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 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].

MB comprises molecular disease groups which form the basis of the genetically-defined medulloblastoma classification in the 2021 WHO classifications of CNS tumours, alongside its histologically-defined MB entities. These are: WNT-activated (MBWNT), Sonic Hedgehog-activated (MBSHH) TP53-wildtype, MBSHH TP53-mutant and non-WNT/non-SHH (comprising Group 3 (MBGrp3) and Group 4 (MBGrp4)) [19].

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. MBWNT patients aged 3–16 years old consistently achieve favourable outcomes (> 95% 5-year progression free survival) [7, 12, 28] and, within MBSHH, 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 MBWNT 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 MBGrp3; MYCN in MBSHH) and current disease-wide risk-stratification models do not adequately characterise risk specifically within MBGrp4, 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 MBGrp4 patients, thus represents an urgent unmet clinical need.

Molecular substructure within MBGrp3 and MBGrp4 was recently described in three independent studies, each proffering different solutions [3, 23, 33, 34]. Cavalli et al. identified that MBGrp3 and MBGrp4 each partition into three subgroups (α, β, and γ) with specific transcriptional profiles [3], whilst Northcott et al. identified and mutationally characterised eight subgroups (I-VIII) that are shared by MBGrp3/Grp4 [23]. Schwalbe et al. highlighted the prognostic relevance of MBGrp3 and MBGrp4 molecular substructure by defining clinically relevant high- and low-disease risk subgroups (MBGrp3-LR/HR, MBGrp4-LR/HR) [33]. A subsequent international meta-analysis resolved this disparity and supported a definition comprising eight robust, clinically-relevant, second-generation methylation MBGrp3/Grp4 subgroups (1–8) [34]; these second-generation methylation subgroups have been adopted into the 2021 WHO classification of CNS tumours [19] and their clinical behaviour has been further confirmed in contemporary, clinically controlled, cohorts [14, 22]. Furthermore, transcriptomic profiling of MBGrp3/Grp4 has also shown that they can be represented as a bipolar continuum between archetypical MBGrp3 and MBGrp4, denoted by a MBGrp3/Grp4 continuum score [37]. Second-generation methylation subgroups can be ordered on this expression continuum from prototypic MBGrp4 subgroups (8, 7 and 6), through mixed subgroups (5, 1), to MBGrp3 subgroups (3, 4, 2).

The underlying biology of MBGrp4 is complex. In contrast to MBWNT and MBSHH, MBGrp4 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 MBGrp4 cohorts. Cytogenetically-defined prognostic features have been described within MBGrp4, 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 whole-chromosome 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 MBGrp4 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 MBGrp4 have been limited by cohort size and/or comprehensive clinical annotation [3, 23, 24, 34, 35] and have not permitted the integrated characterisation of MBGrp4 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 MBGrp4, integrating disease subgroups, mutational and copy-number events, and assess their translational potential in large, clinically-annotated discovery and validation cohorts, comprising > 1000 MBGrp4 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 MBGrp4 patients. These findings refine our understanding of MBGrp4 biology and provide a foundation for personalised therapies, improved therapeutic strategies and future clinical trials.

Materials and methods

Study design and participants

A discovery cohort of 362 primary MBGrp4 tumours (Supplementary Fig. S1a, online resource) and a comparator cohort of 489 non-MBGrp4 tumours (Supplementary Table 1, online resource) were assembled from UK-Children’s Cancer and Leukaemia Group (CCLG) institutions, collaborating international centres and the SIOP-UKCCSG-PNET3 [11], HIT-SIOP-PNET4 [16], and PNET HR + 5 [10] clinical trials (Supplementary Fig. S1a, online resource). Furthermore, 668 independent MBGrp4 samples from two published studies [3, 23] were used as validation cohorts for clinical and molecular features (Supplementary Fig. S1a, online resource).

Ethical approval and consents were given.


For the discovery and non-MBGrp4 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 1·5cm2 [1].

Molecular groups were assigned using established methods, and only tumours confidently assigned as MBGrp4 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 Chromosome arm-level 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].

Targeted gene panel (n = 168) and whole-exome (n = 4) sequencing was carried out to interrogate the mutational status of 63 putative MB driver genes (Supplementary Table 2, online resource). KBTBD4 mutations in the Kelch motif were assessed by Sanger sequencing [5, 23]. PRDM6 and GFI1/1B expression was evaluated by RNA-sequencing using established methods [23, 25, 33].

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 MBGrp4 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, 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.


