Preprocessing and Biomarker Extraction Methods in Mass Spectrometry

1
Preprocessing and Biomarker Extraction Methods in
Mass Spectrometry

• Julien PRADOS
• Centre Universitaire dInformatique
• Université de Genève
• Geneva Artificial Intelligence Laboratory
• 15-06-2007

2
Introduction(1/2)Mass Spectrometry (MS)

• SELDI-MS is a technology able to reflect on a
  signal (mass spectrum) expression of proteins
  present in a specimen
• Intensity (area) of peaks are linked to quantity
  of protein in the specimen
• x-axis Mass (/Charge) of proteins
• y-axis Number of protein with mass x

Protein A Mass 4770.7Da
Specimen 1 Positive
Protein B Mass 9226.5Da
Specimen 2 Negative
3
Introduction(2/2)MS Learning

• Goals
• Diagnostic recognize the group of a specimen
  from his mass spectrum (classification).
• Biomarker extraction find a small number of
  protein whose expression allow the diagnostic the
  specimens (feature selection).
• General Context knowledge extraction from
  signals ? potential impacts on
• Image analysis, Video analysis, Audio analysis,
  Sismologic data, Cardiac data, Climatic data,
  Financial data, ... All numeric devices with
  captors

4
OutlinePhD Contributions

• Preprocessing of Mass Spectra
• Peak Picking
• Peak Alignment
• Reproducibility of MS technology
• Preprocessing Evaluation
• Learning (Feature Selection)
• feature selection stability
• logRatio kernel for SVM base FS

5
OutlinePhD Contributions

• Preprocessing of Mass Spectra
• Peak Picking
• Peak Alignment
• Reproducibility of MS technology
• Preprocessing Evaluation
• Learning (Feature Selection)
• feature selection stability
• logRatio kernel for SVM base FS

6
Preprocessing (1/6)A General ML Issue

• Preliminary step of learning whose goals are
• structure the data
• clean the information and keep only useful one
• integrate domain knowledge
• choice of a representation
• Included in every machine learning problems, it
  has an heavy impact on learning performances and
  takes most of the data miner time
• ? Many current research try indirectly to
  faciltate preprocessing
• learning from complex structure (graphs, trees)
• kernel methods change the data representation
  just by changing a function

7
Preprocessing (2/6)MS Data Preprocessing

• Ideal dataset is protein expression matrix
  matrix of concentration of proteins in specimens
• Its construction from mass spectra involves
  several steps
• Peak picking, peak alignment, baseline
  correction, peak area estimation, noise level
  estimation

8
Preprocessing (3/6)Peak Picking (1/2)

• CBMS 06
• Based on a relative peak characterization and
  valley notion
• Valley of a point p is the minimum intensity
  point located on the left and right of p s.t.
  their is no point with intensity greater than p
  between them.
• Peak Points with deepest valleys

Valley depth
9
Preprocessing (4/6)Peak Picking (2/2)

• Benefits
• Parameter less
• Provide a confidence in peaks
• No assumption on peak width
• Robust to baseline correction
• Doesnt require signal smoothing
• Extension to 2D peak detection (LCMS images)
  even nD peak detection (object detection in
  videos)
• Fast O(n)
• Drawback
• Cannot resolve overlapping peaks

10
Preprocessing (5/6)MS Experiment Alignment

• Done by agglomerative hierarchical clustering
  with constraints
• at each iteration, merge closest clusters
• except if the new cluster contains two peaks from
  the same spectrum
• or if peak error in the new cluster exceed a
  maximum value
• Two strategies for missing values
• set missing values to zero because there is no
  peak
• retrieve signal intensity in spectra (not obvious
  for peak areas)

11
Preprocessing (6/6)MS Preprocessing
Generalisation

• Peak Picking analyse individually each instance
  and identify the elements in the instances
• Peak Alignment identify elements of the
  different instances related to the same
  information

12
OutlinePhD Contributions

• Preprocessing of Mass Spectra
• Peak Picking
• Peak Alignment
• Reproducibility of MS technology
• Preprocessing Evaluation
• Learning (Feature Selection)
• feature selection stability
• logRatio kernel for SVM base FS

13
SELDI-TOF Reproducibility (1/6)Overview

• The Fact when a experiment is repeated several
  times, the mass spectra look different.
• General problem in MS e.g. over 3 LCMS
  experiments only 20 of protein are identified in
  common !
• JChromatB06 SELDI-TOF Reproducibility
• How reproducible is SELDI technology ?
• How experimental protocol affects reproducibility
  ?
• How similar are protocols ?

