This document discusses using feature selection techniques to address the curse of dimensionality in microarray data analysis. It presents the problem of having many more features than samples in bioinformatics tasks like cancer classification and network inference. It describes filter, wrapper and embedded feature selection approaches and proposes a blocking strategy that uses multiple learning algorithms to evaluate feature subsets in order to improve selection robustness when samples are limited. Finally, it lists several microarray gene expression datasets that are commonly used to evaluate feature selection methods.
1 of 29
Download to read offline
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
More Related Content
Feature selection and microarray data
1. On the use of feature selection to deal
with the curse of dimensionality in
microarray data
Gianluca Bontempi
gbonte@ulb.ac.be
Machine Learning Group
ULB, Université Libre de Bruxelles
Boulevard de Triomphe - CP 212
Bruxelles, Belgium
http://www.ulb.ac.be/di/mlg
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 1/2
2. Outline
• The ULB Machine Learning Group
• The feature selection problem
• Feature selection and bioinformatics.
• Feature selection as a stochastic optimization problem.
• A blocking idea to improve wrappers
• Final considerations.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 2/2
3. ULB Machine Learning Group (MLG)
• 7 researchers (1 prof, 6 PhD students), 4 graduate students).
• Research topics: Local learning, Classification, Computational statistics, Data
mining, Regression, Time series prediction, Sensor networks, Bioinformatics.
• Computing facilities: cluster of 16 processors, LEGO Robotics Lab.
• Website: www.ulb.ac.be/di/mlg.
• Scientific collaborations in ULB: IRIDIA (Sciences Appliquées), Physiologie
Moléculaire de la Cellule (IBMM), Conformation des Macromolécules Biologiques
et Bioinformatique (IBMM), CENOLI (Sciences), Microarray Unit (Hopital Jules
Bordet), Service d’Anesthesie (ERASME).
• Scientific collaborations outside ULB: UCL Machine Learning Group (B),
Politecnico di Milano (I), Universitá del Sannio (I), George Mason University (US).
• The MLG is part to the "Groupe de Contact FNRS" on Machine Learning.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 3/2
4. ULB-MLG: running projects
1. "Integrating experimental and theoretical approaches to decipher the molecular
networks of nitrogen utilisation in yeast": ARC (Action de Recherche Concertée)
funded by the Communauté Fran ˛Açaise de Belgique (2004-2009). Partners:
IBMM (Gosselies and La Plaine), CENOLI.
2. "COMP2
SYS" (COMPutational intelligence methods for COMPlex SYStems)
MARIE CURIE Early Stage Research Training funded by the European Union
(2004-2008). Main contractor: IRIDIA (ULB).
3. "Predictive data mining techniques in anaesthesia": FIRST Europe Objectif 1
funded by the Région wallonne and the Fonds Social Européen (2004-2009).
Partners: Service d’anesthesie (ERASME).
4. "AIDAR - Adressage et Indexation de Documents Multimédias Assistés par des
techniques de Reconnaissance Vocale": funded by Région Bruxelles-Capitale
(2004-2006). Partners: Voice Insight, RTBF, Titan.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 4/2
5. Feature selection and bioinformatics
• The availability of massive amounts of experimental data based
on genome-wide studies has given impetus in recent years to a
large effort in developing mathematical, statistical and
computational techniques to infer biological models from data.
• In many bioinformatics problems the number of features is
significantly larger than the number of samples (high feature to
sample ratio datasets).
• Examples can be found in the following bioinformatics tasks:
• Breast cancer classification on the basis of microarray data.
• Network inference on the basis of microarray data.
• Analysis of sequence/expression correlation.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 5/2
7. Breast cancer classification
• Systematic investigation of dependency between expression
patterns of thousands of genes (measured by DNA microarrays)
and specific features of phenotypic variation in order to provide
the basis for an improved taxonomy of cancer.
• It is expected that variations in gene expression patterns in
different tumors could provide a “molecular portrait” of each
tumor, and that the tumors could be classified into subtypes
based solely on the difference of expression patterns.
• In litterature [20] classification techniques have been applied to
identify a gene expression signature strongly predictive of a short
interval to distant metastases in patients without tumor cells in
local lymph nodes at diagnosis. In this context the number n of
features equals the number of genes (ranging from 6000 to
30000) and the number N of samples is the number of patients
under examinations (about hundreds).
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 7/2
8. Inference of regulatory networks
• Most biological regulatory processes involve intricate networks of
interactions and it is now increasingly evident that predicting their
behaviour and linking molecular and cellular structure to function
are beyond the capacity of intuition.
