Combining gene expression patterns and clinical parameters for risk stratification
in medulloblastoma
Fernandez-Teijeiro Ana MD PhD (4), Betensky Rebecca A. PhD(5), Sturla Lisa M. PhD(1), Kim John YH.
MD PhD(1)(3), Tamayo Pablo PhD(6), Golub Todd R. MD(2)(3)(6) and Pomeroy Scott L. MD PhD(1)
(1)Division of Neuroscience, Department of Neurology, (2) Department of Medicine, Children´s Hospital;
and (3)Department of Pediatric Oncology, Dana-Farber Cancer Institute, Harvard Medical School, Boston,
Massachusetts, USA.
(4)Unidad de Oncologia Pediatrica, Hospital de Cruces-Baracaldo, Basque Country, Spain.
(5)Department of Biostatistics, Harvard School of Public Health, Boston, Massachusetts, USA.
(6)Whitehead Institue/MIT Center of Biomedical Research, Massachusetts Institute of Technology,
Cambridge, Massachusetts, USA.
Correspondence and reprints:
Scott L.Pomeroy
300 Longwood Avenue Enders 230
Boston MA 02115 (USA)
e-mail: scott.pomeroy@tch.harvard.edu
Aknowledgements:
This work was supported in part by NIH grant NS35701 and NIH-supported Mental Retardation Research
Center (SLP), Fulbright Scholar Program, Foundation of Spanish Society of Pediatric Oncology (SEOP)
and a grant from the Department of Health of Basque Gouvernement (Spain) (AFT).
1
ABSTRACT
Purpose: Stratification of risk in patients with medulloblastoma remains a challenge. As clinical parameters
have been proven insufficient for accurately defining disease risk, molecular markers have become the
focus of interest. Outcome predictions based on microarray gene expression profiles have been the most
accurate to date. Here, we ask in a multivariate model whether clinical parameters enhance the prediction
of survival in the presence of gene expression profiles.
Patients and Methods: In a cohort of 60 patients whose medulloblastoma samples have been previously
analyzed for gene expression profile1, associations between clinical and gene expression variables and
survival were assessed using Cox proportional hazard models. Clinical variables available included age,
stage (i.e., the presence of disseminated disease at diagnosis), sex, histological subtype, treatment and
status.
Results: Univariate analysis demonstrated expression profiles to be highly predictive of survival
(p<0.0001), but the only significant clinical prognostic factor was the presence of disseminated disease at
diagnosis (p=0.04). Multivariate analysis suggested that stage, sex, and age are significantly predictive of
survival, even after adjusting for gene expression profile. Further, an exploratory analysis noted a trend
for decreased survival of patients with metastases at diagnosis but favorable gene expression profile.
Conclusion: Although gene profiling alone seems to be the more accurate basis for targeting therapy in
children with this disease, the clinical variables of stage, age, and sex appear to contribute information, as
well. These results need to be validated in a larger prospective study.
Key words: medulloblastoma, risk stratification, gene expression pattern, microarrays
Word count: 240
2
Microarray-based gene expression profiling has opened a wide and promising research field for diagnosis,
classification and treatment of malignant neoplasms 2,
3.
Recently, application of cDNA microarrays to a
subset of 60 similarly treated patients from whom biopsies were obtained before treatment allowed to
generate a classifier capable of predicting clinical outcome of this tumor based on gene expression
profiles1. It is not known whether clinical parameters may add to gene expression prediction or whether it
predicts independently.
Medulloblastomas are the most common malignant brain neoplasm in children. Current management
strategies combining surgery, radiotherapy and different chemotherapy regimens account for survival
rates of 50-70% but unfortunately, postreatment cognitive deficits are common in those who do survive
that long4,5,6,7. One of the challenges of management strategies in medulloblastoma is trying to
differentiate “high risk” as opposed to “low risk” patients in order to enable more precise therapeutic
intervention, tailored to the degree of biological aggressiveness. So, patients with more favorable disease
could be cured with reduced therapy while limiting neuro-cognitive and endocrine sequelae.
