ARTICLE https://doi.org/10.1038/s41467-020-20603-4 OPEN Sarcoma classification by DNA methylation profiling Sarcomas are malignant soft tissue and bone tumours affecting adults, adolescents and children. They represent a morphologically heterogeneous class of tumours and some entities lack defining histopathological features. Therefore, the diagnosis of sarcomas is burdened 1234567890():,; with a high inter-observer variability and misclassification rate. Here, we demonstrate clas- sification of soft tissue and bone tumours using a machine learning classifier algorithm based on array-generated DNA methylation data. This sarcoma classifier is trained using a dataset of 1077 methylation profiles from comprehensively pre-characterized cases comprising 62 tumour methylation classes constituting a broad range of soft tissue and bone sarcoma subtypes across the entire age spectrum. The performance is validated in a cohort of 428 sarcomatous tumours, of which 322 cases were classified by the sarcoma classifier. Our results demonstrate the potential of the DNA methylation-based sarcoma classification for research and future diagnostic applications. A full list of authors and their affiliations appears at the end of the paper. NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications 1 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 S arcomas are a heterogeneous group of tumours, which pose defined by the WHO classification (Fig. 1a). 48 methylation challenges to pathologists. Many entities lack unequivocal classes corresponded to distinct WHO entities (category 1) morphologic or molecular hallmarks and the overall rarity comprising 45 mesenchymal tumour entities, cutaneous mela- of sarcomas result in a widespread lack of experience1,2. A high noma, cutaneous squamous cell carcinoma and Langerhans cell inter-observer variability among pathologists is reflected in con- histiocytosis. Nine methylation classes corresponded to subsets siderable discrepancy rates between primary institutions and within WHO entities (category 2) with conventional chondro- specialized referral centres with access to comprehensive mole- sarcoma dividing into four methylation classes, rhabdomyosar- cular testing3,4. Pathologists often rely on the determination of coma with MYOD1 alteration, plexiform neurofibroma, tumour specific molecular alterations if available4. While the dedifferentiated chordoma and small blue round cell tumours determination of characteristic molecular alterations most often with either BCOR alteration or CIC alteration. Three methylation consisting of translocations that generate gene fusions has classes combined WHO entities (category 3). The methylation become a diagnostic standard for many sarcoma types, approxi- class angioleiomyoma/myopericytoma and the methylation class mately half of the sarcoma entities lack unequivocal molecular atypical fibroxanthoma/pleomorphic dermal sarcoma each com- hallmarks1. Even in some cases defined by specific gene fusions, it bined two entities, while the methylation class undifferentiated may not be possible to identify adequately the fusion by FISH or sarcoma contained undifferentiated (pleomorphic) sarcoma, RNA-based methods for a variety of technical and specimen- myxofibrosarcoma and a fraction of pleomorphic liposarcoma, related limitations. Novel approaches are needed to fill these thereby providing further evidence that these sarcoma subtypes diagnostic gaps5. probably fall into a morphologic continuum of a single entity as DNA methylation is a key epigenetic mark and plays important suggested by previous genetic-based studies24–26. Two methyla- roles in normal development and disease6. In cancer, DNA tion classes point towards novel entities not yet defined by the methylation patterns reflect both the cell type of origin, as well as WHO (category 4)13,19. The methylation class SARC (RMS-like) acquired changes during tumour formation7. Profiling of human was identified in sarcomatous CNS tumours with various brain tumours has demonstrated entity-specific methylation sig- morphologic patterns not matching established tumour cate- natures and has led to the identification of several novel and clini- gories. Unifying features of cases mapping to this class are cally relevant subtypes8–13. On this basis, a comprehensive brain rhabdomyoblast-like cells and DICER1 mutations13. Methylation tumour classifier has been developed14,15. Recently, we have class SARC (MPNST-like) was reported as a subset of malignant extended the principle of methylation-based tumour profiling to peripheral nerve sheath tumours19. Cases assigning to SARC small blue