Three-hundred forty-six of 403 patients of the HIT cohort (n = 163 Sharma, n = 183 extension cohort) met the eligibility criteria (Fig. 1a). Concordance analysis of classifier versions v11b6 (used in this study) and the newer v12.5 for Group 3 or 4 classification, and classifier mb_g34 (used in this study) and the newer v12.5 for subgroup I–VIII classification revealed almost perfect agreements between the results of the respective classifier results (Fig. 1b). For clinical and molecular variables of the HIT cohort (Sharma and extension) see Table 1. 294/346 (73%) were included in further analyses based on Group 3 or 4 score (v11b6) ≥ 0.8 and subgroup I–VIII score (mb_g34) ≥ 0.8 (Figs. 1a and 2a).

Methylation profiling of the study cohort. a t-sne plot of methylation profiles of samples from the HIT cohort (Sharma and extension) included in the study, in reference to the Group 3/4 subtypes consensus cohort [24]. b Ridge plot of age distribution across subgroups I–VIII. c Distribution of metastatic (M-) status across subgroups I–VIII. d Distribution of subgroups I–VIII across Group 3 or 4. Percentages indicate percent in each subgroup. e Distribution of MYC non-amplified (non amp) and amplified (amp) cases across subgroups, numbers indicate n cases. f Distribution of MYCN non-amplified (non amp) and amplified (amp) cases across subgroups, numbers indicate n cases. g Distribution of WCA phenotypes (FR: favorable risk, SR: standard risk) across subgroups, numbers indicate n cases

Distribution of age across subgroups was similar as previously described (Fig. 2b). Subgroups III (29/40; 72.5%) and V (21/26; 80.8%) had the highest, and subgroup IV (14/29; 48.3%) the lowest proportion of M + cases (Fig. 2c). Both distribution across subgroups I–VIII as well as correlation of subgroups with Group 3 and 4 (Fig. 2d; Supplemental Fig. 3, online resource) followed pre-published principles [20, 24]. Subgroups II, III, and IV were exclusively Group 3, while subgroups I and V–VIII were mainly or exclusively Group 4 (Fig. 2d). MYC amplifications were predominantly seen in subgroups II and III (Fig. 2e; Supplemental Fig. 3, online resource), MYCN amplifications in subgroups V and VI (Fig. 2f; Supplemental Fig. 3, online resource). The majority of WCA favorable risk (FR) phenotype cases, characterized by presence of two or more of gain of chr7, or loss of chr8 or chr11, were subgroups IV, VI, and VII (Fig. 2g; Supplemental Fig. 3 and Supplemental Table 2, online resources).

PFS and OS were comparable in both the Sharma and the extension cohort (Supplemental Fig. 4, online resource) and showed only marginal differences in clinical parameters (Table 1). We thus combined both cohorts (HIT cohort) for further analysis. Patients with Group 3 or Group 4 MB differed significantly in both PFS and OS, with Group 3 MB showing a poorer outcome (5-year PFS: 41.4 ± 4.6%, 5-year OS: 48.8 ± 5.0%) compared to Group 4 MB (5-year PFS: 68.2 ± 3.7%, 5-year OS: 84.8 ± 2.8%) (Fig. 3a, b). Group 4 MB showed recurrences at later time points compared to Group 3 MB (Fig. 3a) with a significantly longer mean time to progression (Supplemental Table 3, online resource), as well as significantly longer time to death (Supplemental Table 3, online resource), while all recurrences and most deaths in Group 3 occurred within five years of diagnosis. Comparison by subgroup showed significant differences between subgroups, with subgroups II and III showing the poorest PFS (5-year PFS for II: 27.6 ± 8.2%, 5-year PFS for III: 37.5 ± 7.9%) and OS (5-year OS for II: 28.8 ± 8.7%, 5-year OS for III: 43.3 ± 8.3%), respectively (Fig. 3c, d). Subgroups VI, VII and VIII showed the best PFS (5-year PFS for VI: 76.6 ± 7.9%, 5-year PFS for VII: 75.9 ± 7.2%, 5-year PFS for VIII: 66.6 ± 5.8%; Fig. 3c, d). When separated by age, subgroup IV (all Group 3) in patients ≥ 4 years old was found to have an excellent outcome in the HIT cohort, with no events recorded (5-year PFS and OS: 100%) (Supplemental Fig. 5a, b, online resource). This however was not reproducible in the validation cohort, where no difference in PFS and OS between ≥ 4 and < 4 years of age was detected (Supplemental Fig. 5c, d, online resource). WCA FR showed significantly better PFS and OS compared to WCA SR (5-year PFS: 79.5 ± 5.6% vs 51.9 ± 3.6%, p < 0.001; 5-year OS: 86.6 ± 5.2% vs 67.0 ± 3.4%, p = 0.002), as expected (Fig. 3e, f). When SR only was analyzed, this trend was retained, but significance was not reached (PFS: p = 0.078, OS: p = 0.24; Supplemental Fig. 1 and 2, online resource). WCA-FR was inversely correlated with subgroup II or III (0/34 subgroup II and 2/40 subgroup III MBs belonged to the WCA-FR group), but highly correlated with subtype VII (7/10 clinical SR subgroup VII patients were WCA-FR).