So far, few independent prognostic indicators have been identified to accomplish this goal. Age less than
3 years or 1.5 cm2 of residual disease after surgery8 and/or evidence of metastasis based on the Chang
system9 are clinical characteristics currently used to define "high risk" patients. Recent reports6,10 indicate
that clinical variables are an insufficient means of accurately defining disease risk. High levels of
expression of the neurotrophin-3 receptor (TRKC) have correlated with a favorable outcome 11,12,13 while cmyc expression14 and high Erb2 receptor expression and isolated 17p loss could identify a sub-population
of high-risk patients15. Histopathologic grading has been documented of clinical significance as well 16.
Recently, multiple gene outcome classifiers based on microarrays have been found to provide the most
accurate prediction to date1, although confirmation in an independent dataset is still required. The aim of
this paper is to investigate whether gene profiling outcome predictions are independent of clinical
parameters and if combining clinical and molecular factors increases the accuracy of disease risk
stratification above that afforded by gene expression analysis or clinical stage alone.
3
MATERIAL AND METHODS
Patients
This retrospective study was conducted using the dataset of 60 newly diagnosed patients belonging to
eight different institutions whose medulloblastoma tissue specimens had been previously investigated for
gene profiling expression analysis1. Thirty-five patients were part of a cohort described in previous
publications as well11,12. Available variables for the 60 patients were sex, age, stage, histological subtype,
chemotherapy, follow-up and status when the study was closed. Clinical details of the 60 patients are
summarized in Table 1. The dataset included 39 boys (65%) and 21 girls (35%). The median age was 6
years (range, 0.6-38.2 years). Ten patients (17%) were under 3 years and 5 patients (8%) were over 18
years.
The histologic diagnosis of medulloblastoma was confirmed according to WHO criteria17: 46 (77%) cases
fulfilled criteria for classic medulloblastoma while 14 (23%) had the desmoplastic variant. The distribution
of M stage, as defined by Chang9, was as follow: M0 42, M1 5, M2 2, M3 10 and M4 1. In view to the small
numbers analysis were limited to two stage groups: 42 patients (70%) were M0 while the remaining 18
patients (30%) were considered M+. All patients were treated with radiotherapy and chemotherapy.
Craniospinal irradiation reached to 2400 - 3600 centiGray (cGy) with a tumor dose of 5300 - 7200 cGy.
Chemotherapy consisted of cisplatin and vincristine, and combinations of carboplatin, etoposide,
cyclophosphamide, procarbazine or lomustine (CCNU). The majority of patients were treated with
polichemotherapy with regimens combining three or four drugs (see above). Two patients received high
dose chemotherapy at relapse, including methotrexate and thiotepa, followed by autologous bone marrow
transplantation. The 60 patients included were observed for a median of 3.5 years (range, 5 months to 11
years) from date of diagnosis to last contact or death. With 21 dead patients at the close of the study,
median follow-up for the 39 patients was 4.8 years (range, 2 to 11 years): thus patient death shortened
follow-up time.
4
Outcome gene profiling model
Gene profiling expression classification was based in the 8-gene model, established through k-nearest
neighbors (k-NN) algorithm 18, previously described1 The two resulting groups of tumors exhibit clearly
different prognoses. In the poor outcome group, the 5 year survival rate is 17% (median survival is 1.8
years), while in the favorable outcome group, the 5 year survival rate is 82% (median survival is 8.5
years).
Statistical analysis
Wilcoxon Rank Sum tests were used to compare age at diagnosis among groups defined by categorical
variables. Relationships among categorical markers were assessed using Fisher’s exact tests. Survival
distributions were estimated using Kaplan-Meier curves19. Univariate and multivariate analyses were
conducted using Cox proportional hazard regression models. For each of these tests, the significance
level was taken to be 0.05, and all p-values were two-sided. The software package StatView (version
5.0.1; SAS Institute, Cary, NC) was used for these analyses. Exploratory classification tree analyses20
were conducted using the tssa software for S-plus available at the Statlib website at Carnegie Mellon
University (http://lib.stat.cmu.edu/).
RESULTS
No significant associations between age at diagnosis and sex, stage, subtype, or gene expression profile
were observed. In contingency table analysis no significant associations among variables could be
established. In particular, sex, stage and subtype were not significantly associated with gene profiling
expression pattern.
