The molecular landscape and associated clinical experience in infant medulloblastoma: prognostic significance of second‐generation subtypes

Biomarker‐driven therapies have not been developed for infant medulloblastoma (iMB). We sought to robustly sub‐classify iMB, and proffer strategies for personalized, risk‐adapted therapies.

Aims: Biomarker-driven therapies have not been developed for infant medulloblastoma (iMB). We sought to robustly sub-classify iMB, and proffer strategies for personalized, risk-adapted therapies. Methods: We characterized the iMB molecular landscape, including secondgeneration subtyping, and the associated retrospective clinical experience, using large independent discovery/ validation cohorts (n = 387). Results: iMB Grp3 (42%) and iMB SHH (40%) subgroups predominated. iMB Grp3 harboured second-generation subtypes II/III/IV. Subtype II strongly associated with large-cell/anaplastic pathology (LCA; 23%) and MYC amplification (19%), defining a very-high-risk group (0% 10yr overall survival (OS)), which progressed rapidly on all therapies; novel approaches are urgently required. Subtype VII (predominant within iMB Grp4 ) and subtype IV tumours were standard risk (80% OS) using upfront CSI-based therapies; randomized-controlled trials of upfront radiation-sparing and/or second-line radiotherapy should be considered. Seventy-five per cent of iMB SHH showed DN/MBEN histopathology in discovery and validation cohorts (P < 0.0001); central pathology review determined diagnosis of histological variants to WHO standards. In multivariable models, non-DN/MBEN pathology was associated significantly with worse outcomes within iMB SHH . iMB SHH harboured two distinct subtypes (iMB SHH-I/II ). Within the discriminated favourable-risk iMB SHH DN/MBEN patient group, iMB SHH-II had significantly better progression-free survival than iMB SHH-I , offering opportunities for risk-adapted stratification of upfront therapies. Both iMB SHH-I and iMB SHH-Introduction Medulloblastoma (MB), the most common malignant paediatric brain tumour, accounts for around 10% of childhood cancer deaths. Five-year overall survival (OS) rates of approximately 70% are currently achieved in non-infants (children aged over either 3 or 5 years at diagnosis, depending on national treatment philosophies) using contemporary multimodal therapies (maximal surgical resection, cranio-spinal irradiation (CSI) and adjuvant combination chemotherapy) [1].
Infant medulloblastomas (iMB;~30% of all MB patients) are associated with a poorer prognosis (5-year OS <60%) and are treated using separate approaches. Current iMB protocols aim to minimize the permanently disabling late effects associated with irradiation of the developing brain by avoidance/delay of CSI [2]. However, this must be balanced with morbidity and mortality, and any potential for salvage using CSI at a later stage [3]. Desmoplastic nodular/medulloblastoma with extensive nodularity (DN/MBEN) pathology [4] (~40% of iMB; favourable risk) is the only clinically adopted prognostic risk factor and is used as a basis for de-escalation of treatment [5]; no molecular biomarkers are in current clinical use.
Recent years have seen significant advances in our understanding of the disease-wide molecular pathology of medulloblastoma. The 2016 World Health Organisation (WHO) classification of brain tumours recognizes four consensus molecular subgroups (MB WNT , MB SHH , MB Grp3 , and MB Grp4 ) [4], however, recent studies, enabled by increased cohort sizes and profiling resolution, have identified intra-subgroup heterogeneity and described further molecular subtypes within these subgroups [6][7][8][9][10]. Importantly, subgroup-directed targeted and risk-adapted therapies are now in clinical trials for non-infant medulloblastoma based on evidence from biological studies in large retrospective cohorts and clinical trials [11][12][13]. An equivalent evidence base does not exist for iMB, which has, to date, typically only been considered biologically as part of diseasewide studies.
The first dedicated studies of the genomic landscape of iMB are only now emerging, including first prospective characterization of clinical trials cohorts [7,14,15]. Initial findings with clinical potential have emerged. iMB SHH subtypes have been described, however, studies of their clinical significance have been based on modestly-sized clinical cohorts (n = 25 [14], n = 76 [7] and n = 28 [15]) and findings are inconsistent, potentially due to cohort and treatment differences, and limited statistical power, within these cohorts. These observations now require further investigation. Importantly, these studies have focused on specific subgroups (i.e. DN/MBEN MB SHH [7] , non-metastatic DN/MBEN MB SHH [15]) and have not explored biological and clinical heterogeneity within the majority of iMB (non-DN/ MBEN and non-SHH tumours represent~60-70% of all iMBs).
Critically, large-scale, systematic biological studies are urgently required to establish the molecular landscape across all iMB diseaseincluding incidence, biological and clinical relevance of molecular features and novel subtypes (e.g. Group3/4 subtypes, iMB SHH subtypes [10,13,15])to support future clinical advances. In view of the limited clinical studies with biological annotation which have been undertaken to date, the collection and characterization of retrospective iMB cohorts offers the prime current opportunity to address these challenges. Importantly, in view of current strategies towards treatment of iMB with radiation-sparing approaches [15,16], and the common historical use of radiotherapy, its impact must be carefully considered in such retrospective studies.
We report comprehensive characterization of the molecular pathology of iMB using large historical cohorts, encompassing discovery in 202 patients with full centrally reviewed clinical and pathological annotation, and validation in 185 independent patients. We demonstrate that iMBs harbour distinct biological characteristics and clinically significant molecular subtypes within the core molecular subgroups. Using these factors, reproducible molecular subgroup-directed disease sub-classification and risk-stratification models could be derived which are independent of upfront radiotherapy and outperform current clinico-pathological schemes. These models provide a basis for personalized therapies, improved therapeutic strategies and future clinical trials.

