DNA methylation profiling improves routine diagnosis of paediatric central nervous system tumours: A prospective population‐based study

Abstract Aims Paediatric brain tumours are rare, and establishing a precise diagnosis can be challenging. Analysis of DNA methylation profiles has been shown to be a reliable method to classify central nervous system (CNS) tumours with high accuracy. We aimed to prospectively analyse CNS tumours diagnosed in Sweden, to assess the clinical impact of adding DNA methylation‐based classification to standard paediatric brain tumour diagnostics in an unselected cohort. Methods All CNS tumours diagnosed in children (0–18 years) during 2017–2020 were eligible for inclusion provided sufficient tumour material was available. Tumours were analysed using genome‐wide DNA methylation profiling and classified by the MNP brain tumour classifier. The initial histopathological diagnosis was compared with the DNA methylation‐based classification. For incongruent results, a blinded re‐evaluation was performed by an experienced neuropathologist. Results Two hundred forty tumours with a histopathology‐based diagnosis were profiled. A high‐confidence methylation score of 0.84 or more was reached in 78% of the cases. In 69%, the histopathological diagnosis was confirmed, and for some of these also refined, 6% were incongruent, and the re‐evaluation favoured the methylation‐based classification. In the remaining 3% of cases, the methylation class was non‐contributory. The change in diagnosis would have had a direct impact on the clinical management in 5% of all patients. Conclusions Integrating DNA methylation‐based tumour classification into routine clinical analysis improves diagnostics and provides molecular information that is important for treatment decisions. The results from methylation profiling should be interpreted in the context of clinical and histopathological information.

INTRODUCTION Paediatric central nervous system (CNS) tumours are rare and show great heterogeneity, which makes the diagnosis challenging. The classification is based on histopathologic and molecular criteria as outlined by the World Health Organisation (WHO) as well as the location of the tumour [1]. Today, more than 100 different CNS tumour entities with varying grade of malignancy are distinguished. Determining the neuropathological diagnosis can be difficult, and previous studies have reported high inter-observer variability in the diagnostics for some tumour entities [2,3].
CNS tumours are the second most common group of tumours in children, after leukaemia/lymphoma, but account for the majority of cancer-related deaths [4]. These tumours are currently treated with surgery, often followed by chemo-and/or radiotherapy and in selected cases with targeted therapy. In addition to the relatively high overall mortality [5], childhood brain tumour survivors often suffer from serious side effects, both short-term and long-term, with considerable risks of neurological, endocrine and cognitive health problems [6][7][8]. An accurate diagnosis of paediatric tumours is crucial for the choice of treatment and to achieve the optimal balance between likelihood of long-term cure and avoidance of excessive treatment.
During the last decade, DNA methylation profiling has been shown to be a reliable and robust method to classify paediatric CNS tumours [9,10]. The technique is reproducible for analysing freshfrozen tumour samples as well as formalin-fixed paraffin-embedded (FFPE) tumour samples [11]. DNA methylation profiling is recognised as an important tool to stratify paediatric brain tumour patients into clinically relevant subgroups [12][13][14][15] and to better predict prognosis and response to treatment [16,17]. Recent studies showed that using conventional histopathological and molecular diagnostics in combination with DNA methylation profiling can refine tumour diagnoses and sometimes lead to a change of the final diagnosis [18][19][20].
The impact of using DNA methylation analysis in the diagnostics of paediatric brain tumours has been demonstrated in several studies.
However, some series were enriched for diagnostically challenging cases and high grade tumours [18,21] and several retrospective analyses lacked a diagnostic re-analysis by state of the art neuropathologic evaluation. In this analysis, we fully investigated an unbiased series of unselected cases. Therefore, in this study, we aimed to investigate the impact of performing DNA methylation profiling in routine diagnostics, for all children diagnosed with a CNS tumour in Sweden during a four-year period.

Key points
• DNA methylation has a key role (up-front) in a population-based diagnostic setting for paediatric CNS tumours.
• Methylation-based tumour classification enhances the diagnostic information, helps identify rare entities and allows for a change in the management of patients.
• The data demonstrate that also tumours with low tumour cell content can be well classified.

