In modern clinical neuro-oncology, histopathological diagnosis affects therapeutic decisions and prognostic estimation more than any other variable. Among high-grade gliomas, histologically classic glioblastomas and anaplastic oligodendrogliomas follow markedly different clinical courses. Unfortunately, many malignant gliomas are diagnostically challenging; these nonclassic lesions are difficult to classify by histological features, generating considerable interobserver variability and limited diagnostic reproducibility. The resulting tentative pathological diagnoses create significant clinical confusion. We investigated whether gene expression profiling, coupled with class prediction methodology, could be used to classify high-grade gliomas in a manner more objective, explicit, and consistent than standard pathology. Microarray analysis was used to determine the expression of ∼12,000 genes in a set of 50 gliomas, 28 glioblastomas and 22 anaplastic oligodendrogliomas. Supervised learning approaches were used to build a two-class prediction model based on a subset of 14 glioblastomas and 7 anaplastic oligodendrogliomas with classic histology. A 20-feature k-nearest neighbor model correctly classified 18 of the 21 classic cases in leave-one-out cross-validation when compared with pathological diagnoses. This model was then used to predict the classification of clinically common, histologically nonclassic samples. When tumors were classified according to pathology, the survival of patients with nonclassic glioblastoma and nonclassic anaplastic oligodendroglioma was not significantly different (P = 0.19). However, class distinctions according to the model were significantly associated with survival outcome (P = 0.05). This class prediction model was capable of classifying high-grade, nonclassic glial tumors objectively and reproducibly. Moreover, the model provided a more accurate predictor of prognosis in these nonclassic lesions than did pathological classification. These data suggest that class prediction models, based on defined molecular profiles, classify diagnostically challenging malignant gliomas in a manner that better correlates with clinical outcome than does standard pathology.
Malignant gliomas are the most common primary brain tumor and result in an estimated 13,000 deaths each year in the United States 3 . Glial tumors are classified histologically, with pathological diagnosis affecting prognostic estimation and therapeutic decisions more than any other variable. Among high-grade gliomas, anaplastic oligodendrogliomas have a more favorable prognosis than glioblastomas (1) . Moreover, although glioblastomas are resistant to most available therapies, anaplastic oligodendrogliomas are often chemosensitive, with approximately two-thirds of cases responding to procarbazine, 1-(2-chloroethyl)-3-cyclohexyl-1-nitrosourea, and vincristine (2 , 3) . Paradoxically, recognition of the clinical importance of diagnosing anaplastic oligodendroglioma has blurred the histopathological line separating glioblastoma and oligodendroglioma; to ensure that patients are not deprived of effective chemotherapy, pathologists have loosened their criteria for anaplastic oligodendroglioma. Indeed, this diagnostic promiscuity has recently been described as a “contagion” (4) . As such, there is a critical need for an objective, clinically relevant method of glioma classification.
The most widely used histological system of brain tumor classification is that of the WHO (1) . Gliomas are classified according to defined histological features characteristic of the presumed normal cell of origin. Tumors of classic histology clearly display these features and resemble typical depictions in standard textbooks (5 , 6) ; these cases would be diagnosed similarly by nearly all pathologists. Unfortunately, there are situations in which the WHO classification system is problematic, primarily because pathological diagnosis remains subjective (7) ; intratumoral histological variability is common, and high-grade gliomas can display little cellular differentiation, thus lacking defining histological features. The diagnosis of tumors with such nonclassic histology is often controversial. Consequently, diagnostic accuracy and reproducibility are jeopardized, and significant interobserver variability can occur. Coons et al. (8) found that complete diagnostic concordance among four neuropathologists reviewing gliomas over four sessions peaked at 69%. Giannini et al. (9) , in a study of seven neuropathologists and six surgical pathologists scoring histological features of oligodendroglioma, found that agreement for identifying features ranged from 0.05 to 0.8, confirming that numerous classification parameters are not easily reproduced.
To develop more objective approaches to glioma classification, recent investigations have focused on molecular genetic analyses. Sasaki et al. (10) demonstrated loss of chromosome 1p in 86% of oligodendrogliomas with classic histology and maintenance of both 1p alleles in 73% of “oligodendrogliomas” with astrocytic features. Interestingly, tumor genotypes more closely predicted chemosensitivity, demonstrating an ability of tumor genotype to augment standard pathology. Burger et al. (11) also demonstrated close correlation between classic low-grade oligodendroglioma appearance and allelic losses of 1p and 19q. In gene expression studies, Lu et al. (12) suggested that expression of oligodendrocyte lineage genes (Olig1 and 2) might augment identification of oligodendroglial tumors. Similarly, Popko et al. (13) found three of four myelin transcripts significantly more often in oligodendrogliomas than in astrocytomas.
