Molecular tumor classification using DNA methylome analysis

Abstract

Tumor classifiers based on molecular patterns promise to define and reliably classify tumor entities. The high tissue- and cell type-specificity of DNA methylation, as well as its high stability, makes DNA methylation an ideal choice for the development of tumor classifiers. Herein, we review existing tumor classifiers using DNA methylome analysis and will provide an overview on their emerging impact on cancer classification, the detection of novel cancer subentities and patient stratification with a focus on brain tumors, sarcomas and hematopoietic malignancies. Furthermore, we provide an outlook on the enormous potential of DNA methylome analysis to complement classical histopathological and genetic diagnostics, including the emerging field of epigenomic analysis in liquid biopsies.

Background and Underlying Biology

Epigenetic modifications act in concert with transcription factors to regulate and fine-tune tissue and cell type-specific gene expression patterns (1). During development and in tissue regeneration, epigenetic patterns control the differentiation process of individual cells from immature stem and progenitor cells to terminally differentiated cell types with highly specialized functions (2). While epigenetic patterns are faithfully transmitted to daughter cells during mitosis, it is well documented that environmental stimuli and the simple passing of time confer specific changes to the epigenetic landscape of each individual cell. As a result, the epigenetic landscape reflects the cell’s developmental stage, the cell type, the cell’s history of exposure to environmental stimuli and its age (3,4).

Enzymes that establish, detect and modify epigenetic marks (the so-called epigenetic writers, readers and editors) are frequent and recurrent targets for genetic mutations in cancer (5,6). As a consequence of these mutations, the epigenomic landscape of the affected cells is altered, which contributes to the process of malignant transformation, although the detailed underlying molecular mechanisms are largely unknown. Of note, cancer-specific epigenomic alterations occur in the context of normal development- and differentiation-associated epigenomic changes (7–9). This explains why the same mutation can have different effects depending on the tissue context or the differentiation stage where it is acquired.

DNA methylation is an epigenetic mark that regulates genome interpretation during development, differentiation and disease (10). DNA methyltransferases catalyze the covalent attachment of a methyl group to the 5′ carbon of a cytosine base. In eukaryotes, the resulting 5-methylcytosine (5mC) is mainly found in the context of CpG dinucleotides, although a minor fraction of non-CpG methylation has been reported in some cell types (11,12). DNA methylation patterns of tumor cells or tissues can serve as powerful biomarkers for diagnosis, disease sub-classification, patient stratification and for the prediction of response to treatment (8,13–19).

Here, we review the use of DNA methylome analysis for cancer diagnosis, disease sub-classification and patient stratification. We will further discuss how DNA methylome analysis of cell-free tumor DNA derived from blood might serve as a novel universal biomarker for early diagnosis and monitoring of treatment response in the near future.

Development of Genomic Classifier Models for Potential Diagnostic Purposes from the Statisticians’ Perspective

Most published examples using DNA methylation patterns for the diagnostic classification of malignant tumors are based on Infinium methylation arrays. While this technology allows simultaneous, robust, reproducible and rapid measurements of DNA methylation of hundreds of thousands of CpG positions across the human genome, it is not straightforward to use this information for classification purposes. In the following, we will discuss the most important aspects to consider in this context from a statistics perspective.

The minimal requirement for a classification model that is applied diagnostically to guide clinicians in decision-making is that the classifier, in addition to a predicted class, also generates a ‘well-calibrated’ probability as confidence measure for the prediction (20–23). Besides giving confidence in the prediction, such a probabilistic output allows for easy comparison of different classifier models and might also be of importance for standardization of clinical trials that apply classifiers for stratification. A probabilistic classifier score is called ‘well-calibrated’ if for cases predicted to be in a class with the probability P, the actual proportion of samples truly belonging to this class is equal to P. Like other classifier performance-metrics, the calibration of probability outputs can be assessed on independent test data. Many classifier algorithms provide some score for the classification. For example, for Random Forests (RFs), the number of trees in the ensemble voting for a class is already a score that measures the confidence in the prediction. Like a probability score, it also sums up to one when taking the sum over all classes. However, this score is likely not ‘well-calibrated’, and often the observed score distributions between different classes are not comparable (14). ‘Well-calibrated’ scores can be achieved through various model-updating/post-processing strategies as well as fitting logistic models like the famous Platt scaling (24–26) to map raw classifier scores to a more ‘well-calibrated’ probability scale (27). Measures to evaluate the performance of a probability output are for example the Brier score, the logarithmic loss function or by visualization using calibration plots (21,22).

