100s) if s is small.
  • MM looks to have a mixture distribution some of
    the MMs do capture noise and nonspecific
    binding, while others also capture signal.
  • Assume signal is exponential and background is
    normal ?B().

7
RMA Irizarry et al. (Biostatistics, 2003)
  • Let Yij be background adjusted, normalized,
    log-transformed PM. Let
  • Yij ?i ?j ?ij
  • ?i is the expression (log scale)
  • ?j is the probe effect
  • ?ij is noise
  • Fit a statistical model to estimate ?i , use this
    in subsequent analysis.
Write a Comment
User Comments (0)
About PowerShow.com