Practical implementation of DNA methylation and copy-number-based CNS tumor diagnostics: the Heidelberg experience

Prior to presenting and discussing our experience with methylation analysis on more than 2000 diagnostic CNS tumor cases, we would like to briefly outline the concept of the methylation-based classification (the Classifier) and CNV analysis.

The principal output of the Classifier is a list of the predicted class membership probabilities for every class currently included in the Classifier (v11b4: 82 tumor classes, 9 non-tumorous classes). These probabilities are referred to as the “calibrated” Classifier scores. The combined calibrated scores of all 91 methylation classes add up to 1. The automatically generated website report currently only lists calibrated scores above 0.3, but a listing of all scores can also be downloaded from the website. For a valid classification, we have proposed a default threshold value of ≥ 0.9 that has to be reached. As all calibrated scores add up to 1, this implies that the calibrated scores of the remaining 90 classes add up to less than 0.1 in such a classifiable tumor.

Over the last 4 years, the Classifier experienced continuous modification with more tumor entities being recognized by each update. Such an evolution will also continue in future, as our knowledge on novel classes grows. A table of the tumor methylation classes recognized by the current version of the Classifier can be obtained at www.​molecularneuropa​thology.​org.

During cross validation of the Classifier, eight methylation class families were designated consisting of two-to-six individual closely related methylation classes. For cases falling into a methylation class family, two output scores are generated: one “class score” (not different to cases without a family) and a second “family score” representing the sum of the combined scores belonging to a methylation class family. Since the calibrated scores represent probability estimates, it is straightforward to apply the sum rule in probability theory to sum up the individual class probabilities to get a probability estimate for the family. For example, an IDH wild-type (IDH wt) glioblastoma may have a class score of 0.6 for the methylation class GBM, RTK I. The case, additionally, scored with 0.2 for GBM, MES, and 0.07 for GBM, RTK II, 0.04 for GBM, MID, 0.01 for GBM, RTK III, and 0.01 for GBM, MYCN. The GBM family score would then be the sum of all the above class scores (0.93). As this score would be ≥ 0.9, the case is considered classifiable as belonging to the GBM IDH wt family (score has to be ≥ 0.9 as usual). Thus, the introduction of methylation class families allows a single threshold level for all methylation classes and families. If the subclass within a family is of interest (e.g., in IDH gliomas), we use a calibrated cut-off score of 0.5 for prediction of the most likely subclass. Therefore, the case in the above example would be considered a glioblastoma, IDH wt (family score 0.93) of the RTK I subclass (class score 0.6). The subclass cut-off value of 0.5 was chosen arbitrarily. The validity of this cutoff is not easy to assess as for most subclasses DNA methylation profiling is currently the only available method to identify the subclass. The exception to this is 1p/19q codeleted oligodendroglioma that is part of the methylation class family IDH glioma. These tumors can also be identified by the CNV profile. For these, the 0.5 cutoff seems to perform well.

For the determination of the common calibrated score threshold, the trade-off between sensitivity and specificity has to be considered. For the cases of the Classifier reference cohort, the optimal trade-off between sensitivity and specificity (maximization of the Youden index) was reached at a calibrated score of 0.84, maximum specificity was reached at 0.96 [6]. After some non-reference cohort test cases, we decided to implement a threshold of ≥ 0.9 that is in the middle between these two values. However, it should be mentioned that, for many medical tests, the Youden index is preferred, so there is also a rational for using the less conservative cutoff of 0.84 as a threshold.

In our diagnostic experience, calibrated scores between 0.3 and 0.9 will be encountered on a regular basis. If this occurs within the setting of a low-tumor cell content, we may accept a lower score as an indication of a specific diagnosis (see also the below paragraph “DNA methylation analysis of infiltration zone of diffuse gliomas or highly inflamed tumors”). For cases with high tumor cell content and low scores, this seems more problematic. As indicated above, we would likely accept a calibrated score down to 0.84 (maximized Youden index) as valid classification if nothing else strongly speaks against such an interpretation. Scores below 0.5, we would generally discard. For cases with high tumor cell content and scoring between 0.5 and 0.84, it seems problematic to make general recommendations. We would likely see such a score as a suggestion that a case may be in some way related to a certain methylation class and would try to find further evidence of such a relation [e.g., sequencing of BRAF in a case with a methylation class (anaplastic) pleomorphic xanthoastrocytoma (PXA) calibrated score of 0.75].

Genome-wide DNA methylation array data can also be used to perform analysis of copy-number variations (CNV), for example using the ‘conumee' R package in bioconductor [15]. By standard, every interrogated CpG is represented by two probes on the array (one for methylated and one for unmethylated). For the analysis of DNA methylation, the ratio of the intensity signal of the methylated and the sum of the methylated and unmethylated probe intensities are calculated (methylated/(methylated + unmethylated); beta-values). In contrast, for the calculation of CNV, the methylated and unmethylated signal intensities are added together and a ratio is formed against healthy reference samples that have a flat genome. This copy-number ratio is then plotted in a graph according to chromosomal location (Fig. 1a). Areas with high copy-number ratios correspond to areas with a gain of chromosomal material (e.g., by a trisomy or an amplification), areas with low copy-number ratios represent lost DNA (e.g., by a deletion). The results from the CNV analysis can be considered as independent from results of the methylation classifier, and both readouts can independently contribute to the final diagnostic interpretation.

