Molecular characterisation defines clinically-actionable heterogeneity within Group 4 medulloblastoma and improves disease risk-stratification

Group 4 tumours (MBGrp4) represent the majority of non-WNT/non-SHH medulloblastomas. Their clinical course is poorly predicted by current risk-factors. MBGrp4 molecular substructures have been identified (e.g. subgroups/cytogenetics/mutations), however their inter-relationships and potential to improve clinical sub-classification and risk-stratification remain undefined. We comprehensively characterised the paediatric MBGrp4 molecular landscape and determined its utility to improve clinical management. A clinically-annotated discovery cohort (n = 362 MBGrp4) was assembled from UK-CCLG institutions and SIOP-UKCCSG-PNET3, HIT-SIOP-PNET4 and PNET HR + 5 clinical trials. Molecular profiling was undertaken, integrating driver mutations, second-generation non-WNT/non-SHH subgroups (1–8) and whole-chromosome aberrations (WCAs). Survival models were derived for patients ≥ 3 years of age who received contemporary multi-modal therapies (n = 323). We first independently derived and validated a favourable-risk WCA group (WCA-FR) characterised by ≥ 2 features from chromosome 7 gain, 8 loss, and 11 loss. Remaining patients were high-risk (WCA-HR). Subgroups 6 and 7 were enriched for WCA-FR (p < 0·0001) and aneuploidy. Subgroup 8 was defined by predominantly balanced genomes with isolated isochromosome 17q (p < 0·0001). While no mutations were associated with outcome and overall mutational burden was low, WCA-HR harboured recurrent chromatin remodelling mutations (p = 0·007). Integration of methylation and WCA groups improved risk-stratification models and outperformed established prognostication schemes. Our MBGrp4 risk-stratification scheme defines: favourable-risk (non-metastatic disease and (i) subgroup 7 or (ii) WCA-FR (21% of patients, 5-year PFS 97%)), very-high-risk (metastatic disease with WCA-HR (36%, 5-year PFS 49%)) and high-risk (remaining patients; 43%, 5-year PFS 67%). These findings validated in an independent MBGrp4 cohort (n = 668). Importantly, our findings demonstrate that previously established disease-wide risk-features (i.e. LCA histology and MYC(N) amplification) have little prognostic relevance in MBGrp4 disease. Novel validated survival models, integrating clinical features, methylation and WCA groups, improve outcome prediction and re-define risk-status for ~ 80% of MBGrp4. Our MBGrp4 favourable-risk group has MBWNT-like excellent outcomes, thereby doubling the proportion of medulloblastoma patients who could benefit from therapy de-escalation approaches, aimed at reducing treatment induced late-effects while sustaining survival outcomes. Novel approaches are urgently required for the very-high-risk patients. Supplementary Information The online version contains supplementary material available at 10.1007/s00401-023-02566-0.


Introduction
Medulloblastoma (MB) is the most common malignant embryonal tumour of the central nervous system (CNS, WHO grade 4) in children accounting for approximately 10% of all paediatric cancer deaths. Current multi-modal 1 3 treatments for non-infants comprise surgical resection and cranio-spinal radiation (CSI), followed by adjuvant chemotherapy [26]. These treatments commonly cause long-term neurological, neurocognitive and neuroendocrine deficits as well as increased risk for second malignancies [6].
