J. Biomedical Science and Engineering, 2010, 3, 59-64
doi:10.4236/jbise.2010.31009 Published Online January 2010 (http://www.SciRP.org/journal/jbise/
JBiSE
).
Published Online January 2010 in SciRes. http://www.scirp.org/journal/jbise
Identifying predictive markers of chemosensitivity of breast
cancer with random forests
Wei Hu
Department of Computer Science, Houghton College, Houghton, NY, USA.
Received 15 September 2009; revised 10 October 2009; accepted 20 October 2009.
ABSTRACT
Several gene signatures have been identified to build
predictors of chemosensitivity for breast cancer. It is
crucial to understand how each gene in a signature
contributes to the prediction, i.e., to make the predic-
tion model interpretable instead of using it as a black
box. We utilized Random Forests (RFs) to build two
interpretable predictors of pathologic complete re-
sponse (pCR) based on two gene signatures. One sig-
nature consisted of the top 31 probe sets (27 genes)
differentially expressed between pCR and residual
disease (RD) chosen from a previous study, and the
other consisted of the genes involved in Notch sin-
gling pathway (113 genes). Both predictors had a
higher accuracy (82% v 76% & 79% v 76%), a
higher specificity (91% v 71% & 98% v 71%), and a
higher positive predictive value (PPV) (68% v 52% &
73% v 52%)) than the predictor in the previous study.
Furthermore, Random Forests were employed to
calculate the importance of each gene in the two sig-
natures. Findings of our functional annotation sug-
gested that the important genes identified by the fea-
ture selection scheme of Random Forests are of bio-
logical significance.
Keywords: Random Forests; Breast Cancer; Chemosen-
sitivity; Gene Signature; Notch Signaling Pathway;
Pathologic Complete response; Predictor
1. INTRODUCTION
Breast cancer is a clinically heterogeneous disease that
demonstrates a wide variation in its clinical courses and
response to chemotherapy. This complexity is a reflec-
tion of the molecular oncogenic aberration in DNA re-
pair, cell cycle control, cell survival, and signal trans-
duction in breast tumors. Microarray analysis has identi-
fied breast cancer subtypes with distinct gene expression
profiles and clinical behavior [1,2,3]. There are several
major molecular classes of breast cancers indentified by
different research groups. Some studies [2,3] suggested
five major classes of breast cancer: normal breast-like,
luminal-A, luminal-B, basal-like, and human epidermal
growth factor receptor 2 (HER2)-positive cancers. An-
other study [4] proposed three major classes: ER+/
HER2-, ER-/HER2-, and HER2+. The heterogeneity of
breast cancer characterized by these subtypes brings
great challenge to its research. In a significant proportion
of breast cancer patients, chemotherapy does not result
in response, but can induce significant side effects and
financial costs. The ability to identify predictors of re-
sponse or resistance to cancer drugs will provide better
treatment to the individual patient.
Several studies have suggested that the gene-exp-
ression profiles of chemo sensitive tumors are different
from those of chemo resistant ones [5]. Gene expression
profiling with a measurement of thousands of mRNA
transcripts in a single experiment is widely used in hu-
man cancer research. Due to the high dimensionality of
microarray data, a feature selection step to find a subset
of discriminative genes, referred to as a signature, is
often necessary for building robust predictors [6,7].
Ayers et al. [8] developed a multigene predictor of
pCR to sequential weekly paclitaxel and FAC (T/FAC)
neoadjuvant chemotherapy for breast cancer patients.
The study involved 42 patients: 24 patients were used in
the training set and 18 patients in the validation set. pCR
was obtained in 13 patients (31%). A gene set of 74
markers (P < 0.09) was built using data from the training
set and tested on the validation set. Overall, a 78% pre-
dictive accuracy was achieved, with a 100% positive
predictive value for pCR, a 73% negative predictive
value, a sensitivity of 43%, and a specificity of 100%.
Later, a follow-up study [9] included 133 patients with
stage I-III breast cancer, with a pCR rate of 26% (n=34).
A 30-probe set Diagonal Linear Discriminant Analysis
(DLDA-30) classifier was selected for independent vali-
dation. It showed a significantly higher sensitivity (92%
v 61%) than a clinical predictor including age, grade,
and estrogen receptor status. This 30-probe set pharma-
cogenomic predictor correctly identified all but one of
the patients who achieved pCR (12 of 13 patients) and
all but one of those who were predicted to have residual
disease had residual cancer (27 of 28 patients).