Study cohorts
A primary discovery cohort of 202 infant tumours, <5.0 years of age (median 2.61 years) on the date of first-line surgery, was assembled from UK Children's Cancer and Leukaemia Group (CCLG) institutions and collaborating centres. All patients had systematic central clinical review and follow up ≥5 years. Central review of histological variants was performed to WHO 2016 criteria [4]. Full demographic and clinical data, including treatment protocols, are given in the Tables S1-S2. Importantly, considering the retrospective nature of the cohort, survival and sub-total resection (STR) rates were equivalent across the ascertainment period (data not shown), and patients collected post-1990 received radiotherapy at equivalent rates. A noninfant comparator cohort (patients ages 5-16 years at diagnosis) is detailed in Table S3. Additional independent iMB cohorts [6,8] were used for the discovery and validation of clinical and molecular features and for these, institutional annotation was used. Full details of external cohorts and subsets used thereof are given in Table S3, including cohort selection criteria.

Procedures
Tumours were assigned to the four consensus medulloblastoma molecular subgroups (MB WNT , MB SHH , MB Grp3 and MB Grp4 ) using established DNA methylation array-based methods [17]. Chromosome arm-level copy number aberrations (CNAs) were derived from these data as previously described [18]. TP53 status was assessed in iMB SHH where possible [4]. To identify heterogeneity within iMB SHH , class discovery was first undertaken using methylation array data for our primary discovery tumour cohort, then applied to two published datasets [6,8], together totalling 147 iMB SHH patients (see Data S1). Tumours were assigned to subgroups using a consensus of non-negative matrix factorization (NMF) [9] and t-SNE/dbSCAN [8] clustering, as previously described. For iMB Grp3 , second-generation subtypes were assigned to the combined primary discovery and validation cohorts (detailed in the Data S1) according to the 'Grp3 and Grp4 Classifier' found at https://www.molecularneuropathology.org/mnp/classif ier/7. Accession numbers for DNA methylation array profiles used for the determination of molecular subgroup/subtype status are GSE93646 [9], GSE85218 [8] (Gene Expression Omnibus) and EGAS00001001953 [6] (European Genome-Phenome Archive).
Copy number status of MYC and MYCN was defined by consensus of ≥2 of the following methods; iFISH [19,20], MLPA, Affymetrix Genome-Wide Human SNP Array 6.0 and/or Illumina HumanMethylation450 DNA methylation array [18]. Mutational data for KMT2D, SUFU, PTCH1 and TP53 in our primary discovery cohort were generated using the SureSelect target capture system (Agilent) and subsequent sequencing on the Illumina HiSeq2500 instrument.