Current clinical risk-stratification models for MB use established clinical and pathological risk-features derived from studies of disease-wide cohorts. Metastatic disease, sub-total surgical resection, and large cell/anaplastic (LCA) histology have long been associated with poor outcomes in such studies [16,36]. Alongside these, molecular features have profoundly improved our ability to predict risk in MB. MB WNT patients aged 3-16 years old consistently achieve favourable outcomes (> 95% 5-year progression free survival) [7,12,28] and, within MB SHH , TP53 mutations are associated with an extremely poor prognosis [38]. In addition, amplification of either the MYC or MYCN oncogenes have strong associations with inferior survival outcomes and are associated with other high-risk features [31]. Together, these risk-features underpin treatment stratifications in current international biomarker-driven clinical trials, which reduce therapy for favourable MB WNT and use intensified regimens for patients with high-risk features (SIOP-PNET5-MB [21] [NCT02066220], SIOP-HR-MB [2] [EudraCT Number: 2018-004250-17], SJMB012 [NCT01878617]). However, these disease-wide risk-features show molecular group dependency (e.g. MYC is prognostic in MB Grp3 ; MYCN in MB SHH ) and current disease-wide risk-stratification models do not adequately characterise risk specifically within MB Grp4 , which accounts for ~ 40% of MB patients and the majority of non-WNT/non-SHH cases. The identification and validation of prognostic biomarkers, which could direct risk-adapted adjuvant therapies for MB Grp4 patients, thus represents an urgent unmet clinical need.
The underlying biology of MB Grp4 is complex. In contrast to MB WNT and MB SHH , MB Grp4 has a relative paucity of defining genetic drivers; analysis of gene-specific mutations and their clinical relevance has not yet been undertaken in large clinically-annotated MB Grp4 cohorts. Cytogeneticallydefined prognostic features have been described within MB Grp4 , first as specific whole-chromosome aberrations (11 loss or 17 gain) [35] and, more recently, through the recognition of a concerted whole chromosomal aberration (WCA) signature within the clinically defined standard-risk (i.e. non-metastatic, non-LCA and non-MYC amplified) HIT-SIOP-PNET4 trial [NCT01351870] non-WNT/non-SHH MB cohort. This signature (defined by ≥ 2 of wholechromosome 7 gain, chromosome 8 loss, and chromosome 11 loss (WCA-FR)) is associated with increased ploidy, multiple non-random WCAs, and predicted a favourable prognosis in both trial and validation cohorts (5-year PFS, 100%). Remaining tumours (WCA-HR) had much poorer outcomes (68% 5-year PFS) [16]. Mynarek et al. recently incorporated WCA signatures as risk stratification parameters in data from the German Paediatric Brain Tumour (HIT) trials [22]. Importantly, inter-relationships between the different MB Grp4 molecular characteristics (i.e. second-generation subgroups, cytogenetic groups, mutations), their utility and performance as prognostic biomarkers, and potential for integration into risk-stratification schemes, remains to be established. Previous studies of MB Grp4 have been limited by cohort size and/or comprehensive clinical annotation [3,23,24,34,35] and have not permitted the integrated characterisation of MB Grp4 molecular pathology, alongside assessment of its translational potential to improve clinical sub-classification and risk-stratification.
Here, we report a comprehensive characterisation of the molecular pathology of primary MB Grp4 , integrating disease subgroups, mutational and copy-number events, and assess their translational potential in large, clinically-annotated discovery and validation cohorts, comprising > 1000 MB Grp4 tumours. We describe non-random, clinically-actionable biological heterogeneity, which forms the basis of novel biomarker-driven risk-stratification models. These models improve outcome prediction and reassign risk-status for ~ 80% of MB Grp4 patients. These findings refine our understanding of MB Grp4 biology and provide a foundation for personalised therapies, improved therapeutic strategies and future clinical trials.
Ethical approval and consents were given.

Procedures
For the discovery and non-MB Grp4 comparator cohorts, MB diagnosis was confirmed by methylation-based classification [33] and/or central neuropathology review (CPR, 81%) which provided histological sub-classification. In the absence of CPR, institutional annotation was used. Metastatic staging was assigned according to Chang's criteria (M + ; M stages 1-4, M0; local disease only) [4]. Extent of resection was evaluated institutionally and tumours were considered sub-totally resected if residual disease exceeded 1·5cm 2 [1].