CNV analysis provides a good overview of gross structural alterations in the tumor genome. High-level amplifications (e.g., EGFR, Fig. 1a, e) and homozygous deletions (e.g., CDKN2A/B, Figs. 1a, 4d) are usually obvious when present. Numerical chromosome aberrations [e.g., trisomies (Fig. 1a Chr.7, 19, 20), monosomies, or larger sub-chromosomal gains and losses] are unambiguous to interpret in most cases, although this may become less clear when there is an abundance of changes and/or a low-tumor cell content in the analyzed sample. Even the presence of gene fusion events may be suggested if they are associated with focal duplications or deletions (Fig. 6a–c).

The current CNV generation includes the highlighting of 29 CNS tumor relevant genes marked by a blue circle and blue gene nametags (Fig. 1a). The identical set of 29 genes is highlighted by default in each case for easier identification of copy-number alterations. The highlighting does not indicate relevant changes per se. For the analysis of regions that are not marked by default (e.g., for the identification of genes within an unusual amplified region or deletion), an additional.igv file is automatically generated and can be downloaded from the website (“download complete analysis results”, then in the “cnvp” folder). This can, for example, be visualized with the Integrative Genomics Viewer [37].

In comparison to arrays that probe single-nucleotide polymorphisms (SNP arrays), CNV analysis from DNA methylation arrays cannot rely on allele frequencies to define a copy-number neutral state baseline. Instead, in the current version of the conumee package, the baseline is defined as the line where the median absolute deviation to all data points is minimal. Thus, the baseline is close to the predominant copy-number state of a sample. This represents one limitation of using methylation arrays for CNV analysis, but this is likely to interfere only with cases that show a substantial degree of numerical chromosomal aberrations (e.g., choroid plexus tumors or esthesioneuroblastoma) where identification of the copy-number neutral state may be challenging. Despite this limitation, for most cases, CNV can be easily interpreted, and are of considerable added value in tumor entities that exhibit characteristic chromosomal alterations.

To give an overview of typical chromosomal aberrations of the DNA methylation classes, we have created summary CNV plots of all classes included in the Classifier (Supplementary File 1). Selected summary CNV plots are further demonstrated in the main manuscript. It is important to note that these summary CNV plots have to be read differently in comparison to single-case CNV plots. The y-axis of the summary plots does not contain information on the intensity of signal that is related to the amount of DNA material in each sample, but gives the % frequency of changes at that location either as gain or loss (irrespective of whether, for example, a gain is a single-copy gain or an amplification) of many combined cases. Thus, both, the focal duplication of BRAF on Chr.7 as part of a BRAF fusion in pilocytic astrocytomas (Fig. 5a) and the high-level amplification of C19MC on Chr.19 in embryonal tumors with multilayered rosettes (ETMR; Fig. 13f), look relatively the same in this depiction.

For simplification, we defined CNV alterations as either simple chromosomal changes (numeric chromosomal changes, e.g., a trisomy or monosomy, and whole chromosome-arm loss or gain) or complex chromosomal changes (breakage within chromosomal arm, focal loss, focal gain, amplifications, and chromothripsis) throughout the text.

The status for MGMT promoter methylation can be derived directly from the array data. We are employing an approach previously published [2] [1] which has been adapted for the EPIC (850k) array platform. Based on this algorithm which is also underlying the MGMT promoter methylation analysis provided by our webpage for uploaded cases, we receive the readouts “methylated”, unmethylated, or “not determinable”. DNA from cases with the classifier-based result “not determinable” is subjected to MGMT pyrosequencing frequently resolving these cases. The classifier results “methylated” and “unmethylated” show a very high concordance with MGMT pyrosequencing results [2].

Before we discuss our experience with individual tumor classes, we provide a description of how we approach the data as soon as the reports have been generated by the classifying algorithm (www.​molecularneuropa​thology.​org). These steps are general quality assessment steps that we have found valuable for the detection of handling errors that may occur during the process of array preparation and data upload.

In the following paragraphs, we will briefly discuss all entities presented in the WHO classification of CNS tumors [28] in the order which they appear in the blue book in relation to the corresponding methylation classes as well as the most frequent copy-number variants.

The Classifier recognizes tumors with an IDH mutation-associated CIMP phenotype, and thus both IDH1- and IDH2-mutated tumors are detected independent of the mutation type (including variants not detected by, for example, the IDH1 R132H antibody). Because of their close relation, a methylation class family for IDH mutant gliomas was introduced (“IDH glioma”). This family includes the methylation subclasses astrocytoma (A IDH), high-grade astrocytoma (A IDH, HG), and 1p/19q codeleted oligodendroglioma (O IDH). The association of the DNA methylation class IDH glioma and the presence of an IDH mutation in a glioma is so strong that, in our experience, this methylation class can be interpreted as proof of IDH mutation status.

The WHO diagnosis diffuse astrocytoma IDH mutant more or less completely falls into the methylation class “A IDH”. The CNV pattern in “A IDH” is not highly characteristic (Fig. 2a). Anaplastic astrocytoma WHO grade III partially fall into the “A IDH” and partially into the “A IDH, HG” methylation subclasses. Of note, the “A IDH, HG” group also contains most of the IDH mutant glioblastomas. The CNV in “A IDH, HG” are comparable to those in “A IDH” but with frequently a higher number of total changes and also a higher degree of complex changes and frequent loss on Chr.9p including the CDKN2A/B locus (Fig. 2b). However, recent data provide evidence that grading of IDH mutant astrocytoma by assessing copy-number alterations is more powerful than grading by DNA methylation analysis only [44].