The advent of expression microarray techniques now allows simultaneous analysis of thousands of genes. We hypothesized that this approach could identify molecular markers capable of refining the current method of malignant glioma classification. We therefore investigated whether gene expression profiling, coupled with the computational methodology of class prediction (14) , could be used to define subgroups of high-grade glioma in a manner more objective, explicit, and consistent than standard pathology. To this end, a subset of gliomas with classic histology was used to build a class prediction model, and this model was then used to predict the classification of samples with nonclassic histology.
MATERIALS AND METHODS
Glioma Tissue Samples.
These investigations have been approved by the Massachusetts General Hospital Institutional Review Board. Tissue samples were collected from Canadian Brain Tumor Tissue Bank (London, Ontario, Canada), Massachusetts General Hospital (Boston, MA), Brigham and Women’s Hospital (Boston, MA), and Charité Hospital (Berlin, Germany). Samples were collected immediately after surgical resection, snap frozen, and stored at −80°C. H&E-stained frozen sections were reviewed histologically for every specimen (D. N. L.); samples containing significant regions of normal cell contamination (>10%) and/or excessively large amounts of necrotic material were excluded. Using these criteria, 50 high-grade glioma samples were selected (Table 1) ⇓ , 28 glioblastomas and 22 anaplastic oligodendrogliomas; all were primary tumors sampled before therapy. All cases had been diagnosed at the primary hospital by board-certified neuropathologists. Original pathology slides were obtained and reviewed centrally by two additional neuropathologists (M. E. M. and D. N. L.) for diagnostic confirmation and selection of the classic tumor subset. Anaplastic oligodendrogliomas designated as having classic histopathology exhibited relatively evenly distributed, uniform, and rounded nuclei and frequent perinuclear halos (10) . In contrast, classic glioblastomas were characterized by irregularly distributed, pleomorphic, and hyperchromatic nuclei, sometimes with conspicuous eosinophilic cytoplasm. The classic subset of tumors were cases diagnosed similarly by all examining pathologists, and each case resembled typical depictions in standard textbooks (5 , 6) . A total of 21 classic tumors was selected, and the remaining 29 samples were considered nonclassic tumors, lesions for which diagnosis might be controversial. Of the 21 classic tumors, 14 were glioblastomas, and 7 were anaplastic oligodendrogliomas.
Gene Expression Profiling.
Tissues were homogenized in guanidinium isothiocyanate, and RNA was isolated using a CsCl gradient. RNA integrity was confirmed by gel electrophoresis. For each sample, 15 μg of total RNA were used to generate biotinylated cRNAs, which were hybridized overnight to Affymetrix U95Av2 GeneChips as described previously (14 , 15) . On the basis of previous experience, one array per sample provided reproducible results with a sample set of the size used in this study (14 , 16) . Arrays were scanned on Affymetrix scanners, and data were collected using GeneChip software (Affymetrix, Santa Clara, CA). Scan quality was assured based on a priori quality control criteria, which included the absence of visible microarray artifacts (e.g., scratches) and significant differences in microarray intensity, and the presence of >30% “present” calls for the ∼12,600 genes and expressed sequence tags on the U95Av2 GeneChips.
Class Prediction Methodology.
The subset of classic gliomas was used to build a class prediction model. This model was then used to predict the classification of the nonclassic samples. Raw expression values were normalized by linear scaling so that mean array intensity for active (present) genes was identical for all scans 4 . Data filtration settings were based on previous studies (14 , 16) . Intensity thresholds were set at 20 and 16,000 units. Gene expression data were subjected to a variation filter that excluded genes showing minimal variation across the samples; genes whose expression levels varied <100 units between samples, and genes whose expression varied <3-fold between any two samples, were removed. The variation filters excluded two-thirds of the genes, leaving ∼3,900 genes for building class prediction models. Further feature (gene) selection was effected, as described previously (14 , 16) , using the S2N 5 statistic. S2N ratio ranks genes based on their correlation to each of the two class distinctions (i.e., classic glioblastoma and anaplastic oligodendroglioma). In addition, the significance of the highly ranked genes was confirmed by random permutation testing; the sample classification labels were permuted, and the S2N ratio was recomputed to compare the true gene correlations to what would have been expected by chance. Five different k-NN class prediction models were built, using different gene numbers (10, 20, 50, 100, and 250 genes), with GeneCluster 6 . Training error (on the classic cases) for these k-NN models was determined using leave-one-out cross-validation, where one sample is withheld, and the class membership of this withheld sample is predicted using a model built on the remaining samples. Class prediction for the withheld sample was the majority class membership of the k (k = 3 in these experiments) closest “neighboring” samples based on the Euclidean distance between the sample under consideration and samples used in training the k-NN model. This process was repeated for each sample in the training set, and a cumulative training error was calculated. Finally, a k-NN model was built using all 21 classic cases (with no samples left out), which was then used to predict classification of the remaining gliomas based on the class labels of the k-NNs of each sample.