Statistical and survival analyses
All clinico-molecular features assessed in the study are listed in Table S4; associations between features were assessed by Chi-squared and Fisher's exact tests. Univariable and multivariable Cox proportional hazards tests were used to investigate the association of features with survival. Analysis was performed using SPSS v23 (SPSS, Chicago, U.S.A.) and the R statistical environment (version 3.2.3).
Expanded methodological and statistical detail can be found in the Data S1.
iMB SHH patients within our cohort were treated heterogeneously, both at diagnosis and relapse (Table S1). We therefore first assessed whether consistent predictors of overall survival were observed across iMB SHH cohorts. STR (HR 6.7, CI 2.5-17.6, P < 0.001) and chromosome 9p gain (HR 3.3, CI 1.1-9.7, P = 0.026) were significantly associated with poorer OS, while DN/MBEN pathology (HR 0.2, CI 0.1-0.5, P = 0.001) conferred a favourable risk ( Figure 2e). No other features tested (Table S4), including the novel intra-iMB SHH molecular subtypes, metastatic disease status or treatment variables, were significantly associated with OS. OS was equivalent between DN and MBEN pathological variants ( Figure S5e).
Given the low frequency of iMB Grp4 , intrasubgroup survival modelling was not possible. Taken together, iMB Grp4 patients belonged to a high-risk group with an OS of 70%, but there were distinct survival outcomes between Grp4-enriched subtypes (subtype VIII, 63% OS, high risk; subtypes VI and VII, OS 80%, standard risk). iMB Grp4 patients were typically older (median, 2.6 years) and almost all (19/20; 95%) were treated with upfront radiotherapy.
We next investigated all features as predictors of progression-free survival (PFS; i.e. relapse following upfront therapy) in iMB (Figure 4a-d). In univariable analysis of iMB Grp3 , a series of molecular (e.g. subtypes II, IV) and clinical (e.g. receipt of CSI radiotherapy (39/ 59, 66%) receipt of focal irradiation (15/59, 25%)) were significantly associated with PFS, with considerable overlap with findings observed for OS. In multivariable analyses, MYC amplification (HR 4.1, CI 1.4-11.6, P = 0.008) and chr11 loss (HR 0.1, CI 0.06-0.5,  (Table S1). Residuals from Fisher's exact or v 2 tests indicate where subgroup enrichment has occurred (darker shades indicate stronger relationships); scale bar for residuals is shown. Remaining second-generation molecular subtypes V, I and VII (rarely iMB Grp3 ) and VI and VIII (exclusively iMB Grp4 ) are shown for reference. Black bar, positive for feature; unfilled, negative for feature; grey bar, data unavailable (b) Kaplan-Meier plot of overall survival for second-generation molecular subtypes with frequency >5% from our primary discovery cohort and the Cavalli et al external discovery/validation cohort. Subtypes with a frequency of >5% of iMB are shown (II, III, IV and VII). At risk table (number censored in parentheses) and p value from log-rank test is shown. (c) Univariable and multivariable Cox proportional hazards regression model of overall survival in discovery cohort iMB Grp3 . Features which entered multivariable analysis (P < 0.1; P < 0.05 shown in bold font) are indicated. Cox proportional hazards test is shown either uncorrected (p) or corrected for multiple testing by the Benjamini-Hochberg method (adjusted p). Additional clinical features are shown for reference (gender, age under 3 years, metastasis). (d) Summary of a novel risk-stratification scheme for overall survival iMB Grp3 (n = 62). E. Kaplan-Meier plot of iMB Grp3 risk stratification. At risk table (number censored in parentheses) and log-rank test are shown. Abbreviations: STR, sub-total resection; CLA, classic; LCA, large cell/anaplastic; CI, confidence interval; HR, high risk; VHR, very high risk; RTX, radiotherapy; CTX, chemotherapy; M+, metastatic disease M2 or above  (Figure 4b) were the only independent risk factors, consistent with OS findings for iMB SHH . When the DN/ MBEN iMB SHH patient group was discriminated and considered in isolation (n = 37), membership of iMB SHH-I was significantly associated with a worse PFS (HR 3.6, CI 1.0-11.8, P = 0.038) (Figure 4c); the only significant predictor among all variables tested (Table S4). In contrast, the iMB SHH-I subtype was not significant when considered across all iMB SHH patients, supporting the importance of diagnosing the DN/MBEN variant within iMB SHH .
Integration of our validated subgroup-dependent OS and PFS prognostication schemes (Figure 2, iMB SHH OS; Figure 3, iMB Grp3 and iMB GRP4 OS; PFS, Figure 4) allow the sub-classification of iMB patients into schema for the stratified delivery of risk-adapted therapies, based on the biomarkers discovered and therapies used in our retrospective cohorts (Figure 5a). Overall risk can be stratified straightforwardly using four validated features; consensus molecular subgroup, pathology variant, extent of resection and MYC amplification. This subgroup-directed model (Figure 5b) significantly outperformed the current, pathology-based, risk stratification [5] (5yr OS AUC 0.744 vs. 0.580) in our cohort, and was independently reproducible (FR 5yr OS 94%; HR 5yr OS 73%; VHR 5yr OS 46%; log-rank P < 0.001; Figure S6). Following definition of DN/ MBEN iMB SHH using this model, further distinction of the iMB SHH-I and iMB SHH-II subtypes enables prediction of PFS (Figure 4), while MB Group3/4 subtypes associated with 60-80% OS following upfront CSI are highlighted ( Figure 3) for further clinical investigation (Figure 5a).