Molecular groups were assigned using established methods, and only tumours confidently assigned as MB Grp4 were included in the discovery cohort [32,33], Second-generation methylation subgroups were assigned using the 'MNP Medulloblastoma classifier group 3/4' version 1·0 at www. molec ularn europ athol ogy. org/ mnp. Chromosome armlevel copy number estimates were derived from Illumina HumanMethylation 450K/EPIC array data, using the package 'Conumee' (R/Bioconductor) as previous described [30,33]. Molecular inversion probe (MIP) array was used for the HIT-SIOP-PNET4 cohort to call arm-level copy number as previously described [16]. MYC(N) copy number status was assessed by iFISH, copy number estimates from methylation array and/or MLPA [17]. Established non-WNT/non-SHH focal copy number variants (CNVs, Supplementary Table 2, online resource) were assessed as described [27,30].

Statistical and survival analysis
In accordance with current treatment protocols, survival analysis in the discovery cohort was restricted to patients aged ≥ 3 years who received CSI and where outcome data was available (323/362 [89%]). Univariable and multivariable Cox proportional hazards models were used to investigate predictors of progression free survival (PFS). Multivariable Cox models (n = 213 patients with available subgroup data) were constructed using backwards selection, considering established clinico-molecular and treatment variables (metastatic disease, extent of resection, LCA pathology, sex, MYC/MYCN amplification, i17q, and dose of CSI) in addition to biologically and clinically significant molecular factors (WCA status, subgroup 7, chromosome 13 loss, subgroup 5, chromosome 18 gain). To assess performance, bootstrapped models were generated using 1000 rounds of resampling to assess calibration and discrimination at 5 years from diagnosis and were tested in an independent validation cohort [23]. From the multivariable Cox model, a novel, clinically-deliverable MB Grp4 risk-stratification scheme was generated from combinations of markers by categorising patients using selected variables into risk groups with established disease cut-offs for projected 5-year PFS (favourable-risk, > 90%; standard-risk, > 75-90%; high-risk, 1 3 50-75%; very-high-risk, < 50%) [29]. We finally compared our risk-stratification scheme to those in current clinical practice [2,21] and previously reported molecular stratification schemes [15,35] by assessing discrimination and calibration performance in the discovery and independent validation cohorts, once again using 1000 rounds of resampling and measuring at 5 years from diagnosis. Proportionality of covariates in Cox modelling were tested using scaled Schoenfeld residuals. Kaplan-Meier curves with log-rank tests were constructed to visualise survival associations.
Fisher's exact and Chi squared tests were used to assess associations between categorical variables. Kruskal-Wallis, Mann-Whitney U, ANOVA and t-tests were used to compare continuous variables between groups. Significant associations were defined as having an adjusted p value of < 0·05 using the Benjamini-Hochberg procedure to correct for multiple testing. Statistical and bioinformatics analyses were done using R statistical environment (version 4.0.4).
Full methodological detail can be found within the Supplementary Material (online resource).

Role of the funding source
The funders had no role in study design.
In contrast, multiple recurrent arm-level and WCAs were common ( Supplementary Fig. S1d, online resource). We sought to re-derive, from first principles, prognostic WCA signatures [16] in our risk-independent all-comer MB Grp4 cohort ( Supplementary Fig. S2, online resource). Analysis recapitulated the previous finding that two or more of chromosome 7 gain, chromosome 8 loss and chromosome 11 loss (WCA-FR) represented the optimum combination of WCAs for predicting PFS (Supplementary Fig. S2i, online resource). Unsupervised hierarchical clustering and association analysis further supported two distinct WCA signatures within MB Grp4 ; i17q negatively associated with other WCAs, while chromosome 7 gain, 8 loss and 11 loss positively associated with most other WCAs ( Supplementary Fig. S3a, b, online resource).
We next considered molecular and clinical heterogeneity within MB Grp4 , first focusing on methylation subgroups. While patients < 5 years are uncommon in MB Grp4 overall, subgroup 7 harboured a significant enrichment in these youngest patients (n = 19/72 [26%], p = 0·001; Fig. 1a, Supplementary Fig. S4b, c, online resource). There was no significant association with extent of surgical resection or metastatic disease (Fig. 1a, Supplementary Fig. S4c, online resource).