Univariate survival analysis
Table 1 lists the p-values for the comparisons of interest. Among the clinical variables analyzed, only
metastatic disease stage at diagnosis was significantly associated with survival, with hazard ratio of 0.39
for M0 versus M+ (p=0.04) (Figure 1A), confirming its well established prognostic significance in
medulloblastoma5. In contrast, neither histologic subtype (p=0.4), sex (p=0.8), nor age at diagnosis
5
(p=0.3) was found to be significantly associated with survival (Figure 1B and 1C). Since there are only 5
patients that were older than 18 at diagnosis, and since their treatment may have differed from that of the
younger patients, we re-analyzed the data without them. When patients over 18 years are excluded, only
gene expression group is significantly predictive of survival. However, as 4 of these 5 patients died,
relative to a total of 21 deaths among all patients, they comprise roughly 20% of the information in the
entire dataset. Thus, their removal entails a substantial decrease in information, and thus statistical power.
Multivariate regression analysis
The principal aim of this study was to establish whether clinical stratification provides prognostic
information for patients with medulloblastoma additional to that afforded by gene profiling expression
analysis. We used backward elimination, with a threshold of 0.10, to select a multivariate model. In
addition to gene expression group, stage, sex, and age remained in the model. The associated hazard
ratios, confidence intervals, and p-values are listed in Table 2. According to this model, the hazard for
death for patients in the high risk gene expression group is twelve times higher than that for patients in the
low risk gene expression group, the hazard for males is 2.8 times that for females, the hazard for patients
with non-metastatic disease (stage M0) is 44% that for patients with metastatic disease, and the relative
increase in the hazard associated with each additional year of life at diagnosis is 8%. Neither subtype nor
chemotherapy group was a significant prognostic factor. Further, we considered several possible cutpoints
for age, but found none of them to be significantly predictive of survival. When the 5 patients over 18
years are excluded, only gene-profiling expression remains significant; this may be because they are truly
different from the other patients, or it may be due to a loss of power.
Survival analysis following risk stratification
An exploratory classification tree analysis20 suggested that the clinical variable of stage might be important
among patients in the high-risk gene expression group. However, in a proportional hazards model, stage
did not significantly distinguish among patients within the high-risk gene group (p=0.10), perhaps because
of the small numbers. Nevertheless, as shown in figure 2, there was a non-significant trend for good risk
patients by gene expression profile but with M+ disease to have more early deaths than those with M0
6
disease. A larger dataset will be needed to determine whether M stage improves outcome predictions in
good risk patients by gene expression profiles.
DISCUSSION
The principal aim of our study was to investigate the hypothesis that the combined assessment of clinical
and molecular markers allows increased accuracy of disease-risk stratification for patients with
medulloblastoma over the use of either type of marker alone. We found that while gene expression profile
is a more powerful predictor of survival than the clinical parameters, some of these clinical parameters
(stage, sex and age) may be predictive of survival also, even in the presence of gene expression
information. In fact, it seems that gene profiling alone allows the most reliable identification of patients who
may be cured with reduced therapy although a subset of better prognosis patients might be more
accurately defined combining clinical and gene expression pattern stratification. In this regard, presence of
disseminated disease in those patients with low-risk gene expression pattern would remain valuable for
outcome prediction as a trend for better prognosis for those patients without metastatic disease is
observed. Our approach is the first attempt of fitting an outcome model including gene expression pattern
and clinical variables. We are currently seeking to confirm these results in a larger prospective study on
patients with medulloblastoma.
For the whole series the only clinical prognostic factor was the presence of disseminated disease at
diagnosis. Patients with no metastatic tumors have a significant better outcome that those with
disseminated disease, although the difference disappears when patients older than 18 years are
excluded. Tumor size, an important element of the T stage of Chang 9 has not been found to be a
prognostic factor, as previously reported.5,7,30,31 (data not shown).
Although in our series boys are clearly predominant, no difference in survival related to sex is observed in
univariate analysis. Previous studies of the effect of gender on clinical outcome in patients with
medulloblastoma reached various conclusions. In at least two large studies, the survival advantage for
girls was statistically significant21,22, whereas other studies showed either borderline significance or no
significance to the association of gender and outcome.57,16,23,24,25,26,27,28 However, in multivariate analysis,
7
sex was found to be a significant predictor, with boys having almost a 3-fold increase in the hazard of
death relative to girls.
Although desmoplastic medulloblastoma, meanly characterized for desmoplasia and marked neuronal
differentiation17, often follow a less aggressive course, in our series including only 14 patients with this
variant no difference in survival for both subtypes is observed.