Oligodendroglioma and anaplastic oligodendroglioma, IDH mutant and 1p/19q codeleted are generally classified as methylation subclass “O IDH”. CNV of “O IDH” exhibits 1p/19q codeletion in all cases. The second most frequent alteration is loss of Chr.4p/q (Fig. 2c).

Oligoastrocytomas by WHO definition [28] are considered NOS and the Classifier typically either scores these tumors as “A IDH”, “A IDH, HG”, or “O IDH”. We have so far analyzed one case of a dual genotype astrocytoma/oligodendroglioma [16, 53] and, indeed, observed a different methylation class of A IDH and O IDH, respectively, in different macrodissected tumor areas and an exclusive 1p/19q codeletion in the oligodendroglial tumor regions. Currently, we believe this to be a very rare genetic constellation.

The diagnosis of diffuse astrocytoma, IDH wt and anaplastic astrocytoma, and IDH wt has been severely questioned by several recent publications [4, 35, 47, 52]. Consensus of the critique is the demonstration of mutational profiles in these tumors that closely match those of glioblastoma (or, rarely, a more pediatric-type diffuse glioma). In concordance with this, we were not able to establish a separate comprehensive IDH wt astrocytoma methylation class and the vast majority of such tumors fall into other DNA methylation classes. Most of these cases resolve as glioblastomas, and frequently, the copy-number pattern can be used to further strengthen this classification. Another group of cases may resolve as a tumor class recently described as “anaplastic astrocytoma with piloid features” that is further described in the paragraph “Other astrocytic tumors” [34].

WHO Glioblastoma, IDH wt is currently represented by seven classes in the Classifier. One (GBM, G34) is characterized by the presence of a G34 H3.3 histone mutation and has previously been defined [23, 46]. The other 6 are more closely related and were grouped into a methylation class family (“Glioblastoma, IDH wt”; members: “GBM, RTK I”, “GBM, RTK II”, “GBM, RTK III”, “GBM, MES”, “GBM, MID”, “GBM, MYCN”). The eighth related class “DMG, K27” directly translates to the diffuse midline glioma, H3 K27 mutant of the WHO classification. However, in cases where this is of relevance, H3F3A, HIST1H3B, and HIST1H3C mutations should be confirmed by other methods, because both methylation classes “GBM, G34” and “DMG, K27” contain a small fraction of tumors without apparent mutations in these three genes. Giant cell glioblastoma and gliosarcoma associate with various members of the methylation class family “Glioblastoma, IDH wt” and do not currently seem to represent distinct molecular entities. Figure 3a–h shows the summary plot for these eight classes. Epithelioid glioblastomas are more problematic. A recent DNA methylation-based analysis of 64 epithelioid glioblastomas demonstrated that they stratify into three molecularly and clinically distinct groups: PXA-like tumors with favorable prognosis, IDH wt glioblastoma-like tumors with poor prognosis, and RTK I pediatric glioblastoma-like neoplasms with intermediate prognosis [24].

Pilocytic astrocytoma is represented by a methylation class family with three subclasses related to tumor location. Of these “low-grade glioma, subclass posterior fossa pilocytic astrocytoma” is a relatively pure pilocytic astrocytoma class, “subclass midline pilocytic astrocytoma” is mostly composed of pilocytic astrocytoma but may also incorporate some cases with appearance of the pilomyxoid variant, whereas “subclass hemispheric pilocytic astrocytoma and ganglioglioma” includes cases of supratentorial pilocytic astrocytoma and cases with additional focal or more widespread ganglioglioma differentiation. Duplication of the BRAF locus is frequent among pilocytic astrocytomas and can be observed as a focal low-level gain indicative for a duplication on Chr.7q (Fig. 6a). The frequency of this duplication varies between the three subclasses (Fig. 5a–c).

A methylation class closely resembling anaplastic pilocytic astrocytoma has recently been described in detail under the name “anaplastic astrocytoma with piloid features” [34] and this term is used throughout this manuscript. The copy-number pattern of these tumors occasionally also demonstrates BRAF duplications but much more frequently harbor CDKN2A/B deletions and further chromosomal changes (Fig. 5d).

Both pleomorphic xanthoastrocytoma and anaplastic pleomorphic xanthoastrocytoma are combined in the “(anaplastic) pleomorphic xanthoastrocytoma” methylation class. The presence of a BRAF V600E mutation seen in approximately 70% of pleomorphic xanthoastrocytomas is not a requirement for classification of these tumors into this class. An established but usually underestimated feature of pleomorphic xanthoastrocytoma is homozygous deletion of CDKN2A previously reported in 50% [51]. Among cases falling into this methylation, class CDKN2A/B deletions are seen in around 70% of cases (Fig. 5e).

Subependymal giant cell astrocytoma completely matches the “low-grade glioma, subependymal giant cell astrocytoma” methylation class. The Classifier does not differentiate between TSC1- and TSC2-mutated (or TSC wild-type) tumors, and CNV are rare in this group (Fig. 5f).

We find considerable differences between WHO and classifier-diagnosed ependymomas. A comprehensive analysis on methylation-based classification of ependymomas has previously been published [32]. The Classifier reproduces and refines the findings of that study. A feature pronounced in ependymomas is the fact that tumors at different sites seem to represent distinct entities despite a similar histology.