• Dynamic regulation models are black-box prediction models
which represent gene activity with continuous values. Examples
of dynamic regulation models are linear relationships of the form
xi(t + δt) = f(x1(t), . . . , xn(t)), i = 1, . . . , n
where xi is the expression level of the i th gene at time t.
• In this case, inferring a dynamic model boils down at learning the
unknown dependency on the basis of expression data. These
problems demand the estimation of a number of predictive
models for each gene, where the number of features equals the
number of measured genes.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 8/2
9. Correlating motifs and expression levels
• These methods consists in directly correlating expression levels
and regulatory motif present in presumptive transcription control
regions [4, 19].
• In [4], the expression of a gene in a single experimental condition
is modelled as a linear function E = a1S1 + a2S2 + · · · + anSn of
scores computed for sequence motifs in the upstream control
region. These sequence motif scores incorporate the number of
occurrences of the motifs and their positions with respect to the
gene’s translation start site.
• In other terms the sequence motifs of a specific gene are
considered as explanatory variables (feature inputs) of a
statistical model which correlates sequence features and
expression of the gene. For instance the number of features is
n = 4m for motifs of length m.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 9/2
10. Feature selection
• In recent years many applications of data mining (text mining,
bioinformatics, sensor networks) deal with a very large number n
of features (e.g. tens or hundreds of thousands of variables) and
often comparably few samples.
• In these cases, it is common practice to adopt feature selection
algorithms [6] to improve the generalization accuracy.
• There are many potential benefits of feature selection:
• facilitating data visualization and data understanding,
• reducing the measurement and storage requirements,
• reducing training and utilization times,
• defying the curse of dimensionality to improve prediction
performance.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 10/2
11. The two issues of f.s.
Two main issues make the problem of feature selection a highly
challenging task:
Search in a high dimensional space: this is known to be a NP-hard
problem.
Assessment on the basis of a small set of samples: this is made difficult by
the high ratio between the dimensionality of the problem and the
number of measured samples.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 11/2
12. Approaches to f.s.
Three are the main approaches to feature selection:
Filter methods: they are preprocessing methods. They attempt to assess the merits of
features from the data, ignoring the effects of the selected feature subset on the
performance of the learning algorithm. Examples are methods that select
variables by ranking them through compression techniques (like PCA) or by
computing correlation with the output (e.g. Gram-Schmidt, mutual information).
Wrapper methods: these methods assess subsets of variables according to their
usefulness to a given predictor. The method conducts a search for a good subset
using the learning algorithm itself as part of the evaluation function. The problem
boils down to a problem of stochastic state space search. Example are the
stepwise methods proposed in linear regression analysis.
Embedded methods: they perform variable selection as part of the learning procedure
and are usually specific to given learning machines. Examples are classification
trees, regularization techniques (e.g. lasso).
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 12/2
13. Wrapper selection
• The wrapper search can be seen as a search in a space
S = {0, 1}n where a generic vector s ∈ S is such that
sj =
0 if the input j does NOT belong to the set of features
1 if the input j belongs to the set of features
• We look for the optimal vector s∗ ∈ {0, 1}n such that
s∗
= arg min
s∈S
Miscl(s)
where Miscl(s) is the generalization error of the model based on
the set of variables described by s.
• the number of vectors in S is equal to 2n
.
• for moderately large n, the exhaustive search is no more possible.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 13/2
14. Wrapping search strategies
Various methods have been developed for evaluating only a small
number of variables by either adding or deleting one variable at a time.
We consider here some greedy strategies:
Forward selection: the procedure starts with no variables. The first input
selected is the one which allows the lowest misclassification error.
The second input selected is the one that, together with the first,
has the lowest error, and so on, till when no improvement is made.
Backward selection: it works in the opposite direction of the forward
approach. We begin with a model that contains all the n variables.
The first input to be removed is the one that allows the lowest
generalization error.
Stepwise selection: it combines the previous two techniques, by testing
for each set of variables, first the removal of features beloning to
the set, then the addition of variables not in the set.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 14/2
15. Open issues
• The wrapper approach to feature selection requires the
assessment of several subset alternatives and the selection of the
one which is expected to have the lowest misclassification error
during generalization.
• To tackle this problem, we need to perform a search procedure in
a very large space of subsets of features aiming to minimize a
leave-one-out or more in general a cross-validation criterion.
• This practice can lead to a strong bias selection in the case of
high dimensionality problems.