round cell sarcomas evading a definite histological (MPNST-like) present similar to MPNST, however, retain diagnosis, thereby resolving these cases into established sarcoma trimethylation at histone 3 lysine 27 (H3K27me3). In addition, entities16. Further, DNA methylation-based profiling showed diag- based on 28 non-neoplastic tissue specimens methylation classes nostic potential for soft tissue and bone sarcoma subtyping17–22. In were established for non-neoplastic skeletal muscle, reactive soft this work, we aimed at developing a DNA methylation-based clas- tissue and leukocytes. Supplementary Data 1 provides basic sification tool for soft tissue and bone sarcomas representing a broad clinical information for each individual case of these methylation range of subtypes and age groups. classes. Supplementary Data 2 indicates characteristic clinical and molecular features for each methylation class. Results Development of the sarcoma classifier. We next developed a DNA methylation profiling of prototypical sarcomas. We classification tool, sarcoma classifier, using a Random Forest subjected prototypical cases of the most common soft tissue and machine learning classification algorithm as described14,27. Cross- bone tumours, non-mesenchymal tumours that might mimic validation, an internal performance metric15, of the sarcoma mesenchymal differentiation, i.e. squamous cell carcinoma or classifier provided an estimated error rate of 1.95% for raw scores melanoma, and non-neoplastic control tissue to DNA methyla- and a discriminating power of 99.9% by area under receiver tion profiling using the Infinium HumanMethylation450K operating characteristic curve analysis. The low rate of mis- BeadChip or EPIC array platform. Following quality control, classifications demonstrates the discriminating power of the methylation data were analysed by unsupervised hierarchical classifier algorithm (Fig. 2, Supplementary Data 2). The dis- clustering and t-Distributed Stochastic Neighbour Embedding crepancies encountered at cross-validation predominantly (t-SNE)23 thereby identifying groups of tumours sharing occurred between the four methylation classes of conventional methylation patterns (methylation classes). To minimize potential chondrosarcoma and between three methylation classes of sar- clustering artefacts at least seven cases were required for defining comas associated with BCOR alterations. Similar to the brain a methylation class, which empirically proved sufficient for tumour classifier we introduced a methylation class family score training a classifier and allowed prediction14,15. Unsupervised combining these closely related methylation classes by adding up clustering, respecting the minimal number of seven cases per their respective prediction scores. This modification reduced the group, led to the designation of 62 tumour methylation classes error rate at cross-validation to 0.65% for the raw scores. We belonging altogether to 54 histological types, and three non- employed a calibration algorithm transforming raw into cali- neoplastic control methylation classes (Fig. 1). Iterative random brated scores thereby ensuring inter-class-comparability. This down-sampling validated the stability of these methylation classes further allowed definition of a general cut-off score of 0.9 as a (Supplementary Fig. 1), and potential confounding factors such as threshold for prediction to a specific methylation class (Supple- sex, patients’ age, type of material, type of array and tumour mentary Fig. 3)14. purity were excluded (Supplementary Fig. 2). Based on 1077 tumour cases, methylation classes were assigned to four categories relating to the WHO classification (Fig. 1a). Classifier performance validated in a clinical cohort. Next, the Category 1 represents methylation classes equaling a WHO sarcoma classifier performance was validated on 428 additional entity. Category 2 represents methylation classes corresponding cases, mostly representing relapsed and refractory soft tissue and to a subgroup of a WHO entity. Category 3 represents bone tumours, enrolled in the MNP2.0, PTT2.0, INFORM or methylation classes that combine WHO entities. Category 4 NCT MASTER trials, which are focused on molecular analysis represents methylation classes of novel entities which are not yet (Supplementary Data 3)28–30. The predicted methylation class by 2 NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 ARTICLE Fig. 1 Establishing the DNA methylation-based sarcoma reference cohort. a Overview of 62 tumour and three control DNA methylation classes included in the sarcoma classifier reference cohort. The methylation classes