As anticipated, MBGrp4 tumours typically arose in older children (median 8 years [0·2–20], p = 0·0002), displayed classic pathology (n = 277/328 [84%], p < 0·0001), and were enriched for isochromosome 17q (i17q) (n = 196/353 [56%], p < 0·0001; Supplementary Table 1, online resource) compared to a non-MBGrp4 comparator cohort [26]. Subgroup 8 predominated (n = 92/248 [37%]), followed by subgroup 7 (n = 73/248 [29%]), subgroup 6 (n = 38/248 [15%]), subgroup 5 (n = 36/248 [14%]) and subgroup 1 (n = 9/248 [3%]; Table 1, Supplementary Fig. S4a, online resource). Outside of WCAs, PRDM6 overexpression was the most common molecular event (n = 18/83 [22%]; Supplementary Fig. S1c, online resource). We found low frequencies of mutations and focal CNVs, however these converged on common biological ontologies (Supplementary Table 2, online resource). Aberrations in genes involved in transcriptional regulation (30%) were most frequent followed by mutations in chromatin remodelling (29%) genes; incorporating KMT2C (n = 16/172 [9%]), KMT2D (n = 13/172 [8%]), KDM6A (n = 13/182 [7%]) and ZMYM3 (n = 12/172 [7%])) with coalescing functions as modifiers of H3K4 and H3K27-methylation. Aberrations in genes involved in genome maintenance (13%) and PI3K/AKT signalling (6%) were also observed (Supplementary Fig. S1c, online resource). No detectable mutation was seen in ~ 20% of patients.

Table 1 Clinicopathological and molecular features of our discovery (n = 362) and validation (n = 668) cohorts

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 MBGrp4 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 MBGrp4; 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 MBGrp4, first focusing on methylation subgroups. While patients < 5 years are uncommon in MBGrp4 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).

Fig. 1
figure 1

Characterisation of MBGrp4 second-generation methylation subgroups. Distribution of a established clinicopathological characteristics and significantly enriched cytogenetic aberrations. b Gene-specific genetic alterations within MBGpr4 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 χ2 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 MBGrp4 methylation subgroup. d Kaplan–Meier plot for PFS in subgroup 5 for chromosome 13 loss. Univariable Cox proportional hazards models of PFS in MBGrp4 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

MBGrp4 methylation subgroups displayed distinct cytogenetic profiles (Fig. 1a, Supplementary Fig. S4c, online resource). Subgroup 5 harbours distinct changes including both 16q loss (21/36, 58%; p < 0·0001) and 13 loss (9/36, 25%; p = 0·03). Subgroups 6 and 7 were highly disrupted cytogenetically, with multiple WCAs (gains/losses; median = 3·5 [IQR 2–5] and 4 [IQR 1·5–6], respectively) and were enriched for the WCA-FR group (subgroup 6, n = 24/38 [63%] and subgroup 7 n = 31/73 [42%], p < 0·0001). Whilst chromosome 7 gain and chromosome 8 loss were equally distributed between subgroups 6 and 7, chromosome 11 loss was largely restricted to subgroup 6 (subgroup 6 n = 19/38 [50%] and subgroup 7 n = 6/73 [8%], p < 0·0001). In contrast, subgroup 8 showed a quiet cytogenetic profile with a lower total number of WCAs (median = 1 [IQR 1–2]) and i17q commonly its sole defining feature (n = 71/92 [77%], p < 0·0001).

Mutations and focal CNVs were less characteristic of subgroups than WCAs (Fig. 1b, Supplementary Fig. S4d, online resource). Subgroup 8 harboured mutations in chromatin remodelling genes: KDM6A (10/11 mutations detected occurred in subgroup 8, 91%), ZMYM3 (9/11, 81%), KMT2C (9/15, 60%) and KMT2D (5/10, 50%) occurred at higher frequencies, although these were not significant. Alterations in genes with overlapping functions relating to transcriptional regulation showed a significant association with subgroup 5 (p = 0·002), primarily driven by the high frequency of MYCN aberrations (p < 0·0001). Lastly, high PRDM6 expression showed significant enrichment within subgroup 8 (p = 0·04; Fig. 1b). The described enrichments of clinico-molecular features validated (Supplementary Fig. S4c, d, online resource) in an independent cohort.

MBGrp4 methylation subgroups also showed distinct survival outcomes (Fig. 1c). Subgroups 6 and 7 were associated with superior survival outcomes (subgroup 6; 5-year PFS 73%, 60–90 [95% CI], subgroup 7; 5-year PFS 82%, 73–94 [95% CI]). In comparison, subgroups 5 and 8 had poorer survival outcomes (subgroup 5; 5-year PFS 50%, 34–71 [95% CI], subgroup 8; 5-year PFS 59%, 50–71 [95% CI]).