Survival analysis of the complete HIT cohort. a and b PFS and OS, comparison between Group 3 and Group 4 (HIT cohort). c and d PFS and OS, comparison between subtypes I–VIII (HIT cohort). e and f PFS and OS, comparison between WCA FR and SR (HIT cohort). Log rank testing, p < 0.05 was considered significant. PFS progression-free survival, OS overall survival, WCA whole-chromosomal aberrations, FR favorable risk, SR standard risk

To quantify biological risk factors in a heterogeneously treated cohort, we used multivariate Cox-regression analysis. We established different models for the quantification of the effect of biological factors on progression-free survival, adjusting for use of radiotherapy (as time-dependent covariate), staging and presence of MYC/MYCN amplifications. On this backbone, we added either molecular group (Group 3/4) alone, subgroup (I–VIII) alone, WCA phenotype alone, WCA phenotype plus molecular group or WCA phenotype plus subgroup, and estimated the model fit of each of the models using the AUC of a ROC-analysis for 5-year PFS (Fig. 4b). Using this approach, we identified the combination of subgroup (I–VIII) with WCA as the model with the highest AUC (0.701), followed by the combination of molecular group (Group 3/4) with WCA (AUC 0.692) and subgroup (I–VIII) alone (AUC 0.684). This led us to select the combination of WCA and subgroup as the most powerful combination of biological parameters for further risk modeling.

Risk factor analysis. a Hazard ratio Forest plot of multivariate Cox-regression analyses of risk factors for PFS in the HIT cohort. Radiotherapy (RT): time-dependent covariate. Number of events: 118. Global p value (Log Rank): 1.0764e–12; AIC: 1184.78; Concordance Index: 0.76. n.s.: not significant. b ROC-curve for 5-year-PFS in different Cox-regression models, including AUC values in parentheses. All models contained the standard clinical parameters radiotherapy as time-dependent covariate, staging, MYC-amplification and MYCN-amplification together with the mentioned biological parameters. Forest plots of the models underlying the ROC-curves can be found in supplemental Fig. 8 and 9 (online resource)