Survival Analyses: Statistical Methods.
Survival distributions were compared between groups defined by pathology or gene expression profiling using permutation Log-rank tests, computed by drawing 50,000 samples from the relevant permutation distribution. The statistical programming language, R, 7 was used to compute permutation Ps. Kaplan-Meier plots were generated with GraphPad Prism (Version 3.02; GraphPad Software, San Diego, CA).
RESULTS AND DISCUSSION
Training of the k-NN Class Prediction Models.
We investigated whether gene expression profiling could be used to define subgroups of high-grade glioma more objectively and consistently than standard pathology. To this end, we examined the expression profile of 14 glioblastomas and 7 anaplastic oligodendrogliomas with classic histology (Fig. 1A) ⇓ . Features (genes) correlating with each of the two class distinctions were ranked according to S2N as described; diagrammatic results for the top 50 features of each class are illustrated (Fig. 1B ⇓ ; the complete list of genes is available online). 4 Because the expression profiles demonstrated robust class distinctions, we proceeded to construct five k-NN class prediction models. The number of features used in the models was chosen to give a range of prediction accuracy; increasing the number of genes in a model can improve prediction accuracy by providing additional biologically relevant input and affording robust signals against noise, whereas using too many genes can increase inaccuracy by generating excess noise. Models were built using 10, 20, 50, 100, or 250 features, and the training error for each model was calculated using leave-one-out cross-validation (Table 2) ⇓ . Although accuracy of the models was comparable, the 20-feature k-NN model was chosen for further study because it predicted most accurately the class distinctions of the classic glioma training set (18 of 21 correct calls; 86% accuracy).
The 20 features used for prediction in this model correspond to 19 genes because of the presence of redundant probe sets (Table 3) ⇓ . Genes highly correlated with glioblastoma included a mixture of metabolic, structural, and signaling proteins. In particular, Rho GTPases (ARHC) and mitogen-activated protein kinases are members of Ras signal transduction pathways known to play a role in tumorigenesis and cell migration (17 , 18) . A large proportion of genes highly correlated with anaplastic oligodendroglioma was found to be involved in protein translation and ribosome biogenesis; translation factors have been implicated previously as effectors of tumorigenesis (19) . Paradoxically, ribosomal protein-encoding genes were found recently to be correlated with poor outcome in medulloblastoma (16) . These models thus provide a substantial number of features that correlate with glioma class distinction, but determination of the biological and clinical significance of these genes requires additional studies.
Training “Errors” of the Class Prediction Model.
Although a class prediction was made for all 21 classic gliomas using the model, such techniques typically classify some samples with more confidence than others. For this reason, confidence values were calculated for all predictions (Table 4) ⇓ . Of the three errors within the classic training set, one prediction was made with relative high confidence (“Brain_CO_4”; ranked 9 of 21), and two were classified as low confidence predictions (“Brain_CG_5” and “Brain_CG_10”; ranked 16 and 18, respectively). “Brain_CO_4,” a classic anaplastic oligodendroglioma, displayed a gene expression profile strikingly more similar to that of glioblastoma (Fig. 1B) ⇓ and was classified as a glioblastoma with relative high confidence in all five k-NN models examined (mean confidence value of 0.17). Reexamination of reports from the initial diagnosis and slides from the central pathology review gave no justification for a histological classification of glioblastoma. Although some evidence of nuclear pleomorphism and hyperchromasia was noted in the original pathology report, the presence of prominent perinuclear halos and a fine capillary network indicated a classic anaplastic oligodendroglioma. Furthermore, glial fibrillary acidic protein, an astrocytic marker, was not expressed in the neoplastic cells. Notably, however, although the histological features of “Brain_CO_4” were consistent with anaplastic oligodendroglioma, clinical data suggested a course more characteristic of a glioblastoma, with survival of only 7 months from diagnosis.
Independent Validation of Class Prediction through Survival Analysis.