W. Hu / J. Biomedical Science and Engineering 3 (2010) 59-64
SciRes Copyright © 2010 JBiSE
60
Chemosensitivity is better predicted by multigene
signatures than by a single molecular discrimination
because biological phenomena occur through the con-
certed expression of multiple genes [10,11,12]. However,
within a signature of genes, the important question of
how each individual gene contributes to the prediction
has not been studied. We attempted in this work to iden-
tify predictors and gene signatures that have better pre-
diction performance than the DLDA-30 and to quantify
the importance of each gene in a signature in the predic-
tion of pCR.
In [9], an exhaustive search of a good predictor of
pCR was conducted. Different machine learning tech-
niques were tested including support vector machines
with linear, radial, and polynomial kernels (SVM), Di-
agnal Linear Discriminant Analysis (DLDA), and K-
nearest neighbor (KNN) using Euclidean distance. One
interesting discovery was that SVM provided the worst
performance of pCR prediction among all these different
techniques in this particular data set. Random Forest has
demonstrated its comparable performances to SVM in
many bioinformatics applications. In the current study,
we sought to explore the utility of Random Forests were
utilized to construct two predictors based on two signa-
tures, the top 31 probe sets and the Notch signature, and
take advantage of the feature selection capability of Ran-
dom Forests to measure the importance of each gene in
these signatures.
2. MATERIALS AND METHODS
2.1. Patient Cohorts and Clinical Information
One breast cancer patient cohort was obtained from a
previous publication [9] (n=133). Needle-biopsy sam-
ples were collected from 133 patients with stage I, II, or
III breast cancer who received preoperative weekly pa-
clitaxel and a combination of fluorouracil, doxorubicin,
and cyclophosphamide (T/FAC). These 133 patients
were divided into two subsets, one training set of size 81
and one validation set of size 52. These data contain
clinical information including patient age, gender, race,
histological classification, stage, nuclear grade, ER (es-
trogen receptor), PR (progesterone receptor), and HER2
(human epidermal growth factor 2) status, pathologic
complete response, and residual disease. These data also
contain each patient’s genome-scale gene expression
profiles generated using Affymetrix U133A chip (Santa
Clara, CA). pCR was defined as no residual invasive
cancer in the breast or lymph nodes. pCR is presently
accepted as a reasonable early indicator for long-term
survival.
2.2. Top 31 Probe Set Signature
To build a predictor of pCR, the genes that are highly
expressed in either the pCR cases or the RD cases need
to be identified. To achieve this goal, t-tests for unequal
variances for all the probe sets on Affymetrix U133A
chip were carried out. The 31 probe sets (27 genes) with
the smallest t-test P values (FDR=0.05%) were selected in
[9], which was used as our first signature.
2.3. Notch Signature
Notch genes encode highly conserved cell surface re-
ceptors. The Notch signaling pathway consists of Notch
receptors, ligands, negative and positive modifiers, and
transcription factors. It plays a key role in the normal
development of many tissues and cell types, through
diverse effects on cell regulation, proliferation, and dif-
ferentiation. Aberrant Notch signaling has been observed
in several human cancers including acute T-cell lym-
phoblastic leukemia and cervical cancer. Recent evi-
dences implied that it might be associated with breast
cancer [13,14].
Selecting a gene signature based on differentially ex-
pressed genes between two conditions, such as pCR and
RD in our study, is a common strategy nowadays. Here
we endeavored to take a quite different approach, i.e., to
identify a signature of genes involved in a particular
pathway that has a key impact on human cancers.
The Oligo GEArray Human Notch Signaling Pathway
Microarray [15] was designed for profiling expression of
113 genes involved in Notch signaling. Our second sig-
nature was these 113 genes as shown in Table 1.
We were particularly interested in uncovering what
genes in these two signatures are important for the pre-
diction of pCR and what biological or medical signifi-
cance they might have.
2.4. False Discovery Rate
The standard P value was designed for testing individual
hypotheses. When applied in a multiple testing problem
such as selecting informative genes in microarray data, it
may result in many false positives. While there are a
number of methods to overcome the problems due to mul-
tiple testing, False Discovery Rate (FDR) approach [16,17]
was used to help select the top 31 probe sets in [9].
2.5. Random Forests
Random Forest, proposed by Leo Breiman in 1999 [18],
is an ensemble classifier based on many decision trees.
