DNA methylation-based classification of sinonasal tumors

Abstract

The diagnosis of sinonasal tumors is challenging due to a heterogeneous spectrum of various differential diagnoses as well as poorly defined, disputed entities such as sinonasal undifferentiated carcinomas (SNUCs). In this study, we apply a machine learning algorithm based on DNA methylation patterns to classify sinonasal tumors with clinical-grade reliability. We further show that sinonasal tumors with SNUC morphology are not as undifferentiated as their current terminology suggests but rather reassigned to four distinct molecular classes defined by epigenetic, mutational and proteomic profiles. This includes two classes with neuroendocrine differentiation, characterized by IDH2 or SMARCA4/ARID1A mutations with an overall favorable clinical course, one class composed of highly aggressive SMARCB1-deficient carcinomas and another class with tumors that represent potentially previously misclassified adenoid cystic carcinomas. Our findings can aid in improving the diagnostic classification of sinonasal tumors and could help to change the current perception of SNUCs.

Fig. 1. DNA methylation classes of sinonasal tumors.

T-distributed stochastic neighbor embedding dimensionality reduction showing the 18 different DNA methylation classes. The conventional histopathological diagnosis is annotated by color. ACC sinonasal adenoid cystic carcinoma, ADC sinonasal adenocarcinoma, ALV RMS alveolar rhabdomyosarcoma, ATRT adult pituitary atypical rhabdoid/teratoid tumor, CPH craniopharyngioma, CTRL normal sinonasal control tissue, EWS Ewing’s sarcoma, EMB RMS embryonal rhabdomyosarcoma, GPC sinonasal glomangiopericytoma, LECA lymphoepithelial carcinoma, MCC Merkel-cell carcinoma, MELA sinonasal mucosal melanoma, NEC sinonasal neuroendocrine carcinoma, NUT NUT midline carcinoma, ONB olfactory neuroblastoma, PDCA sinonasal poorly differentiated carcinoma, PIT AD pituitary adenoma, SCC sinonasal squamous cell carcinoma, SNUC sinonasal undifferentiated carcinoma.

Fig. 2. Reclassification of sinonasal undifferentiated carcinoma (SNUC) classes.

A Pie chart showing the conventional histopathological diagnosis of cases from the NEC-like IDH2 class. B Sanger-plot showing an example of an IDH2 c.516 G > T (R172S) mutation. The frequencies of R172S, R172T, R172G and R172K mutations are displayed as a pie chart. A bar chart shows the frequency of R172 mutations, which occurred in all of the investigated cases. C Summary copy number profile of cases from the NEC-like IDH2 class showing highly recurrent copy number alterations such as gain of chromosome 1q as well as loss of chromosome 17p in combination with chromosome 17q gain. D Pie chart showing the conventional histopathological diagnosis of cases from the SMARCB1 class. E Detailed copy number plot of chromosome 22 with focal deletion of the SMARCB1 gene locus and subsequent loss of INI1 expression in immunohistochemistry. F Summary copy number profile of cases from the SMARCB1 class, the SMARCB1 gene locus is highlighted by the arrow. G Pie chart showing the conventional histopathological diagnosis of cases from the ACC class. H Example of histomorphological and molecular evidence for adenoid cystic (ACC) differentiation in form of sharply punched-out areas as well as recurrent MYB fusions. The frequency of these findings is shown as bar charts. I Summary copy number profile of cases from the ACC class. J Pie chart showing the conventional histopathological diagnosis of cases from NEC-like SMARCA4/ARID1A class. K Hematoxylin and eosin stain showing an example of rosette-like features that were recurrently observed in this class. The second tile shows an exemplary immunohistochemical stain for neuron-specific enolase (NSE) with very heterogenous cytoplasmic staining as example of evidence for neuroendocrine marker expression. The frequency of these findings is shown as bar charts. L Summary copy number profile of cases from the NEC-like SMARCA4/ARID1A class. ACC sinonasal adenoid cystic carcinoma, ADC sinonasal adenocarcinoma, ATRT adult sellar atypical teratoid/rhabdoid tumor, NEC sinonasal neuroendocrine carcinoma, ONB olfactory neuroblastoma, PDCA sinonasal poorly differentiated carcinoma, SCC sinonasal squamous cell carcinoma.

Fig. 3. Results from mass spectrometry-based proteomics.