Currently, risk-assessment and subsequent clinical decision making regarding therapy intensity are strongly based on clinical parameters, which are furthermore considered to be a function of the underlying biology. We therefore developed a model integrating clinical parameters with subgroups and WCA, reducing the potential number of combined molecular risk groups based on subgroup and WCA phenotype to a clinically applicable classification. Based on stratification schemes currently being evaluated in European trials such as SIOP PNET5, we defined clinically standard risk (clinical SR) patients as completely resected (R < 1.5 cm2), non-metastatic, non-anaplastic MB without MYC or MYCN (unless Group 4) amplification in a child older than 4 years at diagnosis. All other patients were stratified as clinically high-risk (clinical HR) (Fig. 5a). As suggested by the multivariate analysis, risk modeling solely based on clinical criteria did not predict survival in an ideal manner (Fig. 5c, d). For the development of an integrated clinico-molecular risk classification, we added the molecular risk factors subgroup II, III, V for very high risk (CM-VHR), and VII and WCA FR for low risk (CM-LR), based on published data [7, 9] and strong overlap of subgroup VII with the WCA-FR phenotype as described above (Fig. 5a, b), and additional analyses on the interplay between subgroup VII and WCA-FR (supplemental Fig. 6, online resource). Integration of these novel molecular risk factors with the clinical model to a “clinico-molecular” model (Fig. 5a, b) led to a highly informative risk modeling in non-WNT/non-SHH MB: both 5-year PFS and OS in the CM-LR stratum were 94 ± 5.7%, while in the CM-VHR stratum 5-year PFS was as low as 29 ± 6.1% and 5-year OS was 35 ± 6.5% (Fig. 5e, f). Upon bootstrapped cross-validation the new clinico-molecular model outperformed the clinical model in predicting PFS (clinical model: IBS 0.186/C-Index at 5-years 0.549 vs. clinico-molecular model: IBS 0.168/C-Index at 5 years 0.641, Supplemental Fig. 7, online resource).

Risk factor modeling. a Proposed model including clinical and molecular parameters. b Sankey plot of cases allocated to strata of clinical model (left) and to strata of clinico-molecular model (right). c and d PFS and OS comparison between Clinical SR (clinical standard risk: regimen containing primary CSI, R0, and M0) and Clinical HR (clinical high risk: infant-type therapy regimen, R + , or M +) strata of the clinical model (HIT cohort). e and f PFS and OS, comparison between CM-LR, -SR, -HR and -VHR (clinico-molecular model) strata (HIT cohort). g and h PFS and OS comparison between Clinical SR (clinical standard risk: regimen containing primary CSI, R0, and M0) and Clinical HR (clinical high risk: infant-type therapy regimen, R + , or M +) strata of the clinical model (validation cohort). i and j PFS and OS, comparison between CM-LR, -SR, -HR and -VHR (clinico-molecular model) strata (validation cohort). Log rank testing, p < 0.05 was considered significant. PFS progression-free survival, OS overall survival, LR low risk, SR standard risk, HR high risk, VHR very high risk

Of note, both events in the CM-LR stratum were not relapses: one of these patients died of a pneumonia during chemotherapy and one further patient died of a second malignancy. It is of particular interest to note, that addition of information of subgroup VII/WCA-FR in the clinical SR stratum identified patients (CM-SR) clinically regarded as standard risk to have a poor prognosis almost identical to clinically high-risk patients without additional high-risk factors (CM-HR) (Fig. 5e, f).

For validation of the clinico-molecular model, this stratification was applied to a validation cohort of 423 non-WNT/non-SHH MB. For clinical and molecular variables of the validation cohort see Table 2. Compared to the HIT cohort, the validation cohort had a higher proportion of R + /M0, of LCA MB and of WCA FR cases (Table 2). No significant differences were detected for the other parameters, including PFS and OS. Again, risk modeling solely based on clinical criteria did not predict survival well with a 5-year PFS of 75.3 ± 4.7% (SR) vs. 55.3 ± 2.8% (HR) and a 5-year OS of 87.7 ± 3.9 (SR) vs. 67.2 ± 2.6% (HR) (Fig. 5g, h). However, similar to the HIT cohort, application of the clinico-molecular model led to an improved risk prediction: the CM-LR stratum of the validation cohort showed a favorable 5-year PFS and OS of 82.1 ± 6.0% and 90.5 ± 5.3% (Fig. 5i, j), respectively, and the CM-VHR stratum a very poor 5-year PFS and 5-year OS of 47.5 ± 4.1% and 55.0 ± 4.2%, respectively (Fig. 5j, h). The CM-SR and CM-HR strata again had very similar survival outcomes upon addition of the molecular parameters, despite being clinically regarded as standard and high risk, respectively. Confirming the results in the discovery cohort, the clinico-molecular model outperformed the clinical model in both the IBS (0.166 vs 0.180) and the C-index (0.667 vs 0.614), indicating superiority of the clinico-molecular model.

In conclusion, addition of novel molecular risk markers such as methylation subgroups and WCA phenotype identifies prognostic strata clearly distinct from clinical risk categories.