Each tree is built on a bootstrap sample from the original
training set. The variables used for splitting the tree
nodes are a random subset of the whole variable set. The
classification decision of a new instance is made by ma-
jority voting over all trees. About one-third of the in-
stances are left of the bootstrap sample and not used in
the construction of the tree. These instances in the train-
ing set are called “out-of-bag” instances and are used to
evaluate the performance of the classifier.

W. Hu / J. Biomedical Science and Engineering 3 (2010) 59-64
SciRes Copyright © 2010 JBiSE
61
Table 1. Genes in notch signaling related pathways [15].
Notch Signaling Pathway:
Notch Binding: DLL1 (DELTA1), DTX1, JAG1, JAG2.
Notch Receptor Processing: ADAM10, PSEN1, PSEN2, PSENEN (PEN2).
Notch Signaling Pathway Target Genes:
Apoptosis Genes: CDKN1A, CFLAR (CASH), IL2RA, NFKB1.
Cell Cycle Regulators: CCND1 (Cyclin D1), CDKN1A (P21), IL2RA.
Cell Proliferation: CDKN1A (P21), ERBB2, FOSL1, IL2RA.
Genes Regulating Cell Differentiation: DTX1, PPARG.
Neurogenesis: HES1, HEY1.
Regulation of Transcription: DTX1, FOS, FOSL1, HES1, HEY1, NFKB1, NFKB2,
NR4A2, PPARG, STAT6.
Other Target Genes with Unspecified Functions: CD44, CHUK, IFNG, IL17B, KRT1,
LOR, MAP2K7, PDPK1, PTCRA.
Other Genes Involved in the Notch Signaling Pathway:
Apoptosis Genes: AXIN1, EP300, HDAC1, NOTCH2, PSEN1, PSEN2.
Cell Cycle Regulators: AXIN1, CCNE1, CDC16, EP300, FIGF, JAG2, NOTCH2,
PCAF.
Cell Proliferation: CDC16, FIGF, FZD3, JAG1, JAG2, LRP5, NOTCH2, PCAF, STIL
(SIL).
Genes Regulating Cell Differentiation: DLL1, JAG1, JAG2, NOTCH1, NOTCH2,
NOTCH3, NOTCH4, PAX5, SHH.
Neurogenesis: DLL1, EP300, HEYL, JAG1, NEURL, NOTCH2, PAX5, RFNG, ZIC2
(HPE5).
Regulation of Transcription: AES, CBL, CTNNB1, EP300, GLI1, HDAC1, HEYL,
HOXB4, HR, MYCL1, NCOR2, NOTCH1, NOTCH2, NOTCH3, NOTCH4, PAX5,
PCAF, POFUT1, RUNX1, SNW1 (SKIIP), SUFU, TEAD1, TLE1.
Others Genes with Unspecified Functions: ADAM17, GBP2, LFNG, LMO2, MFNG,
MMP7, NOTCH2NL, NUMB, SEL1L, SH2D1A.
Other Signaling Pathways that Crosstalk with the Notch Signaling Pathway:
Sonic Hedgehog (Shh) Pathway: GLI1, GSK3B, SHH, SMO, SUFU.
Wnt Receptor Signaling Pathway: AES, AXIN1, CTNNB1, FZD1, FZD2, FZD3,
FZD4, FZD6, FZD7, GSK3B, LRP5, TLE1, WISP1, WNT11.
Other Genes Involved in the Immune Response: CXCL9, FAS (TNFRSF6), G1P2,
GBP1, IFNG, IL2RA, IL2RG, IL4, IL4R, IL6ST, IRF1, ISGF3G, OAS1, OSM,
STAT5A, STUB1.
Table 2. Performance measures of three predictors: DLDA-30,
RF-31, and RF-Notch.
Measures DLDA-30 RF-31 RF-Notch
Accuracy 0.76 0.82 0.79
Sensitivity 0.92 0.55 0.27
Specificity 0.71 0.91 0.98
PPV 0.52 0.68 0.73
NPV 0.96 0.85 0.80
2.6. Feature Selection Using Random Forests
Random Forest calculates several measures of variable
importance. The mean decrease in accuracy measure
was used in [19] to rank the importance of the feature
in prediction. This measure is based on the decrease of
classification accuracy when values of a variable in a
node of a tree are permuted randomly. In this study, two
packages of R, randomForest and varSelRF [19], were to
compute the importance of the genes in a given signature.