Subependymoma WHO grade I is generally classified according to location into “subependymoma, supratentorial”, “subependymoma, posterior fossa”, and “subependymoma, spine”. The CNV plots of the three subependymoma methylation classes show infrequent chromosomal gains or losses (Fig. 7a, b; not shown for spinal tumors).

Myxopapillary ependymoma WHO grade I corresponds to the “ependymoma, myxopapillary” methylation class, and “ependymoma, myxopapillary”-assigned tumors exhibit very lively CNV plots in terms of whole chromosome changes. Most tumors exhibit multiple CNV with gain of Chr.5, 7, 9, 16, and 18 being most frequent (Fig. 7c).

The majority of the spinal ependymomas WHO grade II fall within the “ependymoma, spine” methylation class. CNV in “ependymoma, spine” are also frequent, with the dominating alteration being chromosome 22 loss (Fig. 7d). Typically, these tumors are of cervical, less frequently of thoracic and very rarely of lumbar localization.

The WHO variants of papillary, clear cell, and tanycytic ependymoma currently do not form a distinct methylation class but rather fall into the existing methylation classes. Ependymomas of the posterior fossa are clearly distinct from supratentorial ependymomas. Both, infratentorial and supratentorial ependymomas are further sub-grouped with clear clinical impact: Posterior fossa ependymomas form two major classes, “ependymoma, posterior fossa group A” and “ependymoma, posterior fossa group B” [32], which previously have also been identified by transcriptional profiling [54]. The two classes have distinct clinical characteristics and a diagnostic separation of these two classes seems highly relevant. As previously indicated, the CNV profile of “posterior fossa group A” tumors appears rather quiet. The most evident alteration is gain of 1q (Fig. 7e). The clinically more benign tumors in the “ependymoma, posterior fossa group B” methylation class exhibit substantially more CNV (Fig. 7f).

The majority of supratentorial WHO grade II and grade III ependymomas fall into the frequent “ependymoma, RELA fusion” or the exceedingly rare “ependymoma, YAP1 fusion” methylation classes. As the RELA fusion can also be identified by FISH or immunohistochemistry, ependymoma, RELA fusion positive has already been introduced to the 2016 update of the WHO classification. As described in the original publication [33], the cases frequently harbor chromothripsis involving Chr.11q13. This can occasionally be identified in the CNV of these tumors with short sections of gains and losses close to the centromere of Chr.11, sometimes, extending further across Chr.11 [and thus resulting in the slightly jagged line of the summary CNV plot (Fig. 7g)]. The “ependymoma, YAP1 fusion” tumors on average have fewer alterations, with loss close to the YAP1 locus on Chr.11q and loss of Chr.22 being most frequent (Fig. 7h).

Of note, methylation analysis of the ependymal tumor group provides a completely reshuffled classification system which, in these tumors, obviates the need for further grading as the individual methylation classes already identify patient groups much more homogenous in respect to prognosis than the previous WHO classification and grading was able to provide. Recent data suggest further subdivision of the methylation classes with likely more distinct ependymoma classes to be described in the future.

The chordoid glioma of the third ventricle shows complete overlap with the methylation class “chordoid glioma of the third ventricle” and all cases so far characterized had a flat CNV profile (Fig. 8a). All cases investigated also harbored the recently published PRKCA hotspot mutation [11].

Angiocentric gliomas are typically classified as “low-grade glioma, MYB/MYBL1” methylation class, however, also other low-grade tumors without prominent angiocentric growth pattern occasionally fall into this class. The CNV plot is typically either flat or shows alterations on Chr.6q with a break point classically within the MYB gene in around 30% of cases (Fig. 8b).

Astroblastoma is a WHO diagnosis with a high degree of inter-observer variation. By methylation analysis, a substantial part of astroblastomas are classified as “CNS high-grade neuroepithelial tumor with MN1 alteration” [45]. Likely, the remaining cases are not a single tumor entity but rather harbor genetic alterations that suggest classification as other tumor entities [55]. The summary CNV of “CNS high-grade neuroepithelial tumor with MN1 alteration” is shown in Fig. 8c.

The three WHO entities of choroid plexus tumors are combined to a methylation class family. Three subclasses exist, “plexus tumor, subclass adult”, “plexus tumor, subclass pediatric A”, and “plexus tumor, subclass pediatric B” that broadly correspond to the three previously defined methylation clusters [50]. “Plexus tumor, subclass adult” predominantly contains adolescent and adult patients with choroid plexus papilloma and a fraction of approximately 15% of atypical plexus papilloma. “Plexus tumor, subclass pediatric A” is composed half each of atypical choroid plexus papilloma and choroid plexus papilloma, and they mostly occur in pediatric patients. “Plexus tumor, subclass pediatric B” incorporates plexus carcinomas and approximately 30% of tumors diagnosed as atypical plexus papilloma. The tumors show numerous non-complex chromosomal changes with different frequency patterns (Fig. 8d, e, f). Especially, the frequent loss of Chr.21 in “plexus tumor, subclass adult” in comparison to “plexus tumor, subclass pediatric A” is noteworthy. In contrast to the other two classes, “Plexus tumor, subclass pediatric B” is characterized by a higher number of chromosomal losses. The exact relation of these three methylation classes with TP53 somatic/germline status has not yet been resolved.

