Felix Naef

1
Felix Naef Marcelo Magnasco, GL meeting, Nov.
19 2001felix_at_funes.rockefeller.eduOutline
Excursions into GeneChip data analysis

• Background subtraction
• Probeset statistics

2
Background estimation

• estimate both mean B and fluctuations s
• needed in low-intensity regime
• includes light reflection from substrate,
  photodetector dark current, some
  cross-hybridization (i.e. small residues)
• by the CLT, background is expected to be a
  Gaussian variable

3

• idea B is insensitive to MM and visible at low
  intensity
• select probes such that PM-MM lt e (locally?)
• use e50 (new) or 100 (old settings)
• P(PM) or P(MM) is convolution of Gaussian and
  step function

4

• example
• (e100)

dependence on e
5
trick for dealing with negative values
6
PM vs. MM distribution
zoom
7
PM vs. MM histogram
8
MMgtPM across different chips
MMgtPM not concentrated at low intensities 27 of
probe pairs with MMgtPM are in the top quartile
9
probe pairs trajectories (80 chips)

• take all (PM, MM) for
• a given probe set
• center of mass (x,y)
• ellipsoid of inertia
• gt s1 and s2
• histogram the cms
• color code acc. to
• s s1 / S(min(x, y))
• noise detrending

10
all probe sets blue large s green mid red
small
11
probes with well defined trajectories
(eccentricity gt 3) 1/3 of probes
blue large green mid red small
12
PM within a probe set
Are the brightness of the probes reasonably
uniform? Or do different probes have very
different hybridization efficiencies?
13
So what can possibly be happening?

• sequence dependent hybridization efficiencies
• are kinetic effects important?
• cross-hybridization beyond what is detectable by
• MM probes
• this is hard to assess without sequence info
• sequence dependent fabrication efficiencies?
• variable probe densities

14
Composite scores

• What have we learned from previous slides?
• MM are not consistently behaving as expected
• What about not using them ?
• The probe set intensities vary over decades
• difficult to estimate absolute intensities using
• averages (alternative Li and Wong)
• - we focus on ratio scores

15
Outline of algorithm

• estimate background (mean and std)
• discard noisy and saturated probes
• use either only PM or PM-MM as raw intensities
• average the remaining log-ratios in an outlier
  robust way (robust regression to intercept), SE
• normalize by centering (event. local) log-ratio
  distribution