14
SELDI-TOF Reproducibility (2/6) Experimental
Setup

• 1 urine sample (pool of 4 individuals)
• 8 experimental protocols with different
• urine dilution
• matrices
• ultra filtration methods
• for each experimental condition 50 spectra 10 x
  5 replicates

15
SELDI-TOF Reproducibility (3/6)Peak Picking and
Alignment
8 x
16
SELDI-TOF Reproducibility (4/6) Results

• Estimate pourcentage of peaks found in at least
  10/10, 9/10, 8/10 ... 1/10 replicates

17
SELDI-TOF Reproducibility (5/6) Protocol
Complementarity

• How similar are protocols ? how do they
  complement themselves ?
• ? Hierarchical clustering of methods using
  Tanimoto distance between sets of peaks found

18
SELDI-TOF Reproducibility (6/6) Final Comment

• Reproducibility is a necessary condition but is
  not sufficient its easy to define a very
  reproducible peak picking but it would extract
  irrelevant information for the diagnostic problem
• ? we need also to assert the quality of the
  preprocessed information

19
OutlinePhD Contributions

• Preprocessing of Mass Spectra
• Peak Picking
• Peak Alignment
• Reproducibility of MS technology
• Preprocessing Evaluation
• Learning (Feature Selection)
• feature selection stability
• logRatio kernel for SVM base FS

20
Preprocessing Evaluation (1/1)Choosing a Data
Representation

• CBMS06
• Loss of information in preprocessing ?
• How the different representations affect
  information content ?
• ? Compare classification performances of
• SVM Support Vector Machines (SMO)
• Decision Tree (J48)
• Nearest Neighbor (IBk)
• on datasets
• Stroke (Stk)
• Prostate Cancer (Pro)
• Ovarian Cancer (Ova)
• for different preprocessing options
• no preprocessing (raw)
• peak intensity retrieve missing values from raw
  data (is)
• peak intensity zero for missing values (iz)
• peak area zero for missing values (az)

Classification error
21
OutlinePhD Contributions

• Preprocessing of Mass Spectra
• Peak Picking
• Peak Alignment
• Reproducibility of MS technology
• Preprocessing Evaluation
• Learning (Feature Selection)
• feature selection stability
• logRatio kernel for SVM base FS

22
Learning - FS (1/10)Problem Presentation (1/1)

• In preprocessing we do our best to reduce the
  information and to focus on relevant one ignoring
  groups of specimens (unsupervised)
• For biomarker extraction task, the learning goal
  is to further reduce the information considering
  class labels of the specimens (supervised FS)

23
Learning - FS (2/10)Feature Selection Stability
(1/3)

• In biomarker extraction from MS stability of
  selected features is also capital
• it is preferable that a FS method select same
  biomarkers when little data perturbations are
  introduced
• ICDM05, KIS07 Introduces a framework for
  estimation of FS stability
• apply FS in different folds of a CV to guaranty
  22 of non overlapping instances
• Estimate mean of attribute overlap in the 10
  selected feature sets

24
Learning - FS (3/10)Feature Selection Stability
(2/3)

• With this framework stability of FS methods can
  be compared
• Information Gain
• ReliefF
• linear-SVMRFE
• linear-SVM (no RFE)
• Random FS
• Random FS can be computed theoretically
• Results on MS
• RF gt SVMONE gt SVMRFE gt IG gt Random

25
Learning - FS (4/10)Feature Selection Stability
(3/3)

• Stability is a necessary condition but is not
  sufficient predictive power of selected feature
  has to be verify
• Results show a general superiority of
  performances with SVM based feature sets
• FS Stability Performance ? Concentrate on SVM
  based FS

26
Learning - FS (5/10)LogRatio Kernel (1/6)
?(X)
?(X)1,3
log(X)
1
3
n
2

• Biological motivations
• Log of Ratio often used to compare expressions
  (genomic)
• Pair-wise combination focus learning on
  protein-protein interactions
• Insensitive to instance normalization
• Similar to polynomial kernel of degree 2, but
  instead of pair wise products of attributes,
  LogRatio feature space consider logarithm of
  their ratios
• The n2 components of ?(x) can be visualized on a
  plot with n bars
• LogRatio centring makes it also insensitive to
  attribute normalization

27
Learning - FS (6/10)LogRatio Kernel (2/6)

• SVM decision function can be simplified like in
  linear SVM case
• Vector W store the model like in linear SVM case
• ?(W) is a normal vector to separating hyperplan
  in the feature space
• ? attributes can be ranked according to ?(W)

28
Learning - FS (7/10) LogRatio Kernel (3/6)

• Less significant attribute the one with median
  log(W) value because it minimizes absolute
  distance to all others
• ? Rank attributes according to their distance to
  median
• Perform RFE to improve ranking ? logRatio-SVMRFE

?(W)
log(W)
1
3
n
2
29
Learning - FS (8/10) LogRatio Kernel (4/6)

• logRatio-SVMRFE appears good in removing
  redundant attributes because of the pair-wise
  attribute combination
• Removing Redundant Attribute is not an easy
  problem
• Attributes which express the same information in
  different units can be safely removed
• In presence of noise, attributes which reflects n
  measures of the same information are
  complementary (by computing their means, we
  improve precision of the measure)
• Redundancy is an important concept in expression
  analysis it is a sign of protein interaction
• logRatio-SVMRFE can distinguish between the two
  type of redundancy