Survival analysis was carried out to further identify risk-features within MBGrp4 methylation subgroups. Chromosome 13 loss was the sole significant predictor of improved PFS within subgroup 5 (HR 0·12, 0·02–0·94 [95% CI], p = 0·04; Supplementary Fig. S4e, online resource; log-rank test, p = 0·016, Fig. 1d). Univariable Cox regression within subgroup 6 (Fig. 1e, 2g, 2h) identified that metastatic disease was associated with poor PFS (HR 4·97, 1·19–20·82 [95% CI], p = 0·03). WCA-FR was associated with improved PFS within subgroup 6 at large (HR 0·25, 0·06–0·99 [95% CI], p = 0·048) or within a M0 restricted cohort (Supplementary Fig. S5a, online resource). Within subgroup 7, metastatic status was marginally significant (HR 4·45, 0·91–21·71 [95% CI], p = 0·06; Fig. 1f, p = 0·04, log-rank test; Fig. 1i). WCA-FR did not associate with improved PFS in subgroup 7, irrespective of metastatic disease status (Supplementary Fig. S5b, c, online resource). Survival heterogeneity in subgroup 8 could not be explained by any clinico-molecular features tested (Supplementary Fig. S4f, online resource).

Fig. 2
figure 2

Characterisation of MBGrp4 WCA groups. Distribution of a established clinicopathological characteristics and significant cytogenetic aberrations. b Genetic alterations in chromatin remodelling genes within MBGpr4 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 χ2 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, online resource. c Kaplan–Meier plot of PFS by MBGrp4 WCA groups. Univariable Cox proportional hazards models of PFS within d WCA-FR and e WCA-HR, assessing clinical and molecular features ≥ 10%. Kaplan–Meier plot of PFS by f metastatic disease and g extent of resection within WCA-FR. Kaplan–Meier plot of PFS by h sex, i metastatic disease, j subgroup 5 and k subgroup 7 within WCA-HR. 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, i17q isochromosome 17q, 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

We next defined the cytogenetic/mutational landscape of the validated MBGrp4 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 subgroup 6 and 7 (p < 0·0001) and was cytogenetically complex, harbouring many significantly enriched WCAs: whole-chromosome 3 loss, 7 gain, 8 loss, 10 loss, 11 loss, 13 loss, 17 gain, 18 gain and 22 gain. The number of WCA gains/losses and total WCAs differed significantly (p < 0·0001), with WCA-HR having few changes (median = 2 [IQR 1–3]) compared to WCA-FR (median = 6 [IQR 4–8]) (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).

The WCA-FR group (n = 94/314 [30%]) conferred excellent outcomes with a 5-year PFS of 89% (0·83–0·96 [95% CI]) compared with the WCA-HR group (n = 220/314 [70%]), which performed worse with a 5-year PFS of 59%, (0·53–0·66 [95% CI], p < 0·0001; Fig. 2c). This survival relationship validated in an external cohort: WCA-FR (n = 46/191 [24%]) 5-year PFS of 98% (0·94–1·00 [95% CI]) compared with the WCA-HR group (n = 145/191 [76%]) 5-year PFS of 73% (0·65–0·81 [95% CI]; Supplementary Fig. S6c, online resource).

Univariable survival analysis was carried out within WCA-FR and WCA-HR groups. Metastatic disease (log-rank p < 0·0001; log rank was used as WCA-FR M0 patients had no events) and STR (HR 6·34, 1·64–24·52 [95% CI], p = 0·007) was significantly associated with poorer PFS within WCA-FR (Fig. 2d, f, g); non-metastatic patients showed extremely favourable survival outcomes (5-year PFS 100%). Within WCA-HR, metastatic disease (HR 1·94, 1·23–3·00 [95% CI], p = 0·003), male sex (HR 1·94, 1·17–3·20 [95% CI], p = 0·01) as well as subgroup 5 (HR 1·86, 1·07–3·24 [95% CI], p = 0·03) were prognostic for poor outcome (Fig. 2e, h, i, j). Subgroup 7 (HR 0·49, 0·24–0·98 [95% CI], p = 0·04) was associated with improved outcomes (Fig. 2e, k).

To assess predictors of risk across MBGrp4, 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). For multivariable Cox regression analysis, we selected established MB features (metastatic disease, LCA histology, STR, sex, i17q, MYCN amplification and CSI dose) as well as biologically and clinically significant molecular factors (WCA status, subgroup 5, subgroup 7, chromosome 13 loss, chromosome 18 gain). In the multivariable analysis, metastatic disease (HR 2·32, 1·45–3·72 [95% CI], p = 0·0005), WCA-FR (HR 0·30, 0·13–0·71 [95% CI], p = 0·006), subgroup 5 (HR 2·09, 1·19–3·66 [95% CI], p = 0·01), subgroup 7 (HR 0·53, 0·28–1·03 [95% CI], p = 0·06), chromosome 13 loss (HR 0·24, 0·07–0·80 [95%CI], p = 0·02), and male sex (HR 1·70, 0·98–2·96 [95% CI], p = 0·06) were incorporated in our Cox model (Fig. 3a). Importantly, histology, MYC(N) amplification and extent of resection showed no prognostic utility and were not selected.