We next defined the cytogenetic/mutational landscape of the validated MB Grp4 WCA groups originally identified in the standard-risk HIT-SIOP-PNET4 cohort [16], and explored their clinical behaviour in our risk-independent all-comers' cohort. WCA-FR strongly associated with  (Fig. 2a, Supplementary Fig. S6a, online resource). The WCA-HR group was cytogenetically quiet, with i17q in isolation commonly representing the single defining cytogenetic feature (Fig. 2a, Supplementary Fig. S6a, online resource). Although the total mutational burden was equivalently low for both groups (Supplementary Fig. S6b, online resource), the WCA-HR group was significantly enriched for mutations in chromatin remodelling genes (p = 0·007); mutations in KMT2C (p = 0·02) and ZMYM3 (not significant) were exclusive to WCA-HR and mutations in other chromatin remodelling genes (KDM6A (10/11 mutations occurred in the WCA-HR group, 91%) and KMT2D (8/12 mutations, 67%)), were also present at high frequencies (Fig. 2b). The majority of these features validated ( Supplementary Fig.  S6a, b, online resource) in an independent cohort. Established clinico-molecular risk-features (age at diagnosis, metastatic disease, histological variants, extent of surgical resection and amplification of MYCN) were equivalently distributed between both groups (Fig. 2a, Supplementary Fig.  S6a, online resource).
To assess predictors of risk across MB Grp4 , we assessed all eligible [see methods] clinical, pathological and molecular features (i.e. subgroups, focal and arm-level CNVs, mutations), in univariable Cox regression analysis and found multiple significant associations (Fig. 3a) (Fig. 3a). Importantly, histology, MYC(N) amplification and extent of resection showed no prognostic utility and were not selected.
We compared the performance of the Cox model in our cohort to its performance in an external validation cohort. The Cox model had a high, bias-corrected c-index (0·72) and showed good calibration (Fig. 3b) in our discovery cohort, however, the model performed poorly when tested against the external validation cohort, with a bias-corrected c-index of 0·60 and poor calibration (Fig. 3c).
We therefore next developed a stratification scheme by selecting well defined molecular disease features [26] from our Cox model, remaining as faithful as possible to the original model, whilst minimising the number of variables upon which risk could be stratified accurately. Combinations of markers were used to categorise patients into risk groups to develop a novel, clinically-deliverable MB Grp4 risk-stratification scheme, from established disease cut-offs for projected 5-year PFS [29]. WCA status was retained over chromosome 13 loss, and subgroup 5 was not included as it represents a small proportion of MB Grp4 .
Our MB Grp4 risk-stratification scheme thus redefined risk within MB Grp4 disease (Fig. 4b); clinical standard-risk (37% of MB Grp4 ) was effectively eliminated and redistributed to favourable (37% of clinical standard-risk) or highrisk (63% of clinical standard-risk) groups. A significant subset of the clinical high-risk group (63% of MB Grp4 ) redistributed to favourable (9% of clinical high-risk) or very-high-risk (60% of clinical high-risk) groups.
Finally, our MB Grp4 risk-stratification scheme was reproducible in the independent external validation cohort (Fig. 4e). The performance of the MB Grp4 risk-stratification scheme (bias corrected c-index: 0·6) was again better than published and/or clinically utilised stratification models and, unlike the Cox model, calibrated well within the validation cohort (Fig. 4d, f, Supplementary Fig. S8a, b, online resource).

Discussion
Our interrogation of the specific molecular pathology of ~ 1000 MB Grp4 tumours strongly confirms the well-established prognostic importance of metastatic disease within this disease group. We have identified clinically-actionable biological heterogeneity, centred on WCA and methylation subgroups, improving upon initial studies of their molecular and prognostic relevance, which were limited by either cohort size or lack of combined clinico-molecular annotation. The novel integration of these features resolves biologically homogeneous risk groups which allow us to derive a risk-stratification model that reassigns risk-status for 80% of paediatric MB Grp4 , which outperforms previous schemes (clinical [2,21], cytogenetic [35] and MB Grp3/ Grp4 [15]) and validates in an independent cohort. Importantly, our findings reject previously established disease-wide risk-features (LCA histology and MYC(N) amplification), showing they have little prognostic relevance in this disease context.