The majority of cases with DNT histology are classified as “low-grade glioma, dysembryoplastic neuroepithelial tumor” methylation class. In our hands, this can be useful for small tumor samples not exhibiting all the histological hallmarks of DNT. With respect to CNV patterns, “low-grade glioma, dysembryoplastic neuroepithelial tumor” is inconspicuous (Supplementary file 1). Interestingly, the proportional distribution of histologically diagnosed DNTs and gangliogliomas shows marked geographical variability across surgical series [49]. The exact reason for this remains unexplained but likely the regionally different interpretation of histological criteria plays a major factor.

Within the Classifier, “ganglioglioma” is mostly classified either as “low-grade glioma, ganglioglioma” or as the above-described “subclass hemispheric pilocytic astrocytoma and ganglioglioma”. The exact relation of these two groups is currently unclear. Cases with lower tumor cell content frequently are non-classifiable and often get elevated scores for “control tissue, reactive tumor microenvironment”. Whole chromosomal gains are occasionally observed, in particular of Chr.7 (Fig. 9a).

Anaplastic ganglioglioma is not included as a separate methylation class in the Classifier, suggesting that the histological diagnostic criteria currently in use are not sufficient to identify a tumor group with a common methylation pattern. The majority of histologically defined anaplastic gangliogliomas either fall into the PXA or into the GBM, IDH wt classes.

Desmoplastic infantile astrocytoma (DIA) and desmoplastic infantile ganglioglioma (DIG) form the methylation class “low-grade glioma, desmoplastic infantile astrocytoma/ganglioglioma”. CNV profiles show no recurrent chromosomal imbalances (Supplementary file 1), while initial data suggest that BRAF mutations may be relatively common in this group [20].

The WHO entity of rosette forming glioneuronal tumor constitutes the methylation class “low-grade glioma, rosette forming glioneuronal tumor”. Chromosomal changes are rare (Supplementary file 1). Likewise, the recently recognized diffuse leptomeningeal glioneuronal tumor also forms the clearly demarcated “diffuse leptomeningeal glioneuronal tumor” methylation class. The copy-number pattern of these tumors is rather specific and frequently demonstrates focal 7q34 gain indicating BRAF fusion in combination with 1p deletion and occasionally additional 19q deletion (Fig. 9b). Very recently, a series of molecularly defined diffuse leptomeningeal glioneuronal tumors demonstrated two sub-groups with distinct clinical and genetic features among these tumors [8]. These sub-groups are currently not separately detected by the Classifier but are expected to be included in a future update.

Central neurocytoma and atypical central neurocytoma are classified as “central neurocytoma” methylation class. Separation between these two is not possible by methylation analysis. In contrast, the few extraventricular neurocytomas which we had the chance to analyze were either non-classifiable or fell into various methylation classes but never to “central neurocytoma”. We take this a caveat for assuming that these tumors are simply extraventricular manifestations of central neurocytoma. Further research is needed to define the identity of these tumors. CNV profiles show no recurrent chromosomal imbalances (Supplementary file 1).

Cerebellar liponeurocytomas again form a unique methylation class termed “cerebellar liponeurocytoma”. The CNV profile indicates recurrent chromosomal losses especially of regions on Chr.2p and Chr.14 (Fig. 9c).

In particular, familial paraganglioma may carry germline SDH mutations associated with a CpG island methylator phenotype (CIMP) [27]. This group is currently not recognized by the Classifier due to a lack of numbers, but preliminary data reveal clear differences in their methylation pattern. Paraganglioma without SDH mutations constitute the vast majority of sporadic tumors form the “paraganglioma, spinal non-CIMP” methylation class. CNV profiles show no recurrent chromosomal imbalances (Supplementary file 1).

Gangliocytoma, dysplastic cerebellar gangliocytoma (Lhermitte–Duclos disease) and papillary glioneuronal tumors currently are not recognized by methylation-based classification, as there were too few cases to build up such classes.

Pineocytoma appears to exhibit a methylation profile very close to that of normal pineal gland, therefore, not allowing the construction of a distinct methylation class. Pineocytomas are thus mostly non-classifiable or may even be classified as “control tissue, pineal gland” methylation class.

Pineal parenchymal tumor of intermediate differentiation is represented by the methylation class “Pineal parenchymal tumor”. CNV patterns in this group are not characteristic (Fig. 11a). This methylation class also contains some cases histologically diagnosed as pineoblastoma likely reflecting the continuous gradient of morphology between these lesions.

Two separate methylation classes are defined for pineoblastoma: the more frequent “pineoblastoma group B” and the rare “pineoblastoma group A/intracranial retinoblastoma”. “pineoblastoma group B” also contains some pineal parenchymal tumors of intermediate differentiation, while “pineoblastoma group A/intracranial retinoblastoma” contains some tumors diagnosed as retinoblastoma. We assume these retinoblastomas to be genetically related to the so-called trilateral retinoblastomas detected in approximately 5% of patients with hereditary retinoblastoma who, in addition to bilateral ocular tumors, develop a pineal tumor with morphological characteristics of pineoblastoma [7]. Typical CNV in “pineoblastoma group A/intracranial retinoblastoma” are gain of Chr.1q, gain of 6p and loss of 16q (Fig. 11b), while “pineoblastoma group B” shows frequent gain of Chr.7, 12, 14, 17, and 19q and a characteristic focal loss on Chr.5p (Fig. 11c).

Two methylation classes (A and B) are defined for papillary tumor of the pineal region. “Papillary tumor of the pineal region group A” exhibits a high number of CNV with gains of 4, 5, 11, and 12 and loss of Chr.10 being most frequent (Fig. 11d). The CNV profile of “papillary tumor of the pineal region group B” also shows characteristic changes with all cases showing loss of Chr.10 among other changes (Fig. 11e). A more aggressive clinical course has been suggested for “papillary tumor of the pineal region group B”, but data on this is still preliminary [13].