Patients and samples
All paediatric patients (<18 years old) diagnosed with a CNS tumour at one of the six paediatric neurosurgery-oncology centres in Sweden between 1 January 2017 and 31December 2020 were eligible for the study, provided sufficient FFPE tumour material was available for DNA methylation analysis. Histopathological reports including immunohistochemistry and molecular analyses were performed at the six pathology departments involved in diagnosing paediatric CNS tumours in Sweden. DNA extraction and DNA methylation array analysis were centralised to Carén lab at Sahlgrenska Centre for Cancer Research in Gothenburg.
Clinical data were obtained from the Swedish Childhood Cancer Registry. For all patients included in the study, complete histopathological and molecular standard diagnostics was performed prior to the methylation analysis.
Additional tumour samples from 32 patients with tumour relapse, diagnosed during the study period, were analysed with methylation array. The primary operation in these cases had been performed years earlier. Tumour samples from the primary operations were also collected and analysed in the same order as above. In total, 66 samples (initial diagnosis and relapse and for two cases also a second relapse) were collected.

Tumour cell content
The proportion of tumour cells in each sample was estimated by two neuropathologists (TOB, SD). The assessment was based on haematoxylin and eosin-stained slides, taken before and after slicing the FFPE block used for the array analysis. We defined a high tumour cell content as ≥70% tumour cells. All tumour samples were analysed with the methylation array regardless of the tumour cell content.

DNA extraction and quantification
Tumour DNA was extracted from sections of FFPE tumour blocks and extracted with QIAamp ® DNA FFPE kit (Qiagen, Hilden, Germany) following the protocol provided by the manufacturer, with an extra proteinase K digestion step as previously described [22]

Methylation-based classification
For methylation-based tumour classification, raw data (idat files) were uploaded to the openly available DNA methylation-based classifier, MNP version 11b4 (www.molecularneuropathology.org). This resulted in a report, indicating the best predicted match of a methylationbased tumour diagnosis and a corresponding calibrated score (CS), calculated using an algorithm from the German Cancer Research Centre (DKFZ) [12] ranging from 0-1, and a chromosomal copy number variation (CNV) plot. The brain tumour classifier v11b4 comprises 82 CNS tumour methylation classes and nine control tissue methylation classes [12]. In line with previous publications [19,21] and as suggested Therefore, all 240 tumour samples were re-analysed using this new version, which includes 184 molecular tumour classes, subclasses and control tissue classes. We applied the same cut-off for successful classification (CS ≥ 0.84), although the optimal cut-off for this version has not been fully investigated.

Diagnostic impact
In order to establish what potential impact, the methylation-based classification would have had on the final diagnosis if used up-front, the original histopathology reports were reviewed. This assessment was done by two authors (TOB, ES) independently of the reporting neuropathologists. When the CS was ≥0.84, the impact was categorised as one of the following: (I) confirmed the diagnosis, that is, the methylation-based classification and the histopathological diagnosis were in agreement; (II) confirmed and refined the diagnosis, that is, providing additional molecular subtyping information to standard diagnostics; (III) altered the initial diagnosis and would have changed the final diagnosis if the methylation-based classification had been included in real time diagnostics; or (IV) considered non-contributing or misleading.
When the CS was <0.84, it was considered as (V) a lower confidence score when the CS was 0.3-0.83 or (VI) unclassified when the CS was <0.3.

Neuropathological re-evaluation
All cases with a high-confidence CS ≥0.84 that differed from the diagnosis in the original neuropathological report, and cases with a lower confidence score (<0.84) but with a high tumour cell content, were re-evaluated by an experienced neuropathologist (TP). These samples were re-evaluated and classified according to WHO 2016 using immunohistochemical and molecular analyses [1]. The reviewing neuropathologist was blinded for the original histopathological reports as well as for the results from the methylation profiling.

Statistics
For statistical analyses, the statistical software R was used [23].
Comparisons of CSs and tumour cell content between groups were performed using Welch two sample t test. The significance level was set to p = 0.05.

Ethics
The study was approved by the regional ethics committee in Gothenburg, Sweden (Dnr 604-12, T1162-16). Informed consent was obtained from the guardians.

Patients' characteristics
In all, 372 paediatric patients (0-18 years old) were diagnosed with a CNS tumour in Sweden during 2017-2020. In 313 cases, a tissue-based diagnosis was obtained. Patients with insufficient FFPE tumour tissue for methylation analysis or when informed consent was not obtained were not included, Figure 1. Two patients were excluded from the analysis of the diagnostic effect since a DNA methylation array had already been performed as part of the clinical diagnostic work-up, influencing the original diagnosis. Ten cases with germ cell tumours (GCTs) were excluded as this tumour class is not included in the classifier version 11b4.  Table 1.