Further typical statistical problems faced when training classification models on high throughput molecular data for cancer classification are the high-dimensionality of the data, the imbalanced class sizes due to common and rare cancer entities and the high number of classes that is expected to increase as research in personalized oncology progresses (21). When training a classification model on high-dimensional data, it is expected that this model is somehow overfitting the data and may not generalize to independent data. This is why a proper validation is needed to assess if the classifier generalizes to new, unseen data and to estimate classifier performance metrics (28). Typical validation strategies are either to simply split the data into training and test data sets or to use more complex approaches for unbiased prediction error estimation like, e.g. cross-validation or other resampling methods (28). The most common pitfall when using cross-validation and resampling methods is that not all of the model building steps are performed within in each fold of the cross-validation, which might lead to an over optimistic prediction error estimate (29).

The extent to which classification models suffer from imbalanced classes depends on the classification algorithm chosen. For example, a RF is known to perform poorly for the minority classes when trained on extremely imbalanced data (30). To deal with such imbalanced class sizes, many different methods such as random under- or oversampling have been developed; however, these techniques are often limited to binary classification tasks (31,32). It is important to keep this potential problem in mind, when training classifiers for personalized oncology where a high number of imbalanced classes are to be expected. Finally, to tackle the replication crisis, many scientific journals already demand that authors share their data as well as their analysis code on public online platforms like GitHub.com or Bitbucket.org. This is even more important when developing genomic classifier models for clinical decision-making, especially, as most machine learning algorithms train black-box models for which a direct biological interpretation of resulting predictions is often not possible. Sharing code used to train and validate such classifiers fosters the acceptance by the community and allows to easily train competing models that might achieve better prediction performance (17,10).

Tissue and Tumor Entity-Focused Approaches

In recent years, tissue and tumor entity-focused DNA methylation classifiers have been developed, and these classifiers are now starting to be incorporated into diagnostic work-up and into clinical decision processes. Here, we discuss successful examples that showcase the clinical utility of DNA methylome analysis for diagnosis, disease sub-classification and patient stratification.

Brain tumor classifier

DNA methylation profiling is emerging as the state of the art for molecular classification of central nervous system (CNS) tumors (14,21,33). The complexity and diversity of brain tissues is reflected by the approximately 80 biologically and clinically distinct tumor entities listed in the current World Health Organization (WHO) classification of CNS tumors (34).

Over the last decade, many studies showed robust and reproducible stratification of previously considered homogenous brain tumor entities into clinically relevant DNA methylation classes (Fig. 1). Among others, these entities comprise medulloblastoma (35), glioblastoma (36), astrocytoma (37,38), ependymoma (39,40), choroid plexus tumors (41), atypical teratoid/rhabdoid tumors (42), craniopharyngiomas (43) and adult-type anaplastic glioma (44). These findings ultimately lead to the development of a large reference cohort of 2801 DNA methylation array profiles covering 82 CNS tumor methylation classes mostly reflecting entities or sub-entities defined by the WHO classification and 9 control tissue classes on which a RF classifier was trained and validated (14,21,33). Statistical cross-validation estimated a high prediction performance. The classifier predictions and the accompanying prediction scores are in strong agreement between samples processed by different laboratories, between different DNA methylation array types (EPIC/450k) as well as between array profiles and DNA methylation profiles measured by whole-genome bisulfite sequencing. Applying this classifier to a prospective cohort of over 1000 samples resulted in a revised diagnosis in about 12% of cases, highlighting the improvement in diagnostic accuracy.