Discussion
Our analysis of almost 400 iMB tumours provides critical insights into their subgroup-dependent molecular heterogeneity, its clinical relevance and potential for exploitation towards disease sub-classification and  improved, risk-adapted, therapies. Assessment of these large retrospective cohorts has enabled robust definition of the nature and reproducibility of molecular subtypes within iMB SHH (types I, II) and iMB Group3 (types I-VII), and their interaction with established disease biomarkers. Consideration of their clinical associations across independent cohorts provides strong supporting evidence for their incorporation into future research studies and clinical application. Radiation-sparing approaches have been postulated for iMB in an effort to obviate patients from deleterious, often life-limiting, late effects caused by treatment. Any biomarker discovery study based on retrospective cohorts must therefore consider the impact of radiation therapy. Our study identified risk-stratification groups that were reproducible and independent of receipt of radiotherapy. First, DN/MBEN was confirmed as a favourable-risk biomarker in our cohorts, highly associated with iMB SHH . Importantly, a notable proportion of CLA and LCA iMB SHH tumours were observed, which were associated with a very poor prognosis in both discovery and validation cohorts irrespective of therapy, clearly demonstrating that defining subgroup alone is insufficient for risk stratification in iMB SHH . Central histological review to WHO 2016 [4] standards was essential for the robust definition of histological variants; in our experience, 8/62 iMB SHH tumours were reclassified to DN/MBEN from their institutional CLA diagnosis.
Two discrete and reproducible molecular subtypes -iMB SHH-I and iMB SHH-IIwere discriminated in our analysis of a unified iMB SHH cohort totalling 155 tumours, encompassing patients from three independent studies [6,8]. This further supports the reproducibility of these molecular subtypes, as reported in previous studies [7,8,14,15], and defines their characteristics in large cohorts. Following discrimination of the favourable-risk iMB SHH DN/MBEN group, definition of iMB SHH molecular subtypes enabled prediction of PFS within our cohort, as a potential basis for the stratified delivery of upfront therapy. DN/MBEN iMB SHH-II had significantly improved PFS over iMB SHH-I independent of whether upfront radiotherapy was received. The prognostic significance of iMB SHH-I and iMB SHH-II subtypes differs between reported studies, which likely relates to differences in therapy and statistical power. Our study (large retrospective cohort; n = 37 DN/ MBEN iMB SHH , mixed therapies) and SJYC03 (clinical trial; n = 42 DN/MBEN iMB SHH , risk-adapted therapies; no differences between low-, intermediate-and highrisk strata [7]) showed improved PFS for iMB SHH-II . Conversely, HIT-2000 (clinical trial; n = 28 non-metastatic DN/MBEN iMB SHH , intraventricular methotrexate therapy [15]) reported equivalent PFS for both groups. Two further cohorts, ACNS1221 (clinical trial, n = 25 DN/MBEN iMB SHH , conventional systemic chemotherapy without intraventricular methotrexate [14]) and the HIT group/Burdenko Institute validation cohort (restrospective cohort, n = 48 DN/MBEN iMB SHH , mixed therapies [15]) did not show a statistically significant difference in PFS between iMB SHH-I and iMB SHH-II subtypes.
Controlled clinical trials using stratified therapeutic approaches should be considered for the DN/MBEN iMB SHH-II patient group, aimed at resolving its interaction with different therapies and minimizing therapy-induced late effects, while maintaining OS rates. We observed equivalent rates of rescue, which commonly involved radiotherapy, in both iMB SHH subtypes, further supporting such trials of risk-adapted therapies.
Our studies also reveal clinically actionable subgroups within iMB Grp3 . LCA pathology and MYC amplification are enriched in subtype II. Together or in isolation, they define a VHR group associated with a dismal prognosis (10yr OS 0%) and a short time to death, whether or not upfront radiation was received. These patients are refractory to current conventional treatments and often progress rapidly, with many not surviving to initiation of adjuvant therapy.