The prediction model classified 18 of 21 classic gliomas identically to the pathological classification during leave-one-out cross-validation. The discrepancies in tumor classification could be the result of a class prediction model error or a diagnostic error; preliminary examination of the clinical behavior of “Brain_CO_4” suggested that the class prediction model provided more pertinent tumor classification. Ideally, the designation of error requires independent validation. Differences in survival between patients with glioblastomas and those with anaplastic oligodendrogliomas have been well documented (1) ; consequently, as an independent validation of the gene expression prediction model, prediction model classifications were compared with pathological diagnoses with respect to survival. When the classic gliomas were sorted according to pathology, a clear distinction was found between survival of patients with glioblastoma and those with anaplastic oligodendroglioma (Fig. 2) ⇓ . Although this comparison was not statistically significant (n = 21, P = 0.21), most likely because of the small sample size and relatively short follow-up time on three of the seven anaplastic oligodendrogliomas, statistically significant differences in survival were seen within the pathologically defined classes when all glioblastomas and anaplastic oligodendrogliomas were compared (n = 50, P = 0.009; data not shown). Remarkably, however, when the classic gliomas were sorted using class distinctions according to the model, survival differences were statistically significant (n = 21, P = 0.031; Fig. 2 ⇓ ). These results demonstrate that, even within high-grade gliomas of classic histology, the biologically and clinically relevant information afforded by the genetic profiles augments that provided by pathology alone. Furthermore, the clinical outcome data suggest that the discrepancies in tumor classification are more likely caused by a diagnostic error than a class prediction model error.
Class Prediction of Nonclassic High-grade Gliomas.
Next, we examined the ability of this model to classify the common, nonclassic high-grade gliomas that currently cause such clinical uncertainty regarding therapy and prognosis (Fig. 3A) ⇓ . The ability to identify these lesions in a uniform and reproducible manner would facilitate more accurate therapeutic decisions and prognostic estimation, allowing for improved clinical management of individual patients. The prediction model classifications were compared with pathological diagnoses with respect to survival. When these diagnostically challenging tumors were classified according to pathology, survival of patients with nonclassic glioblastoma was not significantly different from that of patients with nonclassic anaplastic oligodendroglioma (n = 29, P = 0.194; Fig. 3B ⇓ ). These results demonstrate clearly the difficulty in distinguishing these challenging cases in a clinically relevant manner based exclusively on histological parameters. In contrast, class distinctions according to the gene expression-based model trained on the classic gliomas were statistically significant (P = 0.051), giving much better separation between the anaplastic oligodendroglioma and glioblastoma survival curves (Fig. 3B) ⇓ . Thus, gene expression profiles have a remarkable ability to distinguish histologically ambiguous glioblastomas and anaplastic oligodendrogliomas in a clinically relevant manner. Indeed, gene expression profiles provide a more objective and accurate predictor of prognosis in high-grade nonclassic gliomas than does traditional histology. In addition, the ability to distinguish histologically ambiguous gliomas enables appropriate therapies to be tailored to specific tumor subtypes, sparing patients who would not respond from unnecessary treatments. Moreover, uniform and reproducible classification of these nonclassic lesions would provide improved stratification of patients in clinical trials and molecular marker studies.
We investigated whether gene expression profiling, coupled with the computational methodology of class prediction, could be used to define subgroups of high-grade glioma in a manner more objective, explicit, and consistent than standard pathology. Not only was this method effective at classifying high-grade gliomas objectively and reproducibly, it also appeared to provide a more accurate predictor of prognosis. Although the training sample sets for these models were selected based on classic histological features, the biologically and clinically relevant information afforded by the genetic profiles greatly augments that provided by pathology alone. These data therefore suggest that class prediction models, based on defined molecular profiles, classify diagnostically challenging malignant gliomas in a manner that better correlates with clinical outcome than does standard pathology.
We thank Magdalena Zlatescu and Loc Pham for valuable assistance with collecting patient data, Marcela White and Jennifer Roy for accessing tissue samples and information, Lisa Sturla for technical assistance, members of the Program in Cancer Genomics at the Whitehead Institute/Massachusetts Institute of Technology Center for Genome Research for valuable discussions, and Anat Stemmer-Rachamimov for critical review of this manuscript.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
↵1 Supported in part by NIH Grant CA57683 (D. N. L.), Affymetrix and Bristol-Myers Squibb (Whitehead Institute/MIT Center for Genome Research), NIH Grant NS35701 (S. L. P.), and Canadian Institutes of Health Research MOP37849 (J. G. C.).
↵2 To whom requests for reprints should be addressed, at Molecular Pathology Laboratory, CNY7, Massachusetts General Hospital, 149 13th Street, Charlestown, MA 02129. Phone: (617) 726-5690; Fax: (617) 726-5079; E-mail:
↵3 Internet address: http://www.cbtrus.org.
↵4 Internet address: http://www-genome.wi.mit.edu/cancer/pub/glioma.
↵5 The abbreviations used are: S2N, signal-to-noise; k-NN, k-nearest neighbor.
↵6 Internet address: http://www-genome.wi.mit.edu/cancer/software/software.html.
↵7 Internet address: http://www.r-project.org.
- Received October 11, 2002.
- Accepted February 3, 2003.
- ©2003 American Association for Cancer Research.