Attending to age, 4 out of the 5 patients older than 18 years died, which represents nearly 20% of the total
deaths observed. Although a better outcome for adult medulloblastoma has been reported29, because of
the small number of patients it is not possible to establish any conclusion on this group of age. No
significant difference of survival is observed for the 10 patients under three years either, unlike other
recent series5.
Gene profiling expression analysis is just starting to show its utility for diagnosis and classification of
malignant neoplasms. It will be also important in future studies to analyze the interaction between gene
expression pattern and treatment response. Besides, gene expression pattern important in aggressive
tumor phenotype may also represent potential targets for novel therapeutically approaches. If these
results are confirmed with the 8-gene model in a larger dataset, future treatment trials for medulloblastoma
should be based on both gene expression pattern analyses through microarrays technique and clinical
variables in order to achieve the best definition of risk groups.
REFERENCES
1. Pomeroy SL, Tamayo P, Gaasenbeek M, et al: Prediction of central nervous system embryonal tumor
outcome based on gene expression.Nature 415:436-42,2002.
2. Golub TR, Slonim DK, Tamayo P, et al: Molecular classification of cancer: Class discovery and class
prediction by gene expression monitoring. Science 286:531-537,1999.
3. Ramaswamy S, Golub TR. DNA microarrays in Clinical Oncology. J Clin Oncol 20:1932-1941, 2002.
4. Bayley CC, Gnekow A, Wellek S, et al. Propective randomised trial of chemotherapy given before
radiotherapy in childhood meduloblastoma. International Society of Paediatric Oncology (SIOP) and
8
the German Society of Paediatric Oncology (GPO): SIOP 2. Med and Pediatr Oncol 25:166-178,
1995.
5. Zeltzer PM, Boyett JM, Finlay JL, et al. Metastasis stage, adjuvant treatment, and residual tumor are
prognostic factors for medulloblastoma in children: conclusions from the Children's Cancer Group 921
randomized phase III study. J Clin Oncol 17:832-45, 1999.
6. Packer RJ, Goldwein J, Nicholson HS, et al. Treatment of children with medulloblastomas with
reduced-dose craniospinal radiation therapy and adjuvant chemotherapy: A Children's Cancer Group
Study. J Clin Oncol 17:2127-2136, 1999.
7. Jenkin D, Shabanah MA, Shail EA, et al. Prognostic factors for medulloblastoma. Int J Radiation
Oncol Biol Phys 47:573-584, 2000.
8. Albright AL, Wisoff JH, Zeltzer PM, et al. Effects of medulloblastoma resections on outcome in
children: a report from the Children's Cancer Group. Neurosurgery 38:265-271,1996.
9. Chang CH, Housepian EM, Herbert C Jr. An operative staging system and a megavoltage
radiotherapeutic technic for cerebellar medulloblastomas. Radiology. 93:1351-1359, 1969.
10. Thomas PR, Deustch M, Kepner JL et al. Low-stage medulloblastoma. Final analysis of trial
compairing estándar-dose with reduced dose neuraxis irradiation. J Clin Oncol 18: 3004-3011, 2000.
11. Segal RA, Goumnerova LC, Kwon YK, et al. Expression of the neurotrophin receptor TrkC is linked to
a favorable outcome in medulloblastoma. Proc Natl Acad Sci U S A 91:12867-12871, 1994.
12. Kim JY, Sutton ME, Lu DJ, et al. Activation of neurotrophin-3 receptor TrkC induces apoptosis in
medulloblastomas. Cancer Res 59:711-719, 1999.
13. Grotzer MA, Janss AJ, Fung K, et al. TrkC expression predicts good clinical outcome in primitive
neuroectodermal brain tumors. J Clin Oncol 18:1027-35, 2000.
14. Herms J, Neidt I, Luscher B et al. C-myc expression in medulloblastoma and its prognostic value. Int J
Cancer 89:395-402, 2000.
15. Gilbertson R, Wickramasinghe C, Hernan R, et al. Clinical and molecular stratification of disease risk
in medulloblastoma. Br J Cancer 85(5):705-12, 2001.
16. Eberhart CG, Kepner JL, Goldthwaite PT, et al. Histopathologic grading of medulloblastomas: a
Pediatric Oncology Group study. Cancer 94(2):552-60, 2002.