Our MB Grp4 risk-stratification scheme balances the considerations of predictive accuracy with clinically actionable and practical risk-redistribution; i.e. the establishment of three meaningfully-sized risk groups. MB WNT represents around 10% of all MB patients and, to date, is the only Fig. 1 Characterisation of MB Grp4 second-generation methylation subgroups. Distribution of a established clinicopathological characteristics and significantly enriched cytogenetic aberrations. b Genespecific genetic alterations within MB Gpr4 methylation subgroups. For Fig. 2a and b, significance is shown from Fisher's exact or Kruskal-Wallis tests, *depicts significance recapitulated in validation cohort. Residuals from χ 2 test indicate subgroup-enrichment (strong relationships are indicated by darker shades of grey) alongside scale bar. Number of WCA gains (red), losses (blue), total WCAs and number of genetic aberrations (black) are also shown with increasing colour intensity indicating a higher number of changes. Features with a cohort-wide frequency of ≥ 5% or with a subgroup-specific frequency ≥ 10% were included in the analysis. MYC amplifications are shown despite their low frequency. Full data is shown in Supplementary Fig. S4, online resource. c Kaplan-Meier plot of PFS by MB Grp4 methylation subgroup. d Kaplan-Meier plot for PFS in subgroup 5 for chromosome 13 loss. Univariable Cox proportional hazards models of PFS in MB Grp4 e subgroup 6 and f subgroup 7 for clinical and molecular features ≥ 10%. Kaplan-Meier plots of PFS by g metastatic disease in subgroup 6, h WCA groups in subgroup 6 and i metastatic disease in subgroup 7. At-risk tables are shown in two-year increments with number of patients censored in parentheses with significance shown by p value generated from log-rank test. Abbreviations: M + metastatic disease, M0 non-metastatic disease, STR sub-total resection, CLA classic, DN/MBEN desmoplastic/nodular or medulloblastoma with extensive nodularity, LCA large-cell/anaplastic, WCA-FR/HR whole chromosome aberrations-favourable risk/high risk, CTX chemotherapy, HR hazard ratio, CI confidence interval. # Estimates not possible due to group with no events, p values reported from log-rank test ◂ clinically-actioned favourable-risk group in the disease with CSI dose de-escalation from 24 to 18 Gy [21]. Therefore, our identification of a reproducible favourable-risk group in MB Grp4 with excellent outcomes (20% of MB Grp4 , equating to 8% of all MB with a 97% 5-year PFS) almost doubles the proportion of all MBs suitable for therapy de-escalation approaches.
Our model eliminates clinically-defined standard-risk disease within MB Grp4 , representing a significant advancement for the future clinical management of MB Grp4 . Our model redistributes a significant majority (63%) of these patients to a high-risk disease group, potentially explaining why dose reduction strategies within clinically-defined standard-risk MB Grp4 has shown inferior survival outcomes in previous trials [20]. The concept of treatment intensification for patients, previously defined on clinical grounds as SR, is currently a matter of careful consideration within the field. Some intensification of chemotherapy or radiotherapy might be appropriate, however for the latter it is conceivable that doses lower than 36 Gy may be sufficient to induce a survival benefit. In clinically-defined high-risk MB Grp4 patients from the PNET HR + 5 trial (mostly metastatic), high-dose thiotepa conferred a significant survival advantage [10]. Further work is needed to assess chemotherapy intensification and will be addressed in the SIOP-HR-MB trial [2].