DNA methylation-based classification
From the 240 primary tumours, the classifier tool (MNP version11b.4) assigned 187 cases (78%) to a specific DNA methylation class with a F I G U R E 1 Cohort description. FFPE = formalin-fixed paraffin-embedded, GCT = germ cell tumour.
CS of 0.84 or higher, Figure 2. In 40 cases (17%), methylation classification produced lower confidence CSs, between 0.3 and <0.84. For cases with a class-prediction CS < 0.30, a methylation class could not be predicted. This result, here referred to as 'unclassified' cases, was observed in 5% (13/240) of cases and will be discussed below. All samples had a probe failure rate less than 1.5%, with 99% of the samples less than 1%.
Diagnostic impact of methylation profiling I Confirmation of diagnosis and II confirmation and refinement of diagnosis (CS ≥ 0.84) In 165/240 cases (69%), the predicted methylation class confirmed the initial neuropathological diagnosis. In 59 of these cases (25%, 59/240), the initial neuropathological diagnosis was not only confirmed but also refined by the added information gained by DNA methylation, for example, providing molecular subgrouping data not available with standard diagnostics. This refinement of diagnosis  Table 2 and described in detail below for selected cases.
Cases 1 and 2 were initially diagnosed as pilocytic astrocytomas (PA) but classified as diffuse leptomeningeal glioneuronal tumours (DLGNTs) by the classifier-tool. These tumours had clinicopathological features of DLGNT [24]; that is, leptomeningeal enhancement of the spinal cord or loss of chromosome arm 1p visualised on the CNV plot.    [27]. When reviewed, the other five unclassified tumours were considered to be unusual and undifferentiated tumours, most likely representing rare novel entities.

Tumour types and CS
We also investigated how the class prediction scores varied in relation to the different tumour types, based on the initial neuropathological diagnoses, Figure 6. In 18/240 cases (8%), the predicted diagnoses using version 12.5 were incongruent with the original histopathological diagnoses including the cases with altered diagnosis with v11b4 (discussed previously). For 23/240 samples (10%), the class-prediction score was <0.84 but above 0.3 (median 0.63). Still, in almost half of these cases, the suggested methylation class was in concordance with the initial histopathological diagnoses despite the lower confidence score. The other cases were non-concordant or indicated brain control tissue (9/23).
The majority of these samples had a tumour cell content <40%. In 9/240 cases, methylation profiles could not be classified by the MNP version 12.5 with a score of >0.3 (unclassified cases).
No further analyses were performed in these cases as this was beyond the scope of this study.
Methylation-based analysis using the specific medulloblastoma classifier group 3/4 1.0 The 26 non-WNT/non-SHH medulloblastomas were re-analysed with the specific medulloblastoma classifier for group 3 or 4 medulloblastomas. All tumours retained their predicted genetic subgroup between classifier versions 11b4 and 12.5 as well as the medulloblastomas classifier. Four cases classified as subgroup 3; the remaining 22 were classified as subgroup 4. In two samples, the subtypes changed from subtype VIII in the classifier v12.5 to subtype VI in the medulloblastoma classifier (sample-id 8 and 10) (Table S1).