9
17. Giangaspero RJ, Perry RH, Kelly PJ et al.: World Health Organisation classification of tumours
pathology and genetiics tumours of the central nervous system, in Kleihues P and Cavenee WK
(eds): Tumors of the nervous system. Pathology and genetics. Lyon, IARC Press, 2000,
18. Dasarathy VB: Nearest Neighbor (NN) Norms: NN Patteren clasification tecniques. IEEE Computer
Society Press, Los Alamitos, California, 1991.
19. Segal MR. Regression Trees for Censored Data. Biometrics 44:35-47, 1988.
20. Kaplan EL, Meier P. Non parametric estimation from incomplete observations. J Am Stat Assoc
53:457-481, 1958.
21. Weil MD, Lamborn K, Edwards MS, et al. Influence of a child's sex on medulloblastoma outcome.
JAMA 279:1474-1476, 1998.
22. Roberts RO, Lynch CF, Jones MP, et al. Medulloblastoma: a population-based study of 532 cases. J
Neuropathol Exp Neurol 50:134-144, 1991.
23. Caputy AJ, McCullough DC, Manz HJ, et al. A review of the factors influencing the prognosis of
medulloblastoma. The importance of cell differentiation. J Neurosurg 66(1):80-87, 1987.
24. Zerbini C, Gelber RD, Weinberg D, et al. Prognostic factors in medulloblastoma, including DNA ploidy.
J Clin Oncol;11:616-622,1993.
25. Danjoux CE, Jenkin RD, McLaughlin J, et al. Childhood medulloblastoma in Ontario, 1977-1987:
population-based results. Med Pediatr Oncol 26:1-9,1996.
26. Janss AJ, Yachnis AT, Silber JH, et al. Glial differentiation predicts poor clinical outcome in primitive
neuroectodermal brain tumors. Ann Neurol 39:481-489,1996.
27. Michiels EM, Heikens J, Jansen MJ, et al. Are clinical parameters valuable prognostic factors in
childhood primitive neuroectodermal tumors? A multivariate analysis of 105 cases. Radiother Oncol
54:229-38, 2000.
28. Kunschner LJ, Kuttesch J, Hess K et al. Survival and recurrence factors in adult medulloblastoma:
The M.D. Anderson Cancer Center experience from 1978 to 1998. Neuro-oncol 3:167-173, 2001.
29. David KM, Casey AT, Hayward RD, et al. Medulloblastoma: is the 5-year survival rate improving? A
review of 80 cases from a single institution. J Neurosurg 86:13-21,1997.
10
30. Perek D, Perek-Polnik M, Drogosiewicz M, et al. Risk factors of recurrence in 157 MB/PNET patients
treated in one institution. Childs Nerv Syst 14:582-6, 1998.
Table 1. Patients characteristics and survival comparisons
Total
(n=60)
p value
n
%
Age at initial diagnosis, years
0.3
Median
6
Range
0.6-38.2
Mean +SD
7+ 0.9
Less than 3 years
10
17
0.5
Over 18 years
5
8
Sex
Male
39
65
0.8
Female
21
35
Subtype
Classic
46
77
0.4
Desmoplastic
14
23
Stage
M0
42
70
0.04
M+
18
30
Chemotherapy
VC
5
8
0.9
VC+
55
92
Gene profiling expression group
Group 0 (poor prognosis)
12
20
Group 1 (better prognosis)
48
80
<0.0001
Follow-up, years (survivors)
Median
Range
Mean +SD
Hazard ratio
1
0.6
1,1
1,6
0.4
1
7
4.8
2-10.8
4.9 + 2.1
Table 2. Multivariate Cox proportional hazards model
P-value
Exp. (Coef.)
Stage M0
0.075
0.44
Sex Male
0.052
2.79
High risk gene
<0.0001
12.02
expression pattern
Age (years)
0.013
1.08
95% Lower
0.18
0.99
4.00
95% Upper
1.09
7.88
36.17
1.02
1.16
11
Figure 1. Kaplan-Meier survival plots showing the influence of clinical characteristics on
the cumulative survival of patients in the study population: A for stage, B for sex and C
for subtype
Figure 2. Kaplan-Meier survival plot showing the prognostic significance of combined
gene expression pattern (low risk vs. high risk) and clinical stage for metastatic disease
(M0 vs. M+). A non-significant trend for worse outcome is observed in low risk patients
by gene expression profile but with M+ disease
12