3. RESULTS
The first predictor, RF-31, was based on the top 31
probe sets, and the second predictor, RF-Notch, was
based on the Notch signature. As in the case of DLDA-
30, the RF-31 and RF-Notch were trained on the training
data (n=82) and the accuracy of the two predictors was
tested on a separate validation set (n=51).
Random Forests produce non-deterministic outcomes.
To reduce the possible variance of our results, the Ran-
dom Forests algorithm was run multiple times and then
the average of the predictions was taken. The prediction
results of the RF-31 and RF-Notch were based on the
average of 20 repeated predictions, which are shown in
Tables 2. The importance of each gene in the two sig-
natures was based on the averaged calculations by using
the function randomVarImpsRF in varSelRF repeated 10
times, as shown in Figure 1 and Table 3.
The predictions of RF-31 and RF-Notch and the im-

W. Hu / J. Biomedical Science and Engineering 3 (2010) 59-64
SciRes Copyright © 2010 JBiSE
62
Table 3. 31 genes of the highest importance in Notch signature.
Importance Gene
Symbol Probe Set IDP Valuet-Test
Higher
Expression
in
0.000822 CTNNB
1 201533_at 0.46320 0.74521RD
0.000928 SNW1 201575_at 0.044582.07657RD
0.001295 NOTCH
2 202443_x_at 0.03601 -2.23736pCR
0.00185 NOTCH
2 202445_s_at0.02239-2.46026pCR
0.000658 HES1 203395_s_at 0.34906 -0.94767pCR
0.006243 ISGF3G 203882_at 0.01170-2.75847pCR
0.007946 CXCL9 203915_at 0.01268 -2.73059pCR
0.000725 MFNG 204152_s_at0.06783-1.92262pCR
0.000826 LMO2 204249_s_at 0.01955 -2.44963pCR
0.002534 IL6ST 204863_s_at0.011412.66874RD
0.003382 NEURL 204889_s_at 8.33E-05 4.15928RD
0.007544 ADAM1
7 205746_s_at0.00049-4.00315pCR
0.002078 IL2RA 206341_at 0.10542 -1.67547pCR
0.001054 RUNX1 208129_x_at0.632120.48313RD
0.002165 NCOR2 208889_s_at 0.07027 1.85101RD
0.003526 NUMB 209073_s_at0.064761.88058RD
0.000775 MAP2K
7 209952_s_at 0.05390 -1.98674pCR
0.001136 ERBB2 210930_s_at0.03134-2.28195pCR
0.001296 RUNX1 211180_x_at0.19439-1.3392 pCR
0.00223 RUNX1 211181_x_at0.000203.93243RD
0.000769 PSEN2 211373_s_at 0.87130 0.16266RD
0.003479 NOTCH
2 212377_s_at0.000193.98301RD
0.000862 AXIN1 212849_at 0.00131 3.35556RD
0.001427 MYCL1 214058_at 0.01948-2.47946pCR
0.005243 CFLAR 214618_at 0.00984 -2.76622pCR
0.000663 CD44 216056_at 0.22213-1.24048pCR
0.000748 MAP2K
7 216206_x_at 0.43559 0.78876RD
0.001933 CFLAR 217654_at 0.05371-2.02278pCR
0.000858 FZD4 218665_at 0.00180 3.25207RD
0.001928 IL17B 220273_at 0.08425-1.76439pCR
portance of the genes in the two signatures are summa-
rized in Table 2 and Figure 1. The metrics of perform-
ance in Table 2 indicate the different strengths of the
DLDA-30 and our RF-31 and RF-Notch. Both RF-31
and RF-Notch had a higher accuracy, a higher specificity,
and a higher PPV than the DLDA-30.
3.1. Importance of the Genes in Top 31 Probe
Sets
Of these 31 probe sets, five probe sets had a higher ex-
pression value in the pCR cases and 26 probe sets had a
higher expression in the RD cases, demonstrating the
dominance of the highly expressed genes in the patients
with RD.
Figure 1 displays several genes of top importance, in-
cluding MAPT, BBS4, MGC5370, BTG3, MELK, CA12,
FGFR1OP, MTRN, FLJ10916, E2F3, RRM2, and KIF3A.
MAPT, microtubule associated protein tau, was discov-
ered as the best single gene discriminator of pCR to pre-
operative chemotherapy with Paclitaxel, 5-Fluoroutacil,

W. Hu / J. Biomedical Science and Engineering 3 (2010) 59-64
SciRes Copyright © 2010 JBiSE
63
Figure 1. Two plots of the importance of the genes in the top 31 probe set signature and
the Notch signature respectively.
