Embryonal tumours of the central nervous system (CNS) represent a heterogeneous group of tumours about which little is known biologically, and whose diagnosis, on the basis of morphologic appearance alone, is controversial. Medulloblastomas, for example, are the most common malignant brain tumour of childhood, but their pathogenesis is unknown, their relationship to other embryonal CNS tumours is debated1,2, and patients’ response to therapy is difficult to predict3. We approached these problems by developing a classification system based on DNA microarray gene expression data derived from 99 patient samples. Here we demonstrate that medulloblastomas are molecularly distinct from other brain tumours including primitive neuroectodermal tumours (PNETs), atypical teratoid/rhabdoid tumours (AT/RTs) and malignant gliomas. Previously unrecognized evidence supporting the derivation of medulloblastomas from cerebellar granule cells through activation of the Sonic Hedgehog (SHH) pathway was also revealed. We show further that the clinical outcome of children with medulloblastomas is highly predictable on the basis of the gene expression profiles of their tumours at diagnosis.
We first addressed the problem of distinguishing different embryonal CNS tumours from each other. This is important as the classification of these tumours based on histopathological appearance is debated (Fig. 1A). There are two hypotheses regarding the classification of medulloblastomas: the first is that they are part of a larger class of PNETs arising from a common cell type in the subventricular germinal matrix1, the second is that they arise from cerebellar granule cell progenitors2. To begin to generate a molecular taxonomy of CNS embryonal tumours, we analysed the gene expression profiles of 42 patient samples (data set A: 10 medulloblastomas, 5 CNS AT/RTs, 5 renal and extrarenal rhabdoid tumours, and 8 supratentorial PNETs, as well as 10 non-embryonal brain tumours (malignant glioma) and 4 normal human cerebella). RNA extracted from frozen specimens was analysed with oligonucleotide microarrays containing probes for 6,817 genes. The gene expression data are available as Supplementary Information II (see also http://www.genome.wi.mit.edu/MPR/CNS).
To determine whether the different types of tumours could be molecularly distinguished, we used a method of data reduction called principal component analysis in which the high dimensionality of the data was reduced to three viewable dimensions representing linear combinations of variables (genes) that account for most of the variance in the original data set (Fig. 1B)4. Normal brain was easily separable from the brain tumours, and the different tumour types were similarly separable. Separation of tumour types was also seen using hierarchical clustering (Fig. 1D)5. A more appropriate strategy for distinguishing known tumour types, however, is to use supervised learning methods to identify the genes most highly correlated with the tumour type distinctions (Fig. 1C, E). Analysis of 1,000 random permutations of the data failed to yield a separation of tumour classes to the extent observed in Fig. 1C, indicating that the observed gene expression patterns could not be explained by chance (Supplementary Information III). The robustness of these markers for classification was further investigated using a weighted-voting algorithm and evaluated by cross validation testing6. Correct classification of the tumours could be achieved with accuracy (35 out of 42 correct classifications, P < 10-10 compared with random classification; see Supplementary Information III).
As expected, malignant gliomas were clearly separable from medulloblastomas, reflecting the derivation of gliomas from cells of non-neuronal origin. Consistent with this, the gliomas expressed genes typical of the astrocytic and oligodendrocytic lineage (PEA15, SOX2, PMP2, Olig-2, TrkB kinase-negative splice variant, S100, GFAP), genes related to metabolism (fructose 2,6-bisphosphatase, glutamate dehydrogenase), and genes involved in cell differentiation (ID2, GDF1, TYK2; Fig. 1E and Supplementary Information III). The medulloblastomas form a cluster that is also separate from the PNETs (Fig. 1C), supporting the hypothesis that these two classes of embryonal tumours are molecularly distinct. Among the genes most highly correlated with the medulloblastoma class were ZIC and NSCL1, encoding transcription factors that are specific for cerebellar granule cells (Fig. 1E)7,8. This result suggests that medulloblastomas, but not PNETs, arise from cerebellar granule cells, or alternatively, have activated the transcriptional programme of cerebellar granule cells.