Fig. 3
figure 3

Identification of prognostic survival markers within MBGrp4. a Univariable (n = 323) and multivariable (n = 213) Cox regression analyses of progression-free survival in our MBGrp4 survival cohort. Established MB features were considered alongside molecular factors with a frequency ≥ 10%; only significant features (sig.) are shown. For multivariable analysis, HR and p values are shown for variables retained from backwards selection. Dose of chemotherapy was not considered due to extent of missing data. Calibration plots of multivariable Cox models within both b discovery and c validation cohorts for survival at 5 years alongside the bias-corrected c-index. 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

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 MBGrp4 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 MBGrp4.

Our refined MBGrp4 risk-stratification scheme integrates metastatic stage with both subgroup and WCA status (Fig. 4a), to balance the considerations of predictive accuracy with clinically actionable/practical risk-redistribution, whilst ensuring biological homogeneity within risk-groups. Favourable-risk group membership was defined by non-metastatic patients who are WCA-FR or subgroup 7 (40/188 [21%] of MBGrp4, 97% 5-year PFS). Membership of the very-high-risk group was defined by metastatic WCA-HR patients (68/188 [36%], 49% 5-year PFS). Remaining patients (M + patients with WCA-FR, and all M0 patients whose tumours are (i) WCA-HR and (ii) not subgroup 7) displayed a high-risk profile (80/188 [43%], 67% 5-year PFS; Fig. 4c). This scheme performed equivalently (bias-corrected c-index; 0·68) in comparison to the Cox model (bias-corrected c-index; 0·72). Importantly, this scheme returns a favourable-risk group of significant size (21% of MBGrp4), with excellent survival outcomes (97% 5-year PFS).

Fig. 4
figure 4

Refined MBGpr4 risk-stratification. a Summary of MBGrp4 risk-stratification scheme with corresponding biological annotation and clinical implications. b Reclassification of risk groups. Sankey plot showing the relationship between the current clinical schemes (SIOP-PNET5-MB [21] [NCT02066220] and SIOP-HR-MB [2] [EudraCT: 2018-004250-17]) and the MBGrp4 risk-stratification scheme. c Kaplan–Meier plot of PFS by MBGrp4 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 MBGrp4 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 MBGrp3/4 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 MBGrp4 risk-stratification scheme within the external validation cohort (Northcott et al. [23]). f Calibration plot of the MBGrp4 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

Our MBGrp4 risk-stratification scheme thus redefined risk within MBGrp4 disease (Fig. 4b); clinical standard-risk (37% of MBGrp4) was effectively eliminated and redistributed to favourable (37% of clinical standard-risk) or high-risk (63% of clinical standard-risk) groups. A significant subset of the clinical high-risk group (63% of MBGrp4) redistributed to favourable (9% of clinical high-risk) or very-high-risk (60% of clinical high-risk) groups.

To assess the predictive value of our MBGrp4 risk-stratification scheme against published and/or clinically utilised stratification models, we again compared performance at 5 years post-diagnosis (Fig. 4d). Our MBGrp4 stratification scheme (bias-corrected c-index 0·68) outperformed the scheme in current clinical use (‘current’, bias-corrected c-index 0·57), for the SIOP-HR-MB [EudraCT Number: 2018-004250-17] and SIOP-PNET5-MB [NCT02066220] clinical trials [2, 21]. Furthermore, it performed better against a cytogenetic model (‘cytogenetic’, bias-corrected c-index 0·62) proffered by Shih et al. [35], as well as a combined MBGrp3/4 scheme proposed by Gajjar et al. [15] (‘MBGrp3/4’, bias-corrected c-index 0·63). Additionally, the risk-scheme showed good calibration within the discovery cohort (Supplementary Fig. S7, online resource).

Finally, our MBGrp4 risk-stratification scheme was reproducible in the independent external validation cohort (Fig. 4e). The performance of the MBGrp4 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).


Our interrogation of the specific molecular pathology of ~ 1000 MBGrp4 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 MBGrp4, which outperforms previous schemes (clinical [2, 21], cytogenetic [35] and MBGrp3/ 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 MBGrp4 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. MBWNT represents around 10% of all MB patients and, to date, is the only 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 MBGrp4 with excellent outcomes (20% of MBGrp4, 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 MBGrp4, representing a significant advancement for the future clinical management of MBGrp4. 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 MBGrp4 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 MBGrp4 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 MBGrp4 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 MBGrp4. 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 MBGrp4. 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 MBGrp4 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 MBGrp4 and now require urgent assessment in prospective clinical trials, as a basis for improved diagnostics, personalised therapies and risk-adapted therapeutic strategies.