Doxorubicin [20]. Its expression correlates closely with
ER expression in human breast cancer. In the top 31
probe set signature, there were four probe sets of gene
MAPT with very high t-test statistic. The multiple selec-
tion of MAPT in the signature demonstrates its signifi-
cance. BBS4, Bardet-Biedl syndrome 4, is a member of
the Bardet-Biedl syndrome (BBS) gene family associ-
ated with the Bardet-Biedl syndrome. MGC5370 is an
alias name of gene MDM2, which is a target gene of the
transcription factor tumor protein p53. Over expression of
this gene can result in excessive inactivation of tumor
protein p53, diminishing its tumor suppressor function.
This gene had a very high expression value in our patient
with RD. BTG3, a member of the BTG/Tob family, is a
transcriptional target of p53. It has a role in DNA damage
response. Its antiproliferative action through inhibition of
another gene E2F1 was discovered recently [21]. This
gene was highly expressed in our patients with pCR,
which is consistent with this recent discovery. There were
two probe sets of gene BTG3 in the top 31 probe sets,
which further illustrates this gene’s significance. MELK,
maternal embryonic leucine zipper kinase, is a potential
marker of proliferating mammary epithelial progenitor
cells that are highly expressed in multiple human cancers,
including human breast cancer. CA12, Carbonate dehy-
dratase XII, is a member of a large family of zinc metal-
loenzymes that participate in various biological proc-
esses, and was found to be overexpressed in 10% of
clear cell renal carcinomas. FGFR1OP is Fibroblast
Growth Factor Receptor 1 (FGFR1) Oncogene Partner.
Fusing this gene and the FGFR1 gene has been found in
cases of myeloproliferative disorder. This gene plays an
important role in normal proliferation and differentiation
of the erythroid lineage. MTRN, Meteorin, glial cell
differentiation regulator, is a gene clearly involved in
cell differentiation. FLJ10916 is an alias name of gene
THNSL2, threonine synthase-like 2, which functions in
lyase activity, pyridoxal phosphate binding, and meta-
bolic process. E2F3, E2F transcription factor 3, is a
member of the E2F family of transcription factors. The
E2F family is essential in the control of cell cycle and
action of tumor suppressor proteins. RRM2, Ribonucleo-
tide reductase M2 polypeptide, provides the precursors
necessary for DNA synthesis. During mitosis, Kinesin
family member 3A (KIF3A) has a critical function in the
equal segregation of chromosomes between two daugh-
ter cells.
Based on the above functional annotation, it is evident
that these top important genes are not only vital in the
prediction of pCR but also strongly implicated in tu-
mourigenesis.
3.2. Importance of the Genes in Notch
Signature
Of the 31 genes with highest importance values in Notch
signature, 17 probe sets had a higher expression value in
the pCR cases and 14 probe sets had a higher expression
value in the RD cases as seen in Table 3. This somewhat
even distribution of the probe sets between the pCR and
RD cases was in contrast to the top 31 probe set signature,
which could be attributed to the functions of Notch sig-
naling pathway.
The top important genes in Notch signature were the
following: CXCL9, ADAM17, ISGF3G, CFLAR, NUMB,

W. Hu / J. Biomedical Science and Engineering 3 (2010) 59-64
SciRes Copyright © 2010 JBiSE
64
NOTCH2, NEURL, IL6ST, and RUNX1. NOTCH2 is a
multifunctional gene involved in apoptosis, cell prolif-
eration, cell differentiation, neurogenesis, and regulation
of transcription. There are three probe sets of NOTCH2
and two probe sets of CFLAR in Figure 1, reflecting these
genes’ importance. CXCL9, ISGF3G, and IL6ST are all
involved in immune response. Since the functions of
these genes are illustrated through their pathways in Ta-
ble 1, we will not elaborate on them any further here.
There were 15 genes in the top 31 probe set signature
with importance values above 0.003, and there were
seven such genes in Notch signature. This was expected.
Because of their high t-test statistics, the top 31 probe
sets should be more sensitive to the random permutation
employed in the importance calculation than those in the
Notch signature. Nonetheless, in Figure 1 the Notch
signature genes displayed their significance.
4. CONLUSIONS
Random Forests were employed to study the prediction
of pathologic complete response in breast cancer, and the
results improved the predictions of the DLDA-30. Func-
tional annotation demonstrated that the important genes
identified by the feature selection scheme of Random
Forests are of biological significance.