We next analysed the AT/RT tumours, which have only recently been distinguished from medulloblastoma9. Accurate identification of AT/RT is particularly important because patients with these tumours have an extremely poor prognosis. AT/RT tumours arise in the CNS or in other organs such as the kidney, where they are referred to as rhabdoid tumours. Most tumours harbour hSNF5/INI1 mutations, but it is unknown whether AT/RTs arising in different anatomical locations are molecularly distinct9,10,11. As shown in Fig. 1C, the AT/RTs and rhabdoid tumours were easily distinguishable from the other tumour types in the study. Of note, the CNS AT/RTs and abdominal rhabdoid tumours were molecularly similar despite having arisen in different anatomical locations. This finding supports the idea that they arise from a similar cell of origin. Alternatively, a common mechanism of transformation may yield similar transcriptional programmes in cells of distinct origin. Markers of the distinction between AT/RT and rhabdoid tumours include genes specifically expressed during myogenesis, including skeletal β-tropomyosin, neutral calponin, NFAT3, and myosin regulatory light chain (Fig. 1E and Supplementary Information III). This finding is consistent with the hypothesis that the tumours have a mesenchymal origin.
We next focused on molecular heterogeneity within a single tumour type, medulloblastoma. The principal histological subclass of medulloblastoma is desmoplastic medulloblastoma, although its diagnosis is highly subjective (Fig. 1A). Desmoplastic medulloblastoma is of interest because it is seen with high frequency in patients with Gorlin's syndrome, a rare autosomal dominant disorder resulting from mutation of the SHH receptor PTCH12,13. It is unclear whether dysregulation of the SHH pathway, known to be mitogenic for cerebellar granule cells, is also involved in the pathogenesis of sporadic desmoplastic medulloblastoma14,15,16,17,18.
To determine whether desmoplastic and classic medulloblastoma are distinguishable by gene expression, we analysed 34 medulloblastoma samples (data set B) whose histology was scored using World Health Organization (WHO) criteria19. As shown in Fig. 2, a sharp and statistically significant gene expression signature of desmoplastic histology was evident, and this signature was sufficient for correct classification of 33 out of 34 tumours (P = 8.6 × 10-7 compared with random classification; see Supplementary Information III). Notably, among the genes most highly correlated with desmoplastic medulloblastoma was PTCH (itself a transcriptional target of SHH), as well as two other SHH downstream targets: GLI20 and N-MYC (A. Kenney and D. Rowitch, personal communication). Furthermore, insulin-like growth factor II expression was correlated with desmoplastic histology, and its expression is essential for SHH-mediated tumorigenesis in mice21. Taken together, the transcriptional profiling indicates that sporadic desmoplastic medulloblastomas, like tumours associated with Gorlin's syndrome, are characterized by activation of the SHH signalling pathway, further supporting the proposal that SHH dysregulation may be important in the pathogenesis of medulloblastoma.
A clinical challenge concerning medulloblastoma is the highly variable response of patients to therapy. Whereas some patients are cured by chemotherapy and radiation, others have progressive disease3. Currently, the only prognostic factor used in clinical practice is tumour staging; a reflection of postoperative tumour size and the presence of metastases. Unfortunately, staging-based prognostication is imperfect in that many patients with low-stage disease still succumb to their disease. There are currently no molecular markers of outcome used in clinical practice for any brain tumour. High levels of expression of the neurotrophin-3 receptor (TRKC), however, have been reported to correlate with a favourable medulloblastoma outcome22,23, suggesting a molecular basis for the variability of medulloblastoma outcome. Molecular correlates of medulloblastoma metastasis have also been recently reported24.
To explore the heterogeneity in response to treatment of medulloblastomas, we expanded our analysis to include 60 similarly treated patients from whom biopsies were obtained before receiving treatment, and for whom clinical follow-up was available (data set C). We first investigated whether clustering methods would identify biologically distinct subsets of the tumours. The tumours were clustered into two groups using self-organizing maps (SOM); an unsupervised algorithm that groups samples into a predetermined number of clusters on the basis of their gene expression patterns6,25. The genes most highly correlated with the SOM clusters were primarily ribosomal protein-encoding genes (Supplementary Information III), suggesting differences in ribosome biogenesis. Blinded electron microscopic examination of 9 samples by 3 observers confirmed that tumours falling into the cluster characterized by high expression of ribosomal protein genes contained higher numbers of ribosomes (P = 0.03, Fisher's exact test; Fig. 3). We next investigated whether the SOM-derived clusters were correlated with patient survival. No statistically significant difference in the proportion of survivors compared with treatment failures in each cluster was observed (Fisher's exact test, P = 0.1; see Supplementary Information III). Because unsupervised methods are generally not the most appropriate analytical approach to predicting known distinctions such as outcome, we developed a supervised learning outcome predictor based on gene expression in which the classifier ‘learns’ the distinction between patients who are alive after treatment (‘survivors’) compared with those who succumbed to their disease (‘failures’, minimum follow-up of 24 months for surviving patients, overall median 41.5 months).