A t-SNE depicting the global proteomic profile correlation between normal sinonasal tissue (CTRL), olfactory neuroblastoma (ONB), sinonasal squamous cell carcinoma (SCC), NEC-like IDH2, NEC-like SMARCA4/ARID1A, and adenoid cystic carcinoma (ACC). ONB, NEC-like IDH2 and NEC-like SMARCA4/ARID1A had an unexpectedly similar proteomic profile. B Differential expression analysis between normal sinonasal tissue and the investigated tumor classes was performed using a moderated t-test followed by Benjamini-Hochberg multiple testing correction. The results are shown as volcano plots. Recurrent and highly differential expressed proteins are annotated, highlighting overexpression of proteins specific for neurons or neuroendocrine cells in the ONB, NEC-like IDH2 and SMARCA4/ARID1A class. C Combined bar and point plots of 59 biologically independent samples showing the expression of Cytokeratin 18 (KRT18) and Ubiquitin carboxy-terminal hydrolase L1 (UCHL1) in the different tumor classes as determined by proteomics and immunohistochemistry (IHC). D Exemplary hematoxylin/eosin and KRT18 and UCHL1 immunohistochemical stainings are shown, validating the results from the proteomics analysis. In conclusion, the combination of these markers could be useful for histopathological classification if DNA methylation is not available or feasible. E Results from overrepresentation analysis comparing the overall similarity of the global protein expression signatures of tumor classes with various normal cell types of the airways. Differentially expressed genes between normal sinonasal tissue and the investigated tumor classes were subjected to overrepresentation analysis using previously published cell type-specific gene sets that were identified using single-cell RNA sequencing data of the mucosal lining of human airways. The Fisher’s exact test followed by FDR multiple testing correction was used to test for significance. A high similarity between pulmonary neuroendocrine cells (‘PNEC’) and the ONB, NEC-like IDH2 and NEC-like SMARCA4/ARID1A was observed, in line with the overexpression of neuronal/neuroendocrine markers shown in B. ACC specimens mostly resembled serous cells of submucosal glands and SCC specimens mostly resemble Squamous Cells 1. ACC adenoid cystic carcinoma, ADC adenocarcinoma, NEC neuroendocrine carcinoma, CTRL normal sinonasal tissue, ONB olfactory neuroblastoma, PDCA poorly differentiated carcinoma, SCC squamous cell carcinoma, FDR False discovery rate, PNEC Pulmonary neuroendocrine cells.

Fig. 4. DNA sequencing results for NEC-like SMARCA4/ARID1A tumors and outcome analysis of sinonasal undifferentiated carcinoma (SNUC) DNA methylation classes.

A The oncoprint plot shows tumor mutation burden (TMB) and recurrent mutations from whole exome sequencing or next generation panel sequencing (NGS). Potential treatment options and their respective evidence levels according to the classification from the Cancer Genome Interpreter website are shown in the panel below. B Overall survival Kaplan–Meier curve comparing disease-specific survival rates in different SNUC DNA methylation classes. The tables below show the number of patients at risk at different time points as well as a pairwise test for significance using the log-rank test. Patients with tumors from the SMARCB1 class had significantly worse survival compared to adenoid cystic carcinoma (ACC; p = 0.004), NEC-like IDH2 (p = 0.012) and NEC-like SMARCA4/ARID1A (p < 0.001) tumors. There was no significant difference between tumors from the NEC-like SMARCA4/ARID1A and ACC (p = 0.163) and NEC-like IDH2 (p = 0.168) class as well as between ACC and NEC-like IDH2 tumors (p = 0.950). NEC neuroendocrine carcinoma.

Fig. 5. Classifier development and application to an independent test set.

A Confusion matrix of the non-sinonasal samples from the test set, split up in different categories. Overall, the classifier achieved a specificity of 0.982, ranging from 0.763 in salivary gland tumors to 1.0 in normal tissue and brain tumors. B Overview of the classification results from the test set. Out of 52 samples, five cases were assigned to the Unknown class, resulting in an outlier detection sensitivity of 0.904. Out of the 47 remaining specimens, DNA methylation-based classification confirmed the initial diagnosis in 39 samples (83%). Eight specimens (17%) were reclassified to a divergent DNA methylation class. Additional molecular workup supported the DNA methylation-based reclassification in all cases. C Confusion matrix from the sinonasal test set shows an accuracy of 1.0 for classification of sinonasal tumors. D Receiver operating characteristic (ROC) curve of the support vector machine classifier with regards to the binary outlier detection problem with an area under the curve (AUC) of 0.9942.

Fig. 6. Summary of important molecular characteristics and clinical parameters of the four sinonasal undifferentiated carcinoma (SNUC) classes, including proposed type names and potential immunohistochemical (IHC) markers.

SNUC Sinonasal undifferentiated carcinoma, Chr Chromosome, UCHL1 Ubiquitin carboxy-terminal hydrolase L1, CK18 Cytokeratin 18.