are colour-coded and grouped according to the WHO scheme. The relation between methylation classes and the WHO defined subtypes is categorised in 4 tiers: equivalent to a WHO entity (category 1); subgroup of a WHO entity (category 2); combining WHO entities (category 3); non-defined by WHO (category 4). b Visualisation of the reference cohort methylation profiles (n = 1,077) using t-distributed stochastic neighbour embedding (t-SNE) dimensionality reduction. Individual samples are colour-coded in the respective class colour (n = 65) as given in (a). Abbreviations: LIPO, lipoma; MLS, myxoid liposarcoma; WDLS/DDLS, well differentiated liposarcoma/dedifferentiated liposarcoma; NFA, nodular fasciitis; MO, myositis ossificans; MP, myositis proliferans; DTFM, desmoid-type fibromatosis; DFSP, dermatofibrosarcoma protuberans; SFT, solitary fibrous tumour; IMT, inflammatory myofibroblastic tumour; IFS, infantile fibrosarcoma; LGFMS, low-grade fibromyxoid sarcoma; SEF, sclerosing epithelioid fibrosarcoma; LMO, leiomyoma; LMS, leiomyosarcoma; RMS (EMB), embryonal rhabdomyosarcoma; RMS (ALV), alveolar rhabdomyosarcoma; RMS (MYOD1); rhabdomyosarcoma with MYOD1 mutation; ALMO/MPC, angioleiomyoma/myopericytoma; EHE, epithelioid haemangioendothelioma; AS, angiosarcoma; GIST, gastrointestinal stromal tumour; SWN, schwannoma; NFB, neurofibroma; NFB (PLEX), plexiform neurofibroma; MPNST, malignant peripheral nerve sheath tumour; AFX/PDS, atypical fibroxanthoma/pleomorphic dermal sarcoma; AFH, angiomatoid fibrous histiocytoma; OFMT, ossifying fibromyxoid tumour; SYSA, synovial sarcoma; ES, epithelioid sarcoma; ASPS, alveolar soft part sarcoma; CCS, clear cell sarcoma of soft parts; EMCS, extraskeletal myxoid chondrosarcoma; DSRCT, desmoplastic small round cell tumour; MRT, malignant rhabdoid tumour; USARC, undifferentiated sarcoma; CCSK, clear cell sarcoma of the kidney; ESS (LG), low-grade endometrial stromal sarcoma; ESS (HG), high-grade endometrial stromal sarcoma; SCC (CUT), cutaneous squamous cell carcinoma; MEL (CUT), cutaneous melanoma; SARC, sarcoma; CTRL, control; MUS, muscle tissue; REA, reactive tissue; CB, chondroblastoma; CSA, chondrosarcoma; CSA (MES), mesenchymal chondrosarcoma; CSA (CC), clear cell chondrosarcoma; OB, osteoblastoma; OS (HG), high-grade conventional osteosarcoma; SBRCT, small blue round cell tumour; GCTB, giant cell tumour of bone; CHORD, chordoma; DD, dedifferentiated; FDY, fibrous dysplasia; LCH, Langerhans cell histiocytosis. the sarcoma classifier was compared to institutional diagnoses neoplastic cells than estimated by histological examination, (Fig. 3). A calibrated score ≥0.9 was reached for 322 of 428 cases although the mean value for tumour cell purity of 47,4% in non- (75%). The respective methylation class or -family matched with classifiable cases was only slightly lower compared to 51,3% in the institutional diagnosis in 263/428 cases (61%). A discrepant classifiable cases (Fig. 5). However, because some sarcomas with classifier prediction with a calibrated classifier prediction score low calibrated classifier scores carried unique molecular altera- ≥0.9 was encountered in 59/428 cases (14%). In these cases, tions such as ONECUT1-NUTM1 or EWSR1-TFCP2 gene fusions molecular data were screened for subtype-specific alterations. The we favour considering these as epigenetic subsets not yet covered initial diagnosis was revised in favour of the predicted methyla- by the current classifier version31,32. A heatmap for the perfor- tion class in 29/59 cases. In 26/59 cases the discrepancy between mance of the classifier in the validation set is shown in Supple- histological diagnosis and classifier prediction could not be mentary Fig. 4. resolved due to lack of entity specific mutations. The initial diagnosis was retained against the predicted methylation class in Copy number profiling of sarcomas. Independent from the 4/59 cases (Fig. 4). The reason for misleading methylation class methylation patterns used for classification, high-density DNA prediction in the latter cases, all passed the quality control steps, methylation arrays allow for determining copy number altera- remains unclear. The 0.9 threshold was not reached for 106 of 428 tions, the detection of which is of major diagnostic relevance for cases (25%). Consecutive t-SNE analysis demonstrated a position sarcomas25,26. We generated copy number variation (CNV) plots of many of these cases peripheral or outside of the methylation from all sarcomas of the reference cohort as described14. Fre- classes from the reference set. It is possible that some of these quently encountered alterations include MDM2 