We used a k-nearest neighbours (k-NN) algorithm26 that computes the distance of a test sample to each of the training set samples, each of which has an associated class (in this case, survivor or failure), and then predicts the class of the test sample to be that of the majority of the k-closest samples. The k-NN classifier was evaluated by cross-validation, whereby one sample is randomly withheld, a model is trained on the remaining samples, and the model is then used to predict the class of the withheld sample. The process is repeated until all of the samples are tested.
Outcome predictions based on gene expression were statistically significant for k-NN models ranging from 2 to 21 genes, with optimal predictions made by an 8-gene model that made only 13/60 classification errors (Fisher's exact test, P = 0.0002). Shown most clearly by a Kaplan–Meier survival analysis in Fig. 4a, patients that were predicted to be survivors had a 5 year overall survival of 80% compared with 17% for patients predicted to have a poor outcome (P = 0.000003, log rank test). A more conservative method of assessing statistical significance is to attempt to optimize classifiers of random permutations of the survivor/failure class labels. We performed 1,000 such permutations and found only 9 for which prediction accuracy matched or exceeded our observed result (Supplementary Information III), indicating that the result is unlikely to be achieved by chance (P = 0.009). We subsequently tested several other classification algorithms including weighted voting6,27, support vector machines28,29 and IBM SPLASH30, all of which performed with similarly high accuracy (Supplementary Information I and III).
We explored further the clinical value of the predictor by considering existing prognostic factors for medulloblastoma outcome. Patients with localized disease (M0) had a more favourable outcome compared with patients with involvement of the cerebrospinal fluid or with distant metastases (M+) (P = 0.03 comparing M0 with M+ by Kaplan–Meier analysis), although not all M0 patients survived. When our outcome predictor was applied only to the 42 M0 patients, the prediction of outcome remained significant (P = 0.002), indicating that the expression-based predictor substantially improved staging-based prognostication. Similarly, prediction based on TRKC expression was imperfect in this series in that not all patients in the unfavourable (TRKC-low) category died. When our gene expression-based predictor was applied to the 33 TRKC-low patients, the surviving patients could be significantly separated from those who succumbed to their disease (P = 0.01, Supplementary Information III). Of note, not all patients in this study received identical therapy. However, restricting our analysis to the 35 patients that received surgery, vincristine, cisplatin and cyclophosphamide, the predictor continued to yield a significant Kaplan–Meier survival distinction (P = 0.0012). Taken together, these results demonstrate that the outcome predictor based on gene expression exceeds other approaches to prognosis determination.
A number of genes not previously associated with clinical outcome were identified (Fig. 4b). Those correlated with favourable outcome included many genes characteristic of cerebellar differentiation (vesicle coat protein β-NAP, NSCL1, TRKC, sodium channels), and genes encoding extracellular matrix proteins (PLOD lysyl hydroxylase, collagen type V αI, elastin). As expected, TRKC expression was correlated with a favourable outcome, consistent with previous reports of this association22,23. In contrast, genes related to cerebellar differentiation were underexpressed in poor prognosis tumours, which were dominated by the expression of genes related to cell proliferation and metabolism (MYBL2, enolase 1, LDH, HMG1(Y), cytochrome C oxidase) and multidrug resistance (sorcin). Genes correlated with poor outcome included a number of the ribosomal protein-encoding genes identified by the SOM clustering experiments (Fig. 4b). This indicates that whereas this ribosomal signature is correlated with poor outcome, optimal outcome prediction requires not only these genes, but also genes correlated with a favourable outcome that were not identified by the unsupervised clustering analysis.
The routine clinical implementation of genomics-based outcome predictors must await confirmation in independent data sets, and the models may need to be modified as treatment regimens evolve. For patients predicted to have a favourable outcome, efforts to minimize toxicity of therapy might be indicated, whereas for those predicted not to respond to standard therapy, earlier treatment with experimental regimens might be considered. This work illustrates how genomic technologies have the potential to advance treatment planning beyond the empiric, towards a more molecularly defined, individualized approach to medicine.