WNT-activated medulloblastomas constitute the “medulloblastoma, WNT” methylation class. These tumors typically exhibit loss of Chr.6 with, otherwise, only few copy-number alterations (Fig. 13a). Due to differences in the DNA methylation profile, SHH-activated medulloblastoma have been divided into the two subclasses “SHH A (children and adult)” and “SHH B (infant)” together forming a methylation class family. The two methylation subclasses are likely related to the age dependent spectrum of genetic alterations found in SHH medulloblastoma [22]. Both SHH medulloblastoma subclasses do not have highly characteristic CNV (Fig. 13b, c), but are both moderately enriched for cases with loss of 9q. Group 3 and group 4 medulloblastomas also form a methylation class family consisting of “medulloblastoma, subclass group 3” and “medulloblastoma, subclass group 4”. Tumors in both methylation classes frequently show an isochromosome 17q, approximately 60% in group 3 and 80% in group 4. Group 3 medulloblastoma tend to have a higher number of alterations, most frequently gain of 1q and, 7 and loss of 10q (Fig. 13d). Besides isochromosome 17q, most abundant in group 4 medulloblastomas is gain of 7p and loss of 8 (Fig. 13e). The frequency of amplifications of MYCN on Chr.2p or MYC on Chr.8q is not readily extractable from the summary CNV.

Embryonal tumor with multilayered rosettes, C19MC-altered is another genetically defined WHO entity. It completely matches the methylation class “embryonal tumor with multilayered rosettes”. By definition, the CNV plot shows a focal amplification of 19q13. Furthermore, Chr.2 gain is seen in approximately 60% of these tumors (Fig. 13f). The very few cases that we have so far observed of the exceedingly rare non-19q13 amplified embryonal tumors with multilayered rosettes also fall into this group. The distinct methylation class of intraocular medulloepithelioma has not yet been included in the Classifier [17, 25, 40].

CNS neuroblastoma is represented by the methylation class “CNS neuroblastoma with FOXR2 activation”. These tumors show a characteristic gain of Chr.1p and frequently an additional loss of Chr.16q (Fig. 14a). These tumors have recently been defined in more detail [45].

The exceedingly rare class of CNS ganglioneuroblastomas is currently not represented as a distinct class in the Classifier.

CNS embryonal tumor, NOS, has been extensively studied and some of the cases could be allotted to newly defined molecular classes [45]. Besides “CNS neuroblastoma with FOXR2 activation” the three classes of “CNS high-grade neuroepithelial tumor with BCOR alteration”, “CNS Ewing sarcoma family tumor with CIC alteration”, and the previously mentioned “CNS high-grade neuroepithelial tumor with MN1 alteration” may frequently present with primitive embryonal histology. The classes show different CNV profiles (Figs. 8c, 14b, c) and additional defining molecular changes have been described [45]. CNV profiles of “CNS Ewing sarcoma family tumor with CIC alteration” and “CNS high-grade neuroepithelial tumor with BCOR alteration” exhibit no characteristic features, whereas “CNS high-grade neuroepithelial tumor with MN1 alteration” very frequently shows extensive alterations on Chr.X reminiscent of chromothripsis (Chr.X not shown in the summary CNV plots).

Atypical teratoid/rhabdoid tumor (ATRT) is represented by three related methylation classes: “ATRT, MYC”, “ATRT, SHH”, and “ATRT, TYR” forming the methylation class family ATRT [18]. The tumors in all three methylation classes are characterized by loss of INI1 protein (SMARCB1). However, the gross alteration patterns of Chr.22 vary, with whole chromosome loss in almost all “ATRT, TYR”, either whole chromosome or focal loss in almost all “ATRT, MYC” and only 50% of cases with whole chromosome loss (and rather more inactivating point mutations) in “ATRT, SHH” (Fig. 14d for ATRT MYC, for SHH and TYR refer to Supplementary file 1). So far, we have not had the chance to investigate tumors with the rare alternative mutation in SMARCA4.

Tumors listed in this WHO chapter are not thoroughly covered by the CNS tumor Classifier. Schwannoma and cellular schwannoma are combined in the methylation class “schwannoma”. This group exhibits localization-dependent sub-groups that are not included in the Classifier [39]. The most frequent CNV in “schwannoma” is loss of 22q seen in approximately 60% of tumors (Fig. 15a). “Melanotic schwannoma” represents a separate methylation class and may be particularly helpful in the differential diagnosis of other melanocytic lesions. Members of the methylation class “melanotic schwannoma” exhibit loss of Chr.22 in approximately 50% and equally frequently loss of Chr.21 in addition, with a wide spectrum of other CNV observed at lower frequency (Fig. 15b).

Malignant peripheral nerve sheath tumors have been shown to form distinct methylation classes among peripheral nerve sheath tumors [39]. However, these classes have not been included in the CNS tumor classifier, because a sarcoma classifier containing these tumors is currently under development.

