A 16-Gene Signature Distinguishes Anaplastic Astrocytoma from Glioblastoma

Anaplastic astrocytoma (AA; Grade III) and glioblastoma (GBM; Grade IV) are diffusely infiltrating tumors and are called malignant astrocytomas. The treatment regimen and prognosis are distinctly different between anaplastic astrocytoma and glioblastoma patients. Although histopathology based current grading system is well accepted and largely reproducible, intratumoral histologic variations often lead to difficulties in classification of malignant astrocytoma samples. In order to obtain a more robust molecular classifier, we analysed RT-qPCR expression data of 175 differentially regulated genes across astrocytoma using Prediction Analysis of Microarrays (PAM) and found the most discriminatory 16-gene expression signature for the classification of anaplastic astrocytoma and glioblastoma. The 16-gene signature obtained in the training set was validated in the test set with diagnostic accuracy of 89%. Additionally, validation of the 16-gene signature in multiple independent cohorts revealed that the signature predicted anaplastic astrocytoma and glioblastoma samples with accuracy rates of 99%, 88%, and 92% in TCGA, GSE1993 and GSE4422 datasets, respectively. The protein-protein interaction network and pathway analysis suggested that the 16-genes of the signature identified epithelial-mesenchymal transition (EMT) pathway as the most differentially regulated pathway in glioblastoma compared to anaplastic astrocytoma. In addition to identifying 16 gene classification signature, we also demonstrated that genes involved in epithelial-mesenchymal transition may play an important role in distinguishing glioblastoma from anaplastic astrocytoma.


Selection of patient cohort for the analysis of age, survival and molecular markers of discordant samples.
To additionally analyze the clinical features like age and survival as well as the molecular markers, we combined the patients of our cohort (patients of training and test set) as well as those of validation datasets (TCGA, GSE1993 and GSE4422) based on the on the availability of the information. For Age, we the included Authentic AA samples (n=21) and Authentic GBM (n=37) samples of GSE1993 dataset. A total of 8 Discordant AA samples were from our cohort (n=2), GSE1993 dataset (n=5) and GSE4422 dataset (n=1). A total of 20 Discordant GBM samples were from our cohort (n=13), GSE1993 dataset (n=2) and GSE4422 dataset (n=5). For survival analysis, samples of GSE1993 dataset were not included because of lack of censoring status. A total of 13 Authentic AAs were from TCGA (n=9) and GSE4422 (n=4) datasets. For Authentic GBMs, a total of 165 samples were considered from GSE4422 (n=66) and TCGA datasets (n=99). For Discordant GBM samples, a total of 13 samples were considered from our patient cohort (n=8) and GSE4422 dataset (n=5). Since information on CDKN2A/2B loss, EGFR amplification and p53 mutation was available for samples of GSE1993 dataset, Authentic and Discordant samples of GSE1993 dataset alone were used for molecular marker analysis.
EGFR amplification and CDKN2A/2B loss was assessed by PCR analysis and p53 mutation was assessed by Single-Strand Conformation Polymorphism (SSCP) Analysis as described before [1] Supplementary Results

The validation of the 16-gene signature in GSE4271 dataset (Phillips et al dataset)
The GSE4271 dataset comprised of 22 AA samples and 76 GBM samples [2]. Out of 16genes of our signature set, the expression data was available for only 14-genes; the data for DCN and LGALS3 genes were not available. So, we used the expression data for the 14-genes and performed PAM analysis. Using PAM with a 10-fold cross validation (Supplementary Figure   S5A) the 14 genes of the 16-gene signature was able to predict 12 AA samples out of 22 correctly with an error rate of 0.45. Similarly, among 76 GBM samples used, our 16-gene signature predicted 68 samples correctly as GBM with an error rate of 0.1 (Supplementary Figure S5A). Thus, the 16 gene expression signature could discriminate GBM from AAs with an overall diagnostic accuracy of 81.6% (Table 2). The sensitivity for AA is 54.5%, whereas for GBM, it is 89.4%; the specificity for AA is 89.4%, whereas for GBM, it is 54.5% ( Table 2).
While we do not know the exact reason for the low accuracy of 16 gene signature in classifying high grade glioma from GSE4271 (Phillips dataset), one possible reason could be because of the missing data for 2 genes of the 16-gene signature.
Further to see if there is any difference in the clinical features of the authentic and discordant samples as per PAM of 14-genes, we looked at the average age and the survival of discordant and authentic AA and GBM samples. We included all the samples of this dataset: Authentic AAs (n=12), Authentic GBMs (n=68), Discordant AAs (n=10) and Discordant GBMs (n=8). As expected, the average age of Authentic AA (34 years) was significantly (p < 0.0001) lower than that of Authentic GBM (49.7 years) (Supplementary Figure S5B). The Discordant AAs (41.4 years) were older in age as compared to the Authentic AAs (34 years) (though statistically not significant) whereas Discordant GBMs (38.6 years) were significantly (p=0.04) younger in age as compared to the Authentic GBMs (49.7 years) (Supplementary Figure S5B).
In addition, the average age of Discordant AAs was similar to Authentic GBM (p=0.06) whereas the average age of Discordant GBMs was similar to Authentic AA (p=0.36).
With respect to patient survival, as expected, the survival length of Authentic AA (median survival=61 months/ 5.1 years) was significantly (p < 0.004) higher than that of Authentic GBM (median survival=15.5 months/ 1.3 years). Further analysis revealed that the median survival of Discordant AAs (median survival=27 months/ 2.25 years) was significantly (p=0.02) lower than that of Authentic AAs (median survival=61 months/ 5.1 years) Figure S5C). Similarly, the Discordant GBMs (median survival=51 months/ 4.2 years) had statistically significant (p=0.03) better survival as compared to the Authentic GBMs (median survival=15.5 months/ 1.3 years) (Supplementary Figure S5C). This suggests that though the 14-genes of the signature are inadequate for the accurate classification of AA and GBM samples, there is a trend of discordant AA and GBMs belonging to the other group.