• The feature selection problem may be formulated in terms of a
stochastic optimization problem where the selection of the best
subset has to be based on a sample estimate Miscl.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 15/2
17. F.s. and stochastic discrete optimization
We may formulate the feature selection problem as a discrete
optimization problem [10]
s∗
= arg min
s∈S
Miscl(s) (1)
be the optimal solution of the feature selection problem and the
relative optimal generalization error, respectively.
• Unfortunately the Miscl for a given s is not directly measurable but
can only be estimated by the quantity Miscl(s) which is an
unbiased estimator of Miscl(s).
• The wrapper approach to feature selection aims to return the
minimum ˆs of a cross-validation criterion Miscl(s)
ˆs = arg min
s∈S
Miscl(s) (2)
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 17/2
18. Probability of correct selection
Let Miscl(sj), j = 1, . . . , m be the (unknown) misclassification error of
the jth feature subset and assume that the subsets are indexed in
such a way that Miscl(sm) ≤ Miscl(sm−1) ≤ · · · ≤ Miscl(s1) so that
(unknown to us) the feature set sm = s∗ is the best one. Assume also
that Miscl(sm−1) − Miscl(sm) ≥ δ > 0.
Prob {ˆs = s∗
} = Prob Misclm < Misclj, ∀j = m ≥
≥ Prob Zj <
δ
σm 4/N
, ∀j = m
where Misclj − Misclm ≥ δ > 0 for all j = 1, . . . , m − 1 and the random
vector (Z1, Z2, . . . , Zm−1) has a multivariate Normal distribution with
means 0, variances 1 and common pairwise correlations 1/2 [9].
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 18/2
19. Lower bounds
0 200 400 600 800 1000
0.00.20.40.60.81.0
number m of alternatives
ProbCorrectSelection
N=50
N=100
N=150
N=250
N=300
delta=0.1
0 200 400 600 800 1000
0.00.20.40.60.81.0
number m of alternatives
ProbCorrectSelection
N=50
N=300
delta=0.07
Lower bound on the probability of correct selection vs. the number of
alternative subsets for N = {50, 100, 150, 200, 250}
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 19/2
20. Our proposal
• We propose an original blocking strategy for improving feature selection which
aggregates in a paired way the validation outcomes of several learning algorithms
to assess a gene subset and compare it to others.
• This is new with respect to conventional wrappers which commonly adopt only
one learning algorithm to evaluate the relevance of a given set of variables.
• The rationale of the approach is that by increasing the amount of experimental
conditions under which we validate a feature subset, we can lessen the problems
related to the scarcity of samples and consequently come up with a better
selection.
• The idea is quite simple: if we want to search for the best subset of features, this
subset should appear among the best ones whatever the learning algorithm is.
So far, the methods in feature selection end up with one of these two ways: either
avoid any learning algorithm (as in the case of filters), either restrict themselves
to consider only a specific one (as in conventional wrapper). What we advocate
here is that by using more than one algorithm in a feature selection task we may
improve the robustness of the selection procedure.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 20/2
21. Datasets
Name Reference Platform N n C
Golub [5] Affy hu6800 72 7129 2
SRBCT [8] cDNA 63 2308 4
ALL [22] Affy hgu95av2 248 12558 6
Hedenfalk [7] cDNA 22 3226 3
Alon [1] Affy hum6000 62 2000 2
Notterman [11] Affy hu6800 36 7457 2
West [21] Affy hu6800 49 7129 4
9Tumors [17] Affy hu6800 60 7129 9
11Tumors [18] Affy hgu95av2 174 12533 11
14Tumors [14] Affy hu35ksubacdf 308 15009 26
LungCancer [3] Affy hgu95av2 203 12600 5
BrainTumor1 [13] Affy hu6800 60 7129 5
BrainTumor2 [12] Affy hgu95av2 50 12625 4
DLBCL [15] Affy hu6800 77 7129 2
Leukemia2 [2] Affy hgu95a 72 12582 3
Prostate [16] Affy hgu95av2 102 12600 2
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 21/2
22. Experimental session
• A first dimensionality reduction step is carried out by ranking the variables
according to an univariate classification and by selecting the first 1000 gene
probes.
• A three-fold cross-validation strategy is used to measure the generalization
accuracy of the two competitive feature selection strategies. Note that unlike
conventional cross-validation we use here only one third of the samples for
training and two third of the samples for test.
