On Preprocessing SELDIMS and its Evaluation

1
On Preprocessing SELDI-MS and its Evaluation

• Julien Prados, Alexandros Kalousis, Melanie
  Hilario

2
Introduction

• Context
• diagnosis and biomarker extraction from SELDI-TOF
  mass spectra
• Issues
• preprocess mass spectra
• tune preprocessing parameters
• evaluate preprocessing quality

3
Typical Work Flow

• Preprocessing
• Baseline Removal
• Normalisation (TIC)
• Noise Estimation/Elimination
• Peak Detection
• Peak alignment

Spectra with Control Diseased labels
List of Discriminant Features
A learning dataset
Classification model for prediction of patient
state from its mass spectrum
Machine Learning
4
Peak Definition

• Valley definition for a point p The minimum
  points on left and right of p such that their is
  no point with intensity higher than the intensity
  of p between them.
• Peak definition a spectrum point p is considered
  a peak if its left and right valleys are deeper
  than the noise level.
• Remark
• No assumption on peak width

5
Peak Detection
6
Visual Choice of Peak Detection Parameter

• ? We used a peak detection parameter of 2.5

7
Peak Area

• p fixed point found by peak detection
• The signal is splitted into regions according to
  the minima between two consecutive peaks
• In each region, pl and pr are found by mean least
  square fitting of a piecewise linear model in two
  segments (horizontal and oblique)
• Area of the peak is given by area of the triangle
  (pl, p, pr)

8
Peak Alignement Missing Values

• Peak alignment performed by hierachical
  clustering (closest peaks are merged)
• Two strategies for missing values
• set missing values to zero because there is no
  peak
• retrieve signal intensity (not obvious for peak
  area)

?
?
?
9
Data Representation Evaluation

• Error evaluation of 3 classification algorithms
• Instance Base Learning (IBk)
• Decision Tree (J48)
• Support Vector Machine (SMO)
• On 3 data representations
• peak intensity signal intensity for missing
  values (is)
• peak intensity zero filling of missing values
  (iz)
• peak area zero filling of missing values (az)
• For 3 datasets
• Stroke (Stk)
• Prostate Cancer (Pro)
• Ovarian Cancer (Ova)

10
Data Representation Evaluation Results

• is/iz ? filling missing values with signal
  intensity instead of zeroes retains more
  discriminatory informations
• iz/az ? using area or intensity does not result
  in significant differences
• is/raw ? no significant information lost in
  preprocessing, but a much more compact
  representation gain

11
Influence of Peak Detection Parameter

• Choosing is representation and SMO algorithm,
  what is the influence of peak detection parameter
  on the information content of the preprocessed
  datasets ?

12
Conclusion

• We proposed a preprocessing pipeline with a new
  peak detection algorithm and a way of computing
  peak areas.
• The preprocessing has been applied on three
  datasets without significant information loss,
  producing a much more compact representation.
• Classification performance depends heavily on
  parameter tunning, it should be done in an
  informed manner
• manual selection
• automatic selection

13
Perspectives

• Preprocessing pipeline used in a reproducibility
  study of MALDI-TOF MS Zeferos et al. Sample
  Preparation and Bioinformatics in MALDI Profiling
  of Urinary Proteins. Submitted, JChromat, 2006
• Peak detection algorithm has been extended to 2D
  (and even nD) and applied on Nano-LC MS

14
2D Nano-LC Peak Detection
15
Thank you !
16
Preprocessing Objectives

• Correct signal distortions (baseline,
  normalization)
• Reduce dimensionality of learning problem (peak
  detection)
• Avoid removing discriminative informations

17
Signal Distortions Correction

• Baseline estimated with open operator (local
  maxima of the local minima) in a moving window
  of containing 5 of the total number of point in
  the mass spectrum
• Total Ion Current normalization with part of the
  signal gt 2000 Da

18
vd vs Error
19
Advertising