5. ACKNOWLEDGMENT
We thank Houghton College for its financial support.
REFERENCES
[1] Nguyen, P. L., Taghian, A. G. et al. (2008) Breast cancer
subtype approximated by estrogen receptor, progesterone
receptor, and HER-2 is associated with local and distant
recurrence after breast-conserving therapy, J. Clin Oncol.,
26(14), 2373-8.
[2] Perou, C. M., Sorlie, T., et al. (2000)Molecular portraits
of human breast tumours, Nature, 406(6797), 747-752.
[3] Sorlie, T., Perou, C. M. et al. (2001) Gene expression
patterns of breast carcinomas distinguish tumor sub-
classes with clinical implications, Proc Natl Acad Sci U S
A, 98(19), 10869-10874.
[4] Kapp, A. V., Jeffrey, S. S. et al. (2006) Discovery and
validation of breast cancer subtypes. BMC Genomics, 7,
231.
[5] Pusztai, L., (2008) Current status of prognostic profiling
in breast cancer, The Oncologist, 13, 350-360.
[6] van ’t Veer, L. J., Dai, H., et al. (2002) Gene expression
profiling predicts clinical outcome of breast cancer. Na-
ture, 415, 530-536.
[7] van de Vijver, M. J. et al. (2002) A gene-expression sig-
nature as a predictor of survival in breast cancer. N Engl
J Med, 347, 1999-2009.
[8] Ayers, M., Symmans, W. F., Stec, J. et al. (2004) Gene
expression profiles predict complete pathologic response
to neoadjuvant paclitaxel/FAC chemotherapy in breast
cancer. J Clin Oncol, 22, 2284-2293.
[9] Hess, K. R., Anderson, K., Symmans, W. F. et al., (2006)
Pharmacogenomic predictor of sensitivity to preoperative
chemotherapy with paclitaxel and fluorouracil, doxoru-
bicin, and cyclophosphamide in breast cancer. J Clin
Oncol, 24, 4236-4244.
[10] Brenton, J. D., Carey L. A., Ahmed, A. A. et al. (2005)
Molecular classification and molecular forecasting of
breast cancer: ready for clinical application? J Clin On-
col, 23, 7350-7360.
[11] Wang, Y., Klijn, J. G. et al. (2005) Gene-expression pro-
files to predict distant metastasis of lymph-node-negative
primary breast cancer, Lancet, 365, 671-679.
[12] Ma, X. J., Wang, Z., et al. (2004) A two-gene expression
ratio predicts clinical outcome in breast cancer patients
treated with tamoxifen. Cancer Cell, 5(6), 607-16.
[13] Brennan, K. and Anthony Brown, M. C. (2003) Is there a
role for Notch signalling in human breast cancer? Breast
Cancer Res, 5(2), 69-75.
[14] Stylianou, S., Clarke, R. B. et al. (2006) Activation of
notch signaling in human breast cancer, Cancer Research,
66, 1517-1525.
[15] GEArray, O. Human notch signaling pathway microarray.
http://www.sabiosciences.com/gene_array_product/HTM
L/OHS-059.html
[16] Benjamini, Y. and Hochberg, Y. (1995) Controlling the
false discovery rate: a practical and powerful approach to
multiple testing. J. R. Stat. Soc. Ser. B Stat. Methodol, 57,
289-300.
[17] Pounds, S. and Morris, S. W. (2003) Estimating the oc-
currence of false positive and false negatives in microar-
ray studies by approximating and partitioning the em-
pirical distribution of p-values. Bioinformatics 19, 1236-
1242.
[18] Breiman, L. and Random, F. (2001) Machine Learning,
45 (1), 5-32.
[19] Díaz-Uriarte, R. and Alvarez de Andrés, S. (2006) Gene
selection and classification of microarray data using
random forest. BMC Bioinformatics, 7(3).
[20] Rouzier, R., Rajan, R., et al. (2005) Microtubule associ-
ated protein tau is a predictive marker and modulator of
response to paclitaxel-containing preoperative chemo-
therapy in breast cancer. Proc Natl Acad Sci U S A, 102,
8315-8320.
[21] Ou, Y. H., Chung, P. H., et al. (2007)The candidate tumor
suppressor BTG3 is a transcriptional target of p53 that
inhibits E2F1, The EMBO Journal, 26, 3968-3980.