Figure 1

UMAP non-linear dimension reduction of 14 999 DNA methylation array profiles comprising 5915 CNS tumors collected by the molecularneuropathology.org website and 9084 profiles from the Cancer Genome Atlas Project (TCGA). Highlighted are representative CNS and TCGA tumor entities as well as TCGA control samples to demonstrate the specificity of DNA methylome profiles for normal tissues as well as for distinct tumor entities.

In addition to classifier predictions, the Infinium methylation array allows to detect copy number variation profiles (conumee R-package) and MGMT promotor methylation status (45). To share the RF classifier with the community, a freely available website was developed (molecularneuropathology.org). Upon upload of methylation array raw data, this platform automatically reports classifier predictions, copy number variation profiles and MGMT methylation status. Users may contribute their data for further development of the classifier, which up to now, resulted in >30 000 DNA methylation profiles shared. This continuously growing cohort is monitored by applying non-linear dimension reduction methods that identified a very rare, aggressive spinal ependymoma entity that is characterized by MYCN amplification (46–48). This example showcases the potential of website-based, community-driven data-sharing to improve cancer diagnostics.

Recent progress in DNA methylation-based brain tumor classification leads to a further refinement of already known DNA methylation classes for ependymomas (46,49,50), medulloblastoma (51), diffuse leptomeningeal glioneuronal tumors (52), anaplastic astrocytoma (53) as well as meningioma (15) and neuroblastoma (54). These and other DNA methylation subclasses will be included in the upcoming new version of the brain tumor classifier (v12). As the number of these subclasses is constantly increasing, a DNA methylation-based molecular taxonomy of CNS tumors with several hierarchical levels combining subclasses to classes and family classes will be established. Such a molecular taxonomy that is aligned as much as possible to the WHO classification of CNS tumors is needed to translate RF predictions into clinical practice where the most detailed information about molecular subclasses might not yet have practical implications.

Sarcoma classifier

Sarcomas are a rare, clinically and morphologically heterogeneous group of malignant soft tissue and bone tumors that occur across all age groups (55). The diagnosis of sarcomas is often challenging, which is reflected by a high inter-observer variability among pathologists as well as between primary institutions and specialized referral centers (56,57). The detection of gene fusions resulting from genomic translocation events might help here; however, for almost 50% of sarcomas, well-defined molecular markers are lacking (55,58). DNA methylome profiling is emerging as a promising method to overcome these diagnostic limitations. For sarcomas with small blue round cell histology, DNA methylation profiling has been successfully applied to classify these diagnostically challenging tumors (59). Comparable studies showed the diagnostic potential of DNA methylation profiling for the classification of high-grade soft tissue sarcomas (60), rhabdomyosarcomas (61), nerve sheath tumors (62), bone sarcomas (63), spindle cell sarcomas (64) and angiosarcomas (65). Applying the blueprint of a machine learning workflow that was originally developed for the classification of CNS tumors (14,21), a comparable model for the classification of 47 soft tissue and bone sarcoma entities has been established (Koelsche et al., in revision). Like the so-called ‘Heidelberg brain tumor classifier’, this sarcoma classification model is publicly available through a web platform (molecularsarcomapathology.org), and its results can be used to guide differential diagnosis of sarcomas (66). This website also offers users to contribute their data to develop more advanced sarcoma classifiers in the future.