Despite our comprehensive assessment of MB Grp4 molecular pathology, significant survival heterogeneity remains within the high and very-high-risk groups; further biomarkers remain to be identified. Whilst outside the scope of this study, future molecular profiling of the transcriptome and proteome is urgently required to identify novel actionable biological pathways in MB Grp4 . For example, we observed an enrichment of lesions in genes that have coalescing functions as modifiers of H3K4 and H3K27-methylation within a subset of high-risk MB Grp4 . Mutations in these chromatin modelling genes have been shown to induce aberrant expression patterns of their target histone markers and are associated with worse survival outcomes within non-WNT/non-SHH disease [9]. These epigenetic markers have previously been associated with a radiation-resistant phenotype in experimental systems; therapeutic targeting through BET inhibitors in high-risk non-WNT/non-SHH models has been shown to restore radiation sensitivity and may therefore present a potential novel therapeutic intervention [13].
Our analysis was based on retrospective clinical cohorts and heterogeneous treatment protocols. However, the inclusion of patient data from contemporary cohorts (i.e. HIT-SIOP-PNET4, PNET HR + 5) and a discovery cohort of unprecedented size mitigates this limitation. Furthermore, multimodal therapy in non-infants has become standardised in the last three decades (surgical resection, CSI with adjuvant chemotherapy), legitimising the combination of retrospective cohorts for the development of survival models. The clinical implementation of our MB Grp4 risk-stratification scheme is both attractive, feasible and immediately adoptable into clinical studies, given that patients can be mapped using only metastatic status and molecular data routinely collected via DNA methylation array (WCAs/molecular subgroup), which is, increasingly, becoming standard of care in developed countries [8,19]. Profiling using the Illumina 450k/EPIC DNA methylation arrays has been proven to be robust for detecting copy number variation, in particular WCAs [18].
Our findings refine our understanding of the clinical behaviour of MB Grp4 and now require urgent assessment in prospective clinical trials, as a basis for improved Fig. 2 Characterisation of MB Grp4 WCA groups. Distribution of a established clinicopathological characteristics and significant cytogenetic aberrations. b Genetic alterations in chromatin remodelling genes within MB Gpr4 WCA groups. For a and b, significance is shown from Fisher's exact or Mann-Whitney U tests, *depicts significance recapitulated in validation cohort. Residuals from χ 2 test indicate subgroup-enrichment (strong relationships are indicated by darker shades of grey) alongside scale bar. Number of WCA gains (red), losses (blue), and total WCA (black) are also shown, with increasing colour intensity indicating a higher number of changes. Features with a cohort-wide frequency of ≥ 5% or with a subgroup-specific frequency ≥ 10% were included in the analysis. MYC amplifications are shown despite low frequency. All data is shown in Supplementary Fig. S5 [21] [NCT02066220] and SIOP-HR-MB [2] [EudraCT: 2018-004250-17]) and the MB Grp4 risk-stratification scheme. c Kaplan-Meier plot of PFS by MB Grp4 risk-stratification group. At-risk tables are shown in two-year increments with number of patients censored in parentheses and significance shown by p value generated from log-rank test. d Performance of MB Grp4 risk-stratification scheme in comparison to the current clinical-trial risk scheme (SIOP-PNET5-MB [21] and SIOP-HR-MB [2]), a previously reported cytogenetic risk scheme (Shih et al. [35]) and a published combined MB Grp3/4 risk scheme (Gajjar et al. [14]) measured by bias-corrected c-index at 5 years within discovery and validation cohorts. e Kaplan-Meier PFS plot for the MB Grp4 risk-stratification scheme within the external validation cohort (Northcott et al. [23]). f Calibration plot of the MB Grp4 risk-stratification within the validation cohort for survival at 5 years alongside the bias-corrected c-index. Abbreviations: M0 non-metastatic disease, M + metastatic disease, WCA-FR whole chromosome aberrations-favourable risk, WCA-HR whole chromosome aberrations-high risk, PFS progression-free survival diagnostics, personalised therapies and risk-adapted therapeutic strategies.