Proteomic Mass Spectrometry. Outline Previous Research Project Goals Data and Algorithms Experimental Results Conclusions To Do List.

Proteomic Mass Spectrometry

Outline

• Previous Research

• Project Goals

• Data and Algorithms

• Experimental Results

• Conclusions

• To Do List

Motivation

• MS spectra has high dimension– Most ML algorithms are incapable of handling

such high dimensional data

• Dimensionality Reduction (DR)– Preserve as much information as possible,

while reducing the dimensionality.

• Feature Extraction (FE)– Removal of irrelevant and/or redundant features

(information)

Previous research

• Usually applies DR then FE• Does Order matter ?

DR: Down Sampling, PCA, WaveletsFE: T-Test, Random Forests, Manual Peak Extraction

• In [conrads03] show that high resolution MS spectra produces better classification accuracy.– Most previous research down samples spectraCONJECTURE: Down Sampling detrimental to

performance.

Project Goals

• Test Down Sampling Conjecture

• Compare FE algorithms (NOTE: Optimal FE is NP-hard !)– Use a simple but fast classifier to test a number

of FE approaches

• Test across different data sets– Are there any clearly superior FE algorithms ?

Three Data Sets

Heart/Kidney (100/100)– 164,168 features, 2 classes

Ovarian Cancer (91/162)– 15,154 features, 2 classes

Prostate Cancer (63/190/26/43)– 15,154 features, 4 classes

• Normal, Benign, Stage 1, Stage 2 Cancer• Transformed into Normal/Benign Vs Cancer (1&2)

Algorithms

Centroid Classifier– given class means P, Q and sample point s

C = argmin (d(P,s), d(Q,s))C = argmin (d(P,s), d(Q,s))

P Q

d(P,s) d(Q,s)

Algorithms

• T-testT-test – do the means of 2 distr. Differ ?

• KS-testKS-test – do the cdf differ ?

• CompositeComposite – (T-test)*(KS-test)

• IFEIFE - Individual Feature Evaluation using the centroid classifier

• DPCADPCA – discriminative principle component analysis

Preliminary Experiments

• Compare normalization approaches

• Compare similarity metrics– Cross correlation– (-L1)– Angular

• Across 3 data sets => 27 configurations

L1à norm;1 à norm;no normalization

Preliminary Experiments (cont)

• No single norm/metric clearly superior on all data sets

• 2-5% increase in performance if suitable normalization and similarity metric chosen (can be up to 10% increase)

• L1-norm with angle similarity metric worked well on Heart/Kidney and Ovarian Cancer sets (easy sets)

• L1-norm and L1-metric best on Prostate 2-class problem (hard set).

Down Sampling

Statistical Tests

• T-test, KS-test, Composite– Ranks features in terms of relevance

• SFS – Sequential Forward Selection– Selects ever increasing feature sets

• I.e., {1}; {1,2}; {1,2,3}; {1,2,3,4}

Heart/Kidney

Ovarian Cancer

Prostate Cancer

Single Feature Classification

• Use each feature to classify test samples

• Rank features in terms of performance

• SFS

Performance Comparison

Performance of Statistical Feature Extraction on Heart/Kidney Data Set

0.75

0.8

0.85

0.9

0.95

1

None(164168)

KS-test (2601)

T-test (887)

Composite(1810)

Best SingleFeature

SFS-L1(3) SFS-Angle(70)

Feature Extraction M ethod (# of Features)

Ac

cu

rac

y

Performance Comparison

Performance of Statistical Feature Extraction on Ovarian Cancer Set

0.75

0.8

0.85

0.9

0.95

1

None(15154)

KS-test(12)

T-test (91) Composite(20)

BestSingle

Feature

SFS-L1(48)

SFS-Angle(14)

Feature Extraction Method (# of Features)

Acc

ura

cy

Performance Comparison

Performance of Statistical Feature Extraction on Prostate Cancer Set

0.50.55

0.60.65

0.70.75

0.80.85

0.90.95

1

Feature Extraction Method (# of Features)

Ac

cu

rac

y

Summary

• For each data set, for each FE algorithm ran 15,000 3-fold cross validation experiments.

• Total of 810,000 FE experiments ran

• DE experiments ~ 100,000 experiments

• Additional 50,000 experiments using DPCA classifier

• did not produce significantly different results than the centroid classifier

Conclusions• HK and Ovarian Data sets considerably easier to classify than

Prostate Cancer• Feature Extraction (in general) significantly improves

performance on all data sets• No single technique superior on all data sets.

– Best Performance using SFS with feature weighting– Smallest feature set with T-test of KS-Test– Composite test inferior to all others.

• Down Sampling appears to be detrimental – What about other Dim. Red. Techniques ?

• E.g. PCA and Wavelets

Conclusions

• Down Sampling appears to be detrimental – What about other Dim. Red. Techniques ?

• E.g. PCA and Wavelets

• What about FE after Down Sampling ?– On Prostate data performance appears to drop

w.r.t. to best single feature.

To Do List

• Check PCA, Wavelets and other DR techniques• Use other (better) classifiers• General Hypothesis

– Use a simple fast classifier together with FE techniques to extract a good feature set

– Replace classifier with a more effective one.• Need to verify that other classifiers respond well to the extracted

features.