Lymphoid malignancies

Chronic lymphocytic leukemia (CLL) is a B cell neoplasm that is characterized by a highly heterogeneous clinical course ranging from patients who have stable disease over many years to patients with rapidly progressive disease (67). Epigenetic alterations in CLL are known for more than 20 years (68–78). Several recent studies investigated DNA methylation across selected B cell developmental stages obtained from healthy donors and detected massive DNA methylation programming that affects up to 30% of the DNA methylome (7,8,79,80). Subsequently, it was shown that the malignant B cells in CLL patients are derived from a continuum of possible differentiation stages through which healthy B cells transition during normal B cell differentiation (8,9). Furthermore, it was demonstrated that the degree of maturation achieved in the cell-of-origin of CLL closely associates with an increasingly favorable clinical outcome (8,80). Of interest, these studies uncovered that nearly all DNA methylation differences previously reported between tumor/normal or between defined disease subtypes are naturally present in healthy, non-malignant B cells (8,9,79,80). Together, these studies provided proof-of-principle that the epigenomes of CLL patients consist of two different components: (1) the epigenetic imprint of the cell-of-origin of CLL, and (2) the disease-specific epigenetic aberrations acquired during disease pathogenesis and disease progression (9). This information provides useful to identify tumor-specific changes in cell biology that could be used as targets to develop novel therapeutic approaches for CLL (9). Similar principles have been shown to apply to other B cell neoplasms like, e.g. B cell acute lymphoblastic leukemia, diffuse-large B cell lymphoma, mantle cell lymphoma and multiple myeloma (7,79,81). Very recently, a pre-print described the combined use of information of the epigenetic imprint of the cell-of-origin and of disease-specific epigenetic aberrations to train a DNA methylation classifier that identifies 14 different B cell neoplasms and clinically important disease subtypes (82). If this classifier proves to perform confidently also for independent samples, it may provide additional, independent information for the differential diagnosis of B cell neoplasm, especially in unclear diagnostic situations.

Myeloid malignancies

Myeloid malignancies have been studied in great detail at the genomic, transcriptomic and epigenomic level during the last two decades (16,83–88). While results from genomic studies have transformed the classification of myeloid malignancies, epigenomic studies have yet to demonstrate their added value as clinically relevant biomarkers for the classification of myeloid malignancies (89,90). One successful example has been published by Meldi and colleagues, who examined which patients suffering from chronic myelomonocytic leukemia are likely to respond to therapy with the demethylating agent decitabine. Using samples taken at the time of diagnosis, they developed an epigenetic classifier that confidently predicts response to decitabine treatment (91). In addition, recent studies using DNA methylome profiling identified prognostically distinct epigenetic subgroups in patients with myelodysplastic syndromes or chronic myelomonocytic leukemia that associate with clinical, biological and genetic features (92,93). To the best of our knowledge, none of these biomarkers has been implemented into a routine clinical workflow yet.

For juvenile myelomonocytic leukemia (JMML), epigenetic biomarkers have just recently started to be used in clinical practice. Research groups from Europe, the United States and Japan independently reported DNA methylation patterns as a predictive biomarker for outcome in JMML (13,17,19). Subsequently, a machine learning classifier (XGBoost) for the prediction of JMML subclasses was trained based on a consensus definition of DNA methylation subgroups (Schönung et al., under review). This classifier has been validated in an independent patient cohort and demonstrated that DNA methylation subgroups are the single most predictive biomarker for outcome in JMML. This classifier is now used by clinicians in combination with other parameters to guide risk-adapted clinical management of JMML patients.

Together, these examples from different tumor entities show that the era of DNA methylation biomarkers is just about to start, but they also highlight the potential of DNA methylome analysis for cancer diagnosis, identification of disease subclasses and as prognostic/predictive biomarkers.

DNA Methylation Analysis for Early Diagnosis and Disease Monitoring using Liquid Biopsies

In the era of personalized medicine, the identification of tumor-specific genetic and epigenetic aberrations plays an increasingly important role. The detection of potential molecular targets using tumor DNA and RNA is an integral part of the development of targeted and individualized therapeutic concepts. It is not uncommon, however, that extraction of tumor material is impossible due to the localization or the health condition of the patient. In addition, it is often not possible to repeatedly obtain tumor samples by invasive procedures solely for the purpose of disease monitoring while the patient is on therapy or during follow-up after therapy. Analysis of ctDNA from blood plasma and other body fluids (so-called liquid biopsy) has gained considerable attention in recent years and it is the subject of intensive research in almost all tumor entities. Liquid Biopsy offers several advantages over conventional biopsy-based methods: (1) it can be performed as part of a routine blood collection; (2) it can be performed at any time and can therefore be used to assess the course of the disease; (3) the ctDNA maps all tumor manifestations in the body and thus allows the characterization of tumor heterogeneity.