• Given a third of the samples, we run both a conventional forward selection with a
given learning algorithm and an enhanced forward selection based on K = 6
learning algorithms implemented by the R statistical software. The six algorithms
are: a Naive Bayes (NB), a Support Vector Machine with a radial Gaussian kernel
(SVM), a Support Vector Machine with a linear kernel (SVML), a Nearest
Neighbour with one neighbor (NN1), a Nearest Neighbor with 5 neighbors (NN5)
and a Nearest Neighbor with 10 neighbors (NN10).
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 22/2
24. Some considerations
• Bioinformatics applications are known to be characterized by highly noisy data.
• The huge size of the feature space compared to the number of samples makes
hard the problem in terms of:
Optimization techniques to explore the feature space.
Large variance of the resulting model.
• Biologists asks for prediction accuracy but mainly for causual intepretation (gene
signature).
• Biologists are scared of unstable feature selection procedures which change the
sets of relevant genes simply by adding more observations. Examples are clinical
study with different populations of patients.
• Filtering techniques are computational efficient and robust against overfitting.
They may introduce bias but may have considerably less variance.
• Literature have been inflationed by over-optimistic results.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 24/2
25. Closing remarks
• Let us not forget that any learner is an estimator and as such any
outcome it returns, is a random variable.
• If the outcome of our feature selection technique is a set of
variables, this set is also a random set.
• Data miners are used to return confidence interval on accuracy.
• They should start returning confidence intervals also on feature
(gene) subsets.
On the use of feature selection to deal with the curse of dimensionality in microarray data – p. 25/2
26. References
[1] U. Alon, N. Barkai, D.A. Notterman, K. Gish, S. Ybarra, D. Mack,
and A.J. Levine. Broad patterns of gene expression revealed by
clustering analysis of tumor and normal colon tissues probed by
oligonucleotide arrays. Proc. of the National Academy of Sci-
ences, 96(10):6745–6750, 1999.
[2] S.A. Armstrong, J.E. Staunton, L.B. Silverman, R. Pieters, M.L.
den Boer, M.D. Minden, S.E. Sallan, E.S. Lander, T.R. Golub,
and S.J. Korsmeyer. Mll translocations specify a distinct gene
expression profile that distinguishes a unique leukemia. Nature
Genetics, 30(1):41–47, 2002.
[3] A. Bhattacharjee, W.G. Richards, J. Staunton, C. Li, S. Monti,
P. Vasa, C. Ladd, J. Beheshti, R. Bueno, M. Gillette, M. Loda,
G. Weber, E.J. Mark, E.S. Lander, W. Wong, B.E. Johnson, T.R.
Golub, D.J. Sugarbaker, and M. Meyerson. Classification of hu-
man lung carcinomas by mrna expression profiling reveals dis-
tinct adenocarcinoma subclasses. Proc Natl Acad Sci U S A.,
98(24):13790–13795, 2001.
[4] H.J. Bussemaker, H. Li, and E. D. Siggia. Regulatory element
detection using correlation with expression. Nature Genetics,
27:167–171, 2001.
[5] T. R. Golub, D. K. Slonin, P. Tamayo, C. Huard, and M. Gaasen-
beek. Molecular clssification of cancer: Class discovery and
class prediction by gene expression monitoring. Science,
286:531–537, 1999.
25-1
27. [6] I. Guyon and A. Elisseeff. An introduction to variable and feature
selection. Journal of Machine Learning Research, 3:1157–1182,
2003.
[7] I. Hedenfalk, D. Duggan, Y. Chen, M. Radmacher, M. Bittner,
R. Simon, P. Meltzer, B. Gusterson, M. Esteller, O.P. Kallioniemi,
B. Wilfond, A. Borg, and J. Trent. Gene expression profiles in
hereditary breast cancer. New Engl. Jour. Medicine, 344(8):539–
548, 2001.
[8] J. Khan, J. S. Wei, and M. Ringner. Clasification and diagnostic
prediction of cancers using gene expression profiling and artifi-
cial neural networks. Nature Medicine, 7(6):673–679, 2001.
[9] S. H. Kim and B. L. Nelson. Handbooks in Operations Research
and Management Science, chapter Selecting the Best System.
Elsevier, 2005.
[10] A. J. Kleywegt, A. Shapiro, and T. Homem de Mello. The sample
average approximation method for stochastic discrete optimiza-
tion. SIAM Journal of Optimization, 12:479–502, 2001.
[11] D.A. Notterman, U. Alon, A.J. Sierk, and A.J. Levine. Transcrip-
tional gene expression profiles of colorectal adenoma, adenocar-
cinoma and normal tissue examined by oligonucleotide arrays.
Cancer Research, 6:3124–3130, 2001.