Relapses
We Successful (diagnostic) molecular classification by DNA methylation analysis was achieved in 78% of the cohort using the MNP classifier v11b4 [12]. Previous studies have shown classification rates of 49-72% [18][19][20][21]. The higher classification rate in our cohort could be explained by the population-based setup of our study, rather than the investigation of a mixed cohort of paediatric patients composed of diagnostically challenging cases or cases referred for second opinion as in previous studies.
In the majority of cases (69%), the histopathological diagnoses and the methylation-based diagnoses were in agreement. In 25% of the whole cohort, the methylation profiling did not only confirm but also refined the initial diagnosis, for example, by giving a more precise subgroup. This diagnostic refinement is important from a prognostic point of view as different molecularly defined tumour types have a distinct clinical behaviour [16,29] and increasingly also from a therapeutic viewpoint.
In this study, 6% of the initial diagnoses were changed when reevaluated by an experienced reference neuropathologist, who was unaware of the methylation profiling results. All diagnoses were changed in favour of the predicted methylation class showing that the integration of methylation analysis in routine clinical diagnostics improves the diagnostic accuracy. It also provides guidance for additional diagnostic testing. However, methylation-based classification could not help in differentiate in tumour grade; for example, PXA cannot be differentiated from anaplastic PXA or epithelioid GBM by methylation analysis.
Obviously, diagnostic classification is important for correct assignment to a specific treatment, but the change in diagnosis does not always lead to a change in treatment. In our study, we estimated that the change in the diagnosis would have altered the management for 11 (5%) of the patients, that is, a change of treatment or a different follow up. This is of utmost importance in paediatric neurooncology and consistent with previous reports [18]. In our judgement, no methylation classification was misleading when seen in relation to clinical and radiological data.
When the classifier cannot find a class prediction with a high confidence score, the interpretation may be problematic. The lower the score, the higher the rate of misleading diagnosis [18]. In this study, 53/240 cases could not be assigned to a DNA methylation class with a CS of ≥0.84 using MNP v11b4. The reasons for scores below the cut-off cannot always be determined with certainty, but several explanations are possible; for example, the amount of DNA was too low or of poor quality, the tumour cell content was too low or the actual tumour entity was not included in the classifier-algorithm. A recent study [21] found that the tumour cell content was the factor most significantly associated with the classifier score when comparing the amount of DNA in the sample, the tumour cell content and the tumour type. However, a suggested methylation-based classification with a CS <0.84 may give important molecular information, especially when the score is >0.5 [15].
Most previous studies have only included samples with a high tumour cell content, ≥70%, in order to increase the possibility of a match to a predicted class. As our study was population-based, all samples were subjected to methylation profiling independently of the tumour cell content in the samples. Tumour cell content was determined by histology, as this is the standard procedure in clinical work, even though it is known to be inaccurate due to interobserver variability [30].
As in previous studies [15,19], it is clear that the tumour cell content is important in order to achieve a confident class score. Previous studies have shown the benefit of performing methylation profiling mainly in high-grade tumours or challenging cases [18,19,31]. However, this study shows that the possibility of a correct classification also depends on the tumour type itself. Thus, although low-grade gliomas and glioneuronal tumours often showed a low tumour cell content, they were predicted with high confidence scores in 71% of cases compared with our total cohort's average of 78%. This was a higher percentage than expected in comparison with other reports [10][11][12][13][14][15][16][17][18].
Consequently, in our opinion, one should not refrain from performing the analysis based only on a low tumour cell count.
When examining the relapses with paired samples, more than 70% had a defined methylation class, which was unchanged between primary diagnosis and relapse, showing that the methylation profiling is robust despite treatment, probably because it reflects the cell of origin of the tumour [32]. The remaining samples from the relapses mainly consisted of cases that were difficult to classify and unusual cases.
When re-analysing all tumour samples with the newer unpublished MNP brain tumour methylation classifier version 12.5, several more samples reached a high confidence prediction score independently of tumour cell content. Not only were more samples confidently assigned to a specific tumour class, confirming or even refining the diagnosis, but also were new entities identified and difficult cases solved. This shows that the DNA methylation classification is gradually evolving, and even more tumour types with new molecular alterations will be included in future versions. In the recent 2021 WHO classification of CNS tumours, several new tumour types and subtypes were established, and many demand an advanced level of molecular diagnostics [33]. DNA methylation provides an additional molecular layer in this respect.
To our knowledge, no population-based study has been performed that has included all CNS tumours from an entire nation during a defined time-period. Our study shows that DNA methylation analysis has an added value in the diagnostics of paediatric CNS tumours and in treatment decisions, if used as a complement to standard neuropathology. Using a newer version of the classifier, with more diagnoses included, several diagnoses changed, and new entities were identified. One important aspect when using DNA methylation in real time diagnostics is to keep the turnaround time as short as possible in order to integrate the classification result in the pathological diagnosis. A result from a methylation analysis could potentially be obtained within 10 days from operation [12]. In many countries, this requires a centralised analysis also for keeping the costs as low as possible. In one study [19], the use of methylation arrays was considered both cost-effective and tissue-saving for diagnostically difficult cases.
A limitation of the study is that microdissection was not used. In fact, histopathological evaluation prior to dissection for molecular diagnostics (including methylation-based classification) is considered current standard of care for neuropathological diagnostic procedures.
With dissection, the number of cases with a high content of neoplastic cells would probably have increased and hence the probability of high score predications. In spite of this, we got a relatively high proportion of tumours with high-confident scores. Despite a study period of 4 years, there were, for some tumour types, few samples in each subgroup, which may affect the interpretation of the results.
Overall, it is crucial to interpret the results from the methylation profiling in the context of clinical, radiological and histopathological information.
In conclusion, our national population-based study shows that DNA methylation classification is of value for all types of CNS tumours and has an important role also in tumours with a low tumour cell content. DNA methylation can enhance the diagnostic information and potentially help identify rare entities. We find DNA methylation-based classification to be an invaluable tool in the diagnostics of paediatric CNS tumours and advocate its integration in real time standard diagnostics.