Meningiomas, independent of grade and subtype are represented by the broad methylation class “meningioma” in the Classifier. However, recent work shows that meningioma can be sub-grouped into three benign (MC ben-1, 2, 3): two intermediate (MC int-A, B) and one malignant (MC mal) methylation classes. The progression-free survival of benign, intermediate, and malignant MCs is highly distinct. In fact, the power for predicting recurrence of these six methylation sub-groups is higher than that of WHO grading [42]. These epigenetic sub-groups, their mutational characteristics, CNV, and the association with histology and outcome have been previously reported [42]: Cases of MC ben-1 have typically no aberrations besides 22q deletion and NF2 mutation. They mostly display fibroblastic or transitional histology. MC ben-2 is characterized by flat CNPs and an enrichment for KLF4/TRAF7, AKT1/TRAF7, and SMO mutations. Histology is often of meningothelial (AKT1, SMO) or secretory (KLF4/TRAF7) subtype. If needed, detection of the exact AKT1 or SMO mutation holds potential for targeted therapy in case of a biologically low grade but not resectable meningioma. MC ben-3 is the only sub-group harboring many gains of chromosomes, with or without 22q loss and NF2 mutation. Most angiomatous, metaplastic, and microcystic cases fall into this sub-group. The intermediate and malignant MCs mostly have 22q deletion and NF2 mutations, with the number of losses of whole chromosomes (10, 14) or chromosomal arms (1p) increasing with malignancy. These groups, particularly MC mal, can also carry TERT promoter mutations and show CDKN2A deletion. However, a flat copy-number profile does not necessarily indicate a low-grade biology. For example, cases with BAP1 mutations can have no or only focal chromosomal alterations but still be identified as MC mal, in line with their aggressive behavior.

After confirmation of the diagnosis “meningioma” by the CNS tumor classifier, the data can subsequently be analyzed by the meningioma classifier (also available at https://​www.​molecularneuropa​thology.​org/​mnp) for the identification of the respective sub-group.

While evidently malignant meningioma can be readily identified by histology, the power of the epigenetic meningioma sub-groups is the detection of aggressive cases despite inconspicuous histology and a finer separation between benign cases and cases at elevated risk of recurrence than the WHO classification provides. Typical settings in which the epigenetic meningioma profiling is applied, therefore, include cases in which a low-grade histology does not match the clinical impression or if histology is ambiguous between WHO grade I and II.

This tumor group contains tumors which will be incorporated in a separate sarcoma classifier. The CNS tumor classifier contains the methylation classes “hemangioblastoma”, “hemangiopericytoma” “Ewing sarcoma”, and “Ewing sarcoma family tumors with CIC alterations” (the latter already briefly discussed in “embryonal tumors”). The summary CNV plots for these classes can be found in the Supplementary file 1. A more extensive methylation-based classifier for soft tissue and bone tumors is currently under development (unpublished data).

The methylation class “melanoma” consists of melanoma brain metastases of dermal tumors. Primary malignant melanomas of the CNS were not available in sufficient numbers to generate a methylation class. The methylation class “melanocytoma” is clearly separate from both dermal melanoma metastases and melanotic schwannoma. Tumors of the methylation class “melanoma” exhibit many CNV with 1q gain, 6p gain, 6q loss, 7 gain, 9p loss, and 11q loss being most frequent (Fig. 15c). CNV in melanocytoma are less frequent, with a gain of 6p being the most common event (Fig. 15d). In our experience, methylation-based analysis is helpful in sharpening the distinction between dermal melanoma metastasis and melanocytoma, and between melanocytoma and melanotic schwannoma [19]. It should be kept in mind that uveal melanoma and melanocytoma share both methylation signature and specific mutations, most frequently GNAQ and GNA11 [12].

Hematopoietic tumors are represented by the methylation classes “lymphoma” and “plasmacytoma”. The “lymphoma” methylation class was formed exclusively by “diffuse large B-cell lymphoma”, but it was not tested if other lymphomas are also categorized into this class (hence, the broad name of “lymphoma” was preferred for this class). These tumors show multiple CNV, with loss of 6q and gain of 12p seen in more than half of the tumors (Fig. 15e). The methylation class “plasmacytoma” consists of skull and spinal lesions of plasmacytoma. They are characterized by multiple CNV changes (Fig. 15f).

Both of these tumor classes have not yet been included in the Classifier, but will be added in a forthcoming update.

Adamantinomatous craniopharyngiomas are represented in methylation class “craniopharyngioma, adamantinomatous”, while papillary craniopharyngiomas are represented by “craniopharyngioma, papillary”. The corresponding CNV profiles are inconspicuous (Supplementary file 1), but these tumors are characterized by either CTNNB1 or BRAF V600E mutations, respectively [3, 43]. Granular cell tumor, pituicytoma, and spindle cell oncocytoma all exhibit very similar methylation patterns and group together forming the “pituicytoma/granular cell tumor/spindle cell oncocytoma” methylation class. No frequently recurring chromosomal alterations are observed (Supplementary file 1).

Several intracranial primary tumors with relevance to the neuropathologist are included in other books of the WHO series. Of these not included among the CNS tumors, olfactory neuroblastoma/esthesioneuroblastoma, retinoblastoma, chordoma, and several subtypes of pituitary adenoma are included in the current version of the classifier.