30
Learning - FS (9/10) LogRatio Kernel (5/6)

• Data Iris dataset augmented by duplication of
  its attributes in order to express each value in
  cm mm
• logRatio-SVMRFE rank first one of each
  information (petal/sepal length/width)
• linear-SVMRFE rank 2 times the same information
  first once express in cm once in mm
• Noise added only on cm attributes
• logRatio-SVMRFE rank first all mm attributes
• linear-SVMRFE rank first cm mm attributes
• ? Perfect behaviour of logRatio-SVMRFE

31
Learning - FS (10/10) LogRatio Kernel (6/6)

• Concerning classification performances,
  logRatio-SVM classifier performs similar to
  linear-SVM for MS dataset, and better on
  text-mining.
• Essentially because of log transformation
• But logRatio-SVMRFE seems less good than
  linear-SVMRFE in terms of stability selected
  attributes
• inadequate model interpretation ?
• more RFE step should be included ?

32
Conclusions

• Facilitate and Reduce MS preprocessing effort
• FS method focusing on attribute redundancy and
  protein-protein interactions
• Relative data representation helps in reaching
  those goals
• peak picking
• logRatio kernel feature selection
• The role of stability in data reduction and
  simplification
• Preprocessing
• Learning

33
Perspectives

• Extensions of preprocessing to LCMS data
• Harder Alignment Problems
• More Complex Reproducibility Problems
• Needs for Real Time Processing Learning
• Further integration of preprocessing in learning
• We investigated tuning preprocessing parameters
  considering classification performances Pro04
• preprocessing should not be separated from
  learning investigate integration of unsupervised
  methods and supervised methods to automatically
  preprocess information considering the
  classification objective
• Is it possible to extend logRatio kernel s.t. the
  feature space can be visualized in 2D (or nD)
  space instead of 1D ? This would bring enhanced
  information on attribute redundancy.

34
PhD Status

• Problems
• No article on logRatio kernel
• Additional experiments
• Compare preprocessing to other one
• PhD Writting
• Estimation 1/3 written (no corrections included)
• Oral Exam
• Subject Error Evaluation ? ICLP ?
• Date Sep ?
• PhD Defense
• Expected Date End 2007

35
References

• KIS07 A. Kalousis, J. Prados, and M. Hilario.
  Stability of feature selection algorithms.
  Knowledge and Information Systems,
  2007.JChromatB P. Zerefos, J. Prados, S.
  Garbis, A. Kalousis and A. Vlahou. Urinary
  protein profiling by MALDI-TOF-MS Tackling
  sample preparation and bioinformatics issues.
  Journal of Chromatography B, 2006.CBMS06 J.
  Prados, A. Kalousis, and M. Hilario. On
  Preprocessing of SELDI-MS Data and its
  Evaluation. CBMS, 2006. talkICDM05 A.
  Kalousis, J. Prados, and M. Hilario. Stability of
  feature selection algorithms. ICDM,
  2005.PKDD05 A. Kalousis, J. Prados, E.
  Rexhepaj, and M. Hilario. Feature extraction from
  mass spectra for classification. In 6th European
  Conference on Principles and Practice of
  Knowledge Discovery in Databases, 2005.6 M.
  Hilario, A. Kalousis, J. Prados, and P. A. Binz.
  Data mining for mass spectra-based cancer
  diagnosis and biomarker discovery. Drug Discovery
  Today Biosilico, 2(5)214-222, September
  2004.ICTAI04 A. Kalousis, J. Prados, J. C.
  Sanchez, L. Allard, and M. Hilario. Distilling
  classification models from cross-validation runs
  an application to mass spectrometry. ICTAI, 2004.
  talkPro04 J. Prados, A. Kalousis, L. Allard,
  O. Carrette, J. C. Sanchez, and M. Hilario.
  Mining mass-spectra for diagnosis and biomarker
  discovery of cerebral accidents. Proteomics,
  42320-2332, 2004.

36
PreprocessingPeak Picking Peak Area(1)

• Definition The left valley, vl of a point p is
  the minimum intensity point located on the left
  of p s.t. their is no point with intensity
  greater than p between vl and p. (idem for right)
• Peak points with high valley depth

37
PreprocessingPeak Picking Peak Area(2)

• Peak Area
• approximated by the area of the triangle obtained
  after fitting of piecewise linear models in two
  parts
• Benefits
• Parameter less
• Feed back on confidence in a peak
• No assumption on peak width
• Doesnt require baseline correction
• Doesnt require signal smoothing
• Extension to 2D peak detection possible (LCMS)
• Fast O(n)
• Drawback
• Cannot resolve overlapping peaks