A series of better prognosis subgroups within iMB Grp3 and iMB Grp4 were noted. These include subtype IV Figure 5. Subgroup-dependent prognostic models for iMB and candidate groups for risk-adapted therapies. (a) A novel subgroup-directed risk-stratification scheme for iMB, incorporating validated OS and PFS correlates identified in our cohorts. (b) Patients mapped to the OS scheme using four routinely testable features; molecular subgroup, pathology variant, resection status and MYC amplification status. A Kaplan-Meier plot showing our entire iMB cohort stratified for OS using this subgroup-directed model and the current, pathology-based model, alongside an at-risk (defined by many frequent CNAs [21]; 80% 5yr OS), subtype VII (iMB Grp4 enriched; 80% OS) and non-MYC/ non-LCA iMB Grp3 (73% OS). However, the overwhelming majority (>75%) of these patients received standard upfront radiotherapy, and therefore the prognostic relevance of these groups in the radiona€ ıve iMB setting, and the associated use of post-relapse radiotherapy as a rescue strategy, requires further assessment.
We also assessed iMB prognostic factors defined in previous historical studies, in our cohort [5]. Metastatic disease was enriched in iMB Grp3 , but was not significantly associated with poorer survival (Figure 3). Its prognostic relevance, and its interaction with radiotherapy, thus remains unconfirmed in this patient group. Similarly, in a previous iMB cohort, STR was significantly associated with poorer OS [5]. Our analysis demonstrated its independent prognostic significance in iMB SHH ; while this may reflect historic surgical practices, outcomes for, and rates of, STR were equivalent across our collection period. Receipt of upfront focal radiotherapy was not associated with improved PFS within iMB SHH (data not shown). However, receipt of focal radiotherapy was associated with improved PFS (compared to no irradiation) in iMB Grp3 in univariable analyses. This finding is likely contributed to by the high frequency of very-high-risk iMB Grp3 patients in our cohort who received no radiotherapy at all, likely due to a clinical decision to palliate at diagnosis.
To allow maximal inclusion and assessment of clinico-biological relationships we selected patients up to 5 years of age for analysis, however, applying our subgroup-directed survival model reached equivalent findings when restricted to patients under 3 years old in our cohort ( Figure S9). This, coupled with the recent identification of age-dependent molecular subtypes within MB SHH (MB SHH-Infant , <4.3 years vs. MB SHH-Childhood ) [9], suggests that the definition of the infant disease should include patients at the upper end of the 3-5 age range in current clinical use.
As discussed for DN/MBEN iMB SHH and its subtypes, both treatment and prognostic effects may differ between clinical studies. Similarly, our retrospective study encompassed patient cohorts treated with mixed protocols. As far as possible, we controlled for age and therapy type in our survival analyses and risk modelling, and validated findings across large independent cohorts. This has enabled the identification and validation of clinically actionable biological subgroups with distinct and reproducible disease behaviours (favourable-risk DN/ MBEN iMB SHH , very-high-risk LCA/MYC iMB Grp3 ), which are independent of treatment. Cohort-specific or treatment-dependent effects, particularly with regard to emerging therapeutic concepts in iMB (e.g. high-dose chemotherapy, intrathecal therapies) must be considered in future clinically controlled studies.
In summary, assessment of the molecular pathology of iMB in large historic cohorts has allowed the robust characterization of each iMB molecular subgroup and the novel molecular subtypes they harbour. Almost a third of iMB can be reclassified into a VHR group, which, based on their dismal outcomes and rapid disease course, should be urgently considered for novel upfront therapeutic approaches, such as anti-MYC therapeutics [22]. Prognostic subtypes within FR groups (e.g. DN/MBEN iMB SHH-I , iMB SHH-II ) offer opportunities to direct the stratified use of upfront therapies. Identified HR iMB groups, with 60-80% survival rates using CSI-based regimes, are suitable for investigation in randomized clinical trials.

Consent for publication
NA.

Authors' contributions
Conception and design: DH, ECS, DW, SB and SCC. Collection and assembly of data: DH, JL, CIH, RMH, AS, PA, SR, MD, CS, SC and SB. Data analysis and interpretation: DH, GR, ECS, LP, DW, SB and SCC. Central pathological review: AJ, SW and TJ. Provision of study materials or patients: BP, AM and SB. Manuscript writing: All authors. Final approval of manuscript: All authors. Accountable for all aspects of the work: All authors.

Ethics approval and consent to participate
Human tumour samples were provided by the UK CCLG as part of CCLG-approved biological study BS-2007-04; informed consent was obtained from all subjects and investigations conducted with approval from Newcastle/North Tyneside Research Ethics Committee (study reference 07/Q0905/71).

Data Availability Statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.