[12] C. L. Nutt, D.R. Mani, R.A. Betensky, P. Tamayo, J.G. Cairncross,
C. Ladd, U. Pohl, C. Hartmann, M.E. McLaughlin, T.T. Batche-
lor, P.M. Black, A. von Deimling, S.L. Pomeroy, T.R. Golub, and
D.N. Louis. Gene expression-based classification of malignant
25-2
28. gliomas correlates better with survival than histological classifi-
cation. Cancer Research, 63(7):1602–1607, 2003.
[13] S. L. Pomeroy, P. Tamayo, M. Gaasenbeek, L.M. Sturla, Angelo
M, McLaughlin ME, Kim JY, Goumnerova LC, Black PM, Lau C,
Allen JC, Zagzag D, Olson JM, Curran T, Wetmore C, Biegel JA,
Poggio T, Mukherjee S, Rifkin R, Califano A, Stolovitzky G, Louis
DN, Mesirov JP, Lander ES, and Golub TR. Prediction of cen-
tral nervous system embryonal tumour outcome based on gene
expression. Nature, 415(6870):436–442, 2002.
[14] S. Ramaswamy, P. Tamayo, R. Rifkin, S. Mukherjee, C.H. Yeang,
M. Angelo, C. Ladd, M. Reich, E. Latulippe, J.P. Mesirov, T. Pog-
gio, W. Gerald, M. Loda, E.S. Lander, and T.R. Golub. Multiclass
cancer diagnosis using tumor gene expression signatures. Proc
Natl Acad Sci U S A, 98(26):15149–15154, 2001.
[15] M.A. Shipp, K.N. Ross, P. Tamayo, A.P. Weng, J.L. Kutok, R.C.
Aguiar, M. Gaasenbeek, M. Angelo, M. Reich, G.S. Pinkus, T.S.
Ray, M.A. Koval, K.W. Last, A. Norton, T.A. Lister, J. Mesirov
andD.S. Neuberg, E.S. Lander, J.C. Aster, and T.R. Golub.
Diffuse large b-cell lymphoma outcome prediction by gene-
expression profiling and supervised machine learning. Nature
Medicine, 8(1):68–74, 2002.
[16] D. Singh, P.G. Febbo, K. Ross, D.G. Jackson, J. Manola, C. Ladd,
P. Tamayo, A.A. Renshaw, A.V. D’Amico, J.P. Richie, E.S. Lander,
M. Loda, P.W. Kantoff, T.R. Golub, and W.R. Sellers. Gene ex-
pression correlates of clinical prostate cancer behavior. Cancer
Cell., 1(2):203–209, 2002.
25-3
29. [17] J.E. Staunton, D.K. Slonim, H.A. Coller, P. Tamayo, M.J Angelo,
J. Park, U. Scherf, J.K. Lee, W.O. Reinhold, J.N. Weinstein, J.P.
Mesirov, E.S. Lander, and T.R. Golub. Chemosensitivity pre-
diction by transcriptional profiling. Proc Natl Acad Sci U S A.,
98(19):10787–10792, 2001.
[18] A.I. Su, J.B. Welsh, L.M. Sapinoso, S.G. Kern, P. Dimitrov,
H. Lapp, P.G. Schultz, S.M. Powell, C.A. Moskaluk, H.F. Frier-
son Jr, and G. M. Hampton G. Molecular classification of human
carcinomas by use of gene expression signatures. Cancer Re-
search, 61(20):7388–7393, 2001.
[19] S. L.M. van der Keles and M. B. Eisen. Identification of regula-
tory elements using a feature selection method. Bioinformatics,
18:1167–1175, 2002.
[20] L. J. van’t Veer, H. Dai, and M. J. van de Vijver. Gene expres-
sion profiling predicts clinical outcome of breast cancer. Nature,
415:530–536, 2002.
[21] M. West, C. Blanchette, H. Dressman, E. Huang, S. Ishida,
R. Spang, H. Zuzan, J. R. Marks, and J. R. Nevins. Predicting the
clinical status of human breast cancer by using gene expression
profiles. Proc. Nat. Acad. Sci., 98(20):11462–11467, 2001.
[22] E.J. Yeoh, M.E. Ross, S.A. Shurtleff, W.K. Williams, D. Patel,
R. Mahfouz, F.G. Behm, S.C. Raimondi, M.V. Relling, and A. Pa-
tel. Classification, subtype discovery, and prediction of outcome
in pediatric lymphoblastic leukemia by gene expression profiling.
Cancer Cell, 1:133–143, 2002.
25-4