In addition to the sensitive detection of tumor mutations, identification of epigenetic patterns stemming from tumor cells has emerged as one of the key-areas of research in this field. Epigenetic biomarkers like tumor-specific DNA methylation patterns are attractive non-invasive, diagnostic targets in clinical oncology due to their stability in liquid biopsies (94). This development is in part fueled by technological advances that have enabled DNA methylome analysis from problematic samples as ctDNA typically is lowly abundant and highly fragmented (95–101). In addition, the tumor-derived DNA methylation patterns of ctDNA are often contaminated to a variable degree by cell-free DNA originating from non-neoplastic cells like, e.g. blood cells or endothelial cells (102,103). Knowledge on the DNA methylation patterns present in the different tissues from which the cell-free DNA originates can support the deconvolution of the epigenetic patterns detected in the cell-free DNA (102,103). Advances in bioinformatics pipelines, the broad availability of reference epigenomes from different tissues as well as the increasing amount of DNA methylomes available from cancer patients spanning also rare cancer entities have improved the identification and detection of tumor-derived epigenetic patterns, and we predict that this will be a dynamically evolving field in the next few years. Targeted, locus-specific epigenetic assays for early detection of cancer or for disease monitoring are already in clinical use (reviewed by Locke (104)). These assays have the advantage that they are easy to standardize and that they can be quickly adopted by routine laboratories. On the downside, most of these tests suffer from relatively poor specificity for a given tumor entity, which makes them unattractive as screening tools (104). Until now, there have been only few published examples of genome-wide DNA methylome analysis in the context of liquid biopsies for the diagnosis or monitoring of malignant tumors. Successful examples include proof-of-principle studies using whole genome bisulfite sequencing of ctDNA-enriched samples (105), immunoprecipitation-based approaches (103) or 450k array analysis (106). The immunoprecipitation-based approach has recently been applied to the detection of renal cancer in blood and urine samples (107). In addition, a combination of targeted capture BS-seq and low-coverage whole-genome bisulfite sequencing was used to identify prostate cancer-specific DNA methylation patterns in blood plasma and allowed the prediction of response to anti-androgen therapy (108).

Outlook

Genome-wide DNA methylation profiling has proven to be a powerful tool that already impacts on diagnostic work-up and patient stratification. This is foremost true in the brain tumor field, but other tumor classes and entities like, e.g. sarcomas and hematopoietic malignancies are about to catch up. It is tempting to speculate that in the near future, sensitive and accurate DNA methylation classifiers will be developed that support the diagnostic work-up and patient stratification at the pan-cancer scale. Such classifiers might one day facilitate pan-cancer screening from liquid biopsies. The development of sensitive and scalable epigenomic analyses from liquid biopsies will hugely benefit from the dynamically evolving field of single-cell epigenomics in the near future.

Acknowledgements

We thank the German Glioma Network and the Neuroonkologische Arbeitsgemeinschaft for sharing their data. In addition, we thank David Capper, David Jones, Volker Hovestadt and Andreas von Deimling for their fundamental work on the original brain tumor classifier.

Conflict of Interest statement. A patent for a ‘DNA-methylation-based method for classifying tumor species of the brain’ has been applied for by the Deutsches Krebsforschungszentrum Stiftung des öffentlichen Rechts and Ruprecht-Karls-Universität Heidelberg (EP 3067432 A1) with S.M.P. and M.S. as inventors. C.P. and D.B.L. have no potential conflicts of interest to report.

Funding

German Cancer Aid, (DKH 70113869 to C.P., DKH 70112574 to D.B.L.); the José Carreras Leukämie Stiftung (DJCLS R 15/01 to C.P. and D.B.L.); the Helmholtz Foundation (to C.P.); the DKFZ-Heidelberg Center for Personalized Oncology (DKFZ-HIPO_036); the German Childhood Cancer Foundation (DKS 2015.01); an Illumina Medical Research Grant, the DKTK joint funding project ‘Next Generation Molecular Diagnostics of Malignant Gliomas’.