(Supplementary
The validation of Petalidis gene signature in TCGA dataset to see its potential in classifying

AA and GBM
We performed analysis to check the potentiality of the Petalidis gene signature in the classification of AA and GBM samples of the TCGA dataset. Out of 59 genes of the Petalidis signature, the data was available for 54 genes in the TCGA dataset. Thus, we have used the 54 genes for the PAM analysis. The PAM analysis revealed that the 55 genes were able to classify the AA and GBM samples with 100% accuracy in the TCGA dataset at threshold 0.0 Figure S6).

and GBM
We performed additional analysis wherein the ability of Phillips gene signature to classify AAs from GBMs in our dataset was addressed. The results of these analyses are described subsequently. In Phillips et al paper [2], a set of 8 genes are described as a marker for 3 prognostic subclasses of high grade glioma: Olig2, DLL3 and BCN as Proneural markers, PCNA and TOP2A as Proliferative markers and CHI3L1/YKL40, CD44 and VEGF as Mesenchymal markers.
Out of the 8 genes of the signature, we used the expression data (obtained from our dataset) of Olig2, DLL3 and BCAN as markers of Proneural subclass while that of TOP2A and CHI3L1 as markers for other two subclasses. From the PAM analysis (Supplementary Figure   S7A), it is clear that these 5 genes were not adequate for the classification of AA and GBM samples: the classification sensitivity for AA being only 44% (22 of 50 AAs are rightly classified) and the specificity for AA was 99%. However, the classification sensitivity for GBM was 99% (131 of 132 GBMs are rightly classified) and the specificity for GBM was 44% (Supplementary Figure S7B). Thus it appears that Phillips et al gene signature cannot be used classification of AAs from GBMs.
The inadequacy of the Phillips gene signature to accurately classify AA and GBM in our dataset could be because of the lack of expression data for 3 genes. To rule out this possibility, we verified the ability of Phillips gene set to classify AAs from GBMs in Phillips dataset itself and also other datasets, TCGA and GSE4422. In this analysis, we used all the 8 genes as the expression data was available for all the genes.

i) Classification analysis of Phillips gene set in Phillips dataset.
In this dataset, the sensitivity of prediction of AA was only 50% (12 out of 24 AAs were rightly identified) and that for GBM was 93.4% (71 of 76 GBMs are rightly classified) Figure S8A and S8B). The specificity of the signature for AA was 93.4% and for GBM it was 50%. Thus it appears that the 8-gene signature of Phillips was not able to accurately classify the samples of Phillips dataset into AA and GBM. The low classification sensitivity for AAs reiterates the fact that the Phillips gene signature is meant for identifying the prognostic classes, but not meant for distinguishing AA and GBM.

ii) Classification analysis of Phillips gene set in GSE4422 dataset.
We applied Phillips 8-gene signature in GSE4422 (Supplementary Figure S9A and S9B). It is interesting to note that, in this dataset, the Phillips-gene signature could not predict any of the AA samples as AA suggesting that the prediction sensitivity was 0% for AA samples of this dataset and that for GBM was 100%. The specificity of the signature was 100% for AA and 0% for GBM.
iii) Classification analysis of Phillips gene set in TCGA dataset.
Next, we applied Phillips 8-gene signature to TCGA dataset (Supplementary Figure   S10A and S10B). In this dataset, the classification accuracy for AA was 85% (23 out of 27 samples were rightly classified) and the specificity for AA was 100%. In case of GBMs, the sensitivity was 100% and the specificity of prediction was 85%.
Overall, our analyses of Phillips gene-signature across various datasets suggest that the Phillips gene signature fails to consistently predict AA and GBM samples with high accuracy.
The sensitivity of the signature for AA varies greatly across datasets: 0% in GSE4422 dataset; 44% in our dataset, 50% in Phillips dataset and 85% in TCGA dataset. In particular, the Phillips gene signature fails to predict AA samples accurately, thus compromising the sensitivity for AA prediction and specificity for GBM prediction. A possible reason for the inability of Phillips gene signature to classify AAs from GBMs is that the signature was not developed for this purpose.