Patients included 60 children with medulloblastomas, 10 young adults with malignant gliomas (WHO grades III and IV), 5 children with AT/RTs, 5 with renal/extrarenal rhabdoid tumours, and 8 children with supratentorial PNETs (see Supplementary Information I). Medulloblastoma patients were treated with craniospinal irradiation to 2,400–3,600 centiGray (cGy) with a tumour dose of 5,300–7,200 cGy. All patients with medulloblastoma were treated with chemotherapy consisting of cisplatin and vincristine, plus combinations of carboplatin, etoposide, cyclophosphamide or lumustine (1-(2-chloroethyl)-3-cyclohexyl-1-nitrosourea, CCNU) (details in Supplementary Information II). Samples were snap frozen in liquid nitrogen and stored at -80 °C. Studies were done with approval of the Committee for Clinical Investigation of Boston Children's Hospital. The data were organized into three sets: data set A (42 samples containing 10 medulloblastomas, 10 malignant gliomas, 10 AT/RTs, 8 PNETs and 4 normal cerebella); data set B (34 samples containing 9 desmoplastic medulloblastomas and 25 classic medulloblastomas); and data set C (60 samples containing 39 medulloblastoma survivors and 21 treatment failures). The clinical attributes of each of the patients in the study are available in Supplementary Information II. Tissues were homogenized in guanidinium isothiocyanate and RNA was isolated by centrifugation over a CsCl gradient. RNA integrity was assessed either by northern blotting or by gel electrophoresis. Ten–twelve micrograms total RNA was used to generate biotinlylated antisense RNAs, which were hybridized overnight to HuGeneFL arrays containing 5,920 known genes and 897 expressed sequence tags, as described previously6. Arrays were scanned on Affymetrix scanners and the expression value for each gene was calculated using GENECHIP software (Affymetrix, Santa Clara, California). Minor differences in microarray intensity were corrected using a linear scaling method as detailed in Supplementary Information I. Scans were rejected if the scaling factor exceeded 3, fewer than 1,000 genes received ‘present’ calls, or microarray artefacts were visible.
Preprocessing and clustering
The gene expression data were subjected to a variation filter that excluded genes showing minimal variation across the samples being analysed, as detailed in Supplementary Information I.
The data were first normalized by standardizing each column (sample) to mean 0 and variance 1. SOMs were performed using our GeneCluster clustering package (http://www.genome.wi.mit.edu/MPR/software). Hierarchical clustering was performed using Cluster and TreeView software5. Principal component analysis (PCA) was performed by computing and then plotting the three principal components using the S-Plus statistical software package (http://www.insightful.com/products/desktop.asp) using default settings.
Genes correlated with particular class distinctions (for example, classic versus desmoplastic medulloblastoma) were identified by sorting all of the genes on the array according the signal-to-noise statistic (µ0 - µ1)/(σ0 + σ1), where µ and σ represent the median and standard deviation of expression, respectively, for each class. Similar results were obtained using a standard t-statistic as the metric ((µ0 - µ1)/√(σ20/N0 + σ21/N1)), where N represents the number of samples in each class (see Supplementary Information). Permutation of the column (sample) labels was performed to compare these correlations to what would be expected by chance in 99% of the permutations. For classification, we developed a modification of the k-NN algorithm26 that predicts the class of a new data point by calculating the euclidean distance (d) of the new sample to the k nearest samples (for these experiments we set k = 5) in the training set using normalized gene expression data, and selecting the class to be that of most of the k samples. The weight given to each neighbour was 1/d. The k-NN models were evaluated by 60-fold ‘leave-one-out’ cross-validation, whereby a training set of 59 samples was used to predict the class of a randomly withheld sample, and the cumulative error rate was recorded. We tested models with variable numbers of genes (1–200, selected according to their correlation with the survivor versus treatment failure distinction in the training set) in this manner. An 8-gene k-NN outcome prediction model yielded the lowest error rate, and was therefore used to generate Kaplan–Meier survival plots using S-Plus. Predictors using metastatic staging or TRKC expression were constructed by finding the decision boundary half way between the classes, (µclass0 + µclass1)/2 using either the staging values 0 versus 1, 2, 3, 4 or the continuous TRKC microarray gene expression levels, and then predicting the unknown sample according to its location with respect to that boundary.
We thank members of the Whitehead/MIT Center for Genome Research, Program in Cancer Genomics, and J. Volpe for discussions and comments on the manuscript. This work was supported in part by Millennium Pharmaceuticals, Affymetrix and Bristol-Myers Squibb (E.S.L.); NIH grants (S.L.P. and T.C.); NIH-supported Mental Retardation Research Center (S.L.P.) and Cancer Center Support CORE (T.C.); the American Lebanese Syrian Associated Charities (ALSAC); and the Kyle Mullarkey Medulloblastoma Research Fund. We acknowledge the Cooperative Human Tissue Network and the Children's Oncology Group for contributing tumour samples.
About this article
Applied Intelligence (2018)