Olfactory neuroblastoma/esthesioneuroblastoma forms two closely related methylation classes: “esthesioneuroblastoma, subclass A” and “esthesioneuroblastoma, subclass B”. The corresponding CNV profiles are also very similar with both showing a very high incidence of losses of whole Chr.1–4, 8–10, and 12 (shown for subclass B in Fig. 15g, subclass A in Supplementary file 1). Interestingly, in a recent series of 66 institutionally diagnosed olfactory neuroblastomas/esthesioneuroblastomas, only 42 (64%) of cases were classified as DNA methylation esthesioneuroblastoma, subclass A or B [5]. Analysis of the remaining cases demonstrated a mix of cases either with close relation to IDH2-mutant sinonasal carcinomas (11%), other sinonasal carcinomas (20%), or to a possible new class of sinonasal tumors with high global methylation (6%). Thus, caution may be required with the interpretation of the diagnosis of an institutionally diagnosed olfactory neuroblastoma/esthesioneuroblastoma. This study further indicated that the distinction of the two separate subclasses esthesioneuroblastoma A and B may be rather unstable, and thus, it is possible that the two subclasses will be reunited in future versions of the Classifier.

Retinoblastoma is also included as a small group in the CNS tumor Classifier. CNV shows several complex copy-number changes that have not yet been analyzed in detail and that may require further validation (Supplementary file 1).

Chordomas all cluster together forming the “chordoma” methylation class. These tumors frequently exhibit chromosomal gains of 1q, 7 and losses of 1p, 3, 9, 10, 13q, and 14q (Fig. 15h).

The pituitary adenomas form separate methylation classes reflecting the expression of anterior lobe hormones. The CNV profiles of the various subtypes are given in Supplementary file 1. Interestingly, adenomas with growth hormone (somatotropin) expression separate into three different methylation classes: two of which relate to the densely and one to sparsely granulated adenoma. More data on the relevance of these sub-groups are required.

With the exception of melanoma metastasis (see above), other types of solid metastases are currently not included in the Classifier. When run through the Classifier, we have observed that various classes may be scored with low scores, most frequently “Plexus tumor, subclass pediatric B”. Typically, these scores are below 0.5, but further testing is required with a broader spectrum of metastases to exclude that some tumors may also obtain higher scores. A low score for Plexus tumor, subclass pediatric B in an unexpected setting should, therefore, be treated with caution.

Machine learning-based classification is a powerful tool for pathologists to assist in diagnosing CNS tumors. Increasing diagnostic precision is accompanied by more reliable prognostic evaluation which may be appreciated by patients and clinicians. Precise diagnosis is mandatory if subsequent treatment is affected. Some diagnoses per se should immediately be questioned, such as the diagnosis of an IDH wt astrocytoma. We find this a frequent diagnostic problem in stereotactic biopsies, where likely the limited size of the specimen precludes the detection of pathognomonic morphological hallmarks. However, this diagnosis should not be accepted by anyone involved in postoperative therapy of glioma patients, because it comprises glioblastomas and occasionally lower grade astrocytomas in need of very different treatments.

Physicians treating pediatric patients have a high demand for diagnostic refinement. Subtyping of medulloblastoma into four groups with different prognosis has already prompted adaptation of treatment regimens (e.g., in the PNET5 study; NCT02066220). This extends to a demand for sequence data on specific genes for targeted therapy. The TP53 status is of relevance, because an underestimated proportion of medulloblastoma patients carry TP53 germline mutations, putting these patients at very high risk of developing secondary neoplasia upon irradiation or DNA-damaging chemotherapy. Ependymoma diagnosed according to WHO poses an inherent problem as to the relevance of grading. The division of posterior fossa ependymoma in group A and B as well as supratentorial ependymoma in RELA and YAP likely provides a more reproducible categorization than the current WHO grading. While ependymoma, PF A is associated with worse prognosis than ependymoma, PF B, the prognostic power of methylation-based classification in supratentorial ependymomas is not as clear. It has been noted that the outcome of ependymoma, RELA is less favorable than that of ependymoma, YAP [32]; however, this may be in need of further prospective study. In our experience, the classification and grading of glial tumors of children is likely more error prone than that of adult patients, further advocating additional testing for the entirety of this population. Very recent reports have provided convincing evidence for novel tumor entities not represented in the WHO classification. While the current experience with treatment is limited due to low numbers and short observation time, it is reasonable to expect different treatment regimens for these tumors in the future.

Studies designed for testing new drugs or treatment modalities highly depend on biological homogeneity of the tumors of patients included. In many instances, the natural course of tumors exhibiting morphological similarity, but genetically representing different entities can have more of an impact on progression-free or overall survival than the therapeutic effect of the tested treatment regimen. This attaches importance to maximal precision of diagnosis, at the latest upon entry into a clinical trial.

As therapy develops and becomes more expensive, an increase of diagnostic precision will result in a better return on this investment at relatively low costs for testing. Many therapies carry inherent danger of secondary damage. Precise diagnostics will also prevent overtreatment of patients whose tumors may naturally display a more favorable prognosis. Novel findings in meningioma, for example, may point this way. Mainly depending on mitotic count, meningiomas are currently divided into WHO grade I, atypical grade II, and anaplastic WHO grade III. Approximately 8% of meningiomas are considered atypical and many of these patients receive postoperative treatment. Classification into distinct methylation classes appears to better prognosticate recurrence-free survival than WHO grading can accomplish [42]. This allows identification of patients receiving the diagnosis of meningioma WHO grade I with increased risk for recurrence, as well as patients receiving the diagnosis of atypical meningioma WHO grade II who, in fact, have no increased risk for recurrence. This equally holds true for low-grade gliomas in children misdiagnosed as high-grade gliomas. In this setting, an increase in diagnostic resources could be highly economic and beneficial to such patient groups.

In our multidisciplinary setting for diagnosing and treating CNS tumor patients, we have developed an approach for employing methylation-based classification in typical diagnostic and clinical settings.