amplification tumours were contaminated with a higher amount of non- for well-/dedifferentiated liposarcomas, MYC amplification for NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications 3 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 radiation induced angiosarcoma or segmental chromosomal formalin-fixed tissues allowing integration in routine settings. deletions on chromosome 22q encompassing SMARCB1 for This represents a clear advantage over RNA expression profiling rhabdoid tumours. While these alterations often are characteristic dependent on fresh tumour tissue33. The detection of individual for distinct sarcoma entities, they usually are not pathognomonic methylation patterns for sarcoma entities is of special interest for because of their occasional occurrence also in other entities. those entities lacking pathognomonic gene alterations such as However, in combination with methylation profiles, CNV plots entity specific gene fusions. In the spectre of sarcomas currently frequently add to the diagnostic decision process. The frequency recognized by the classifier approximately one third of the entities of chromosomal or subchromosomal numerical alterations within do not exhibit such specific mutational events. the methylation classes/entities can be depicted by summary Heterogeneity on DNA methylation level has been described CNV plots (Supplementary Fig. 5). A systematic overview of between different tumours, but also within individual tumours for frequently observed copy number alterations is provided for each Ewing sarcoma34. On the other hand, that study also reported a methylation class (Supplementary Data 2). Molecular and clinical close to 100% accuracy of distinguishing Ewing sarcoma from characteristics of the predicted methylation class are provided in a other cell types. Nevertheless, the observation of heterogeneity on molecular classifier report (Supplementary Fig. 6). the methylation level within individual tumours contrasts with the high stability of a parameter required for tumour classifica- tion. We here describe a high stability of methylation profiles for Discussion sarcoma entities. In addition, our selection process for CpG sites We established an open-access platform allowing categorization included in the classification algorithm favours those with max- of sarcomas based on machine generated methylation data and imal distinction between tumour entities. A practical example for algorithm driven analysis. Employing DNA methylation-based the high stability of methylation profiles established by this categorization offers highly attractive features. Analyses can be approach has been presented for ependymoma with demonstra- performed on DNA extracted from paraffin-embedded and tion of primary and recurrent tumours from same patients neighbouring in almost all instances upon unsupervised clustering9. While conceptually highly attractive, the current version of the sarcoma classifier could not assign approximately 25% of the cases in the validation cohort to a DNA methylation class. This can be explained: Foremost, in its current stage the sarcoma classifier has not been trained to cover the entire spectrum of sarcoma subtypes. This does account for a portion of the 106/428 unrecognized cases exhibiting a calibrated score <0.9 (Fig. 3). Limited sample numbers for some entities will not allow identi- fying methylation subclasses as done for the chondrosarcomas splitting in four sub-categories. Future increase of the number of cases in the reference set will very likely enable detection of more methylation subgroups. A similar tendency has been observed in pilocytic astrocytomas and medulloblastomas separating now into several methylation subgroups with the clinical impact still remaining unclear7,12,35. Moreover, the DNA methylation-based Fig. 2 Cross-validation of the DNA methylation-based sarcoma classifier. approach is dependent on fairly high tumour cell content in the Heat map showing results of a threefold cross-validation of the Random samples. Our experience is best with 70% or more of all cells in a Forest classifier incorporating information of n = 1077 biologically sample constituting tumour cells36. Many sarcomas, however, independent samples allocated to 65 methylation classes. Deviations from typically contain high proportions of non-neoplastic inflamma- the bisecting line represent misclassification errors (using the maximum tory cells (Fig. 5). This circumstance might have contributed to calibrated score for class prediction). Methylation class families (MCF) are classifier output scores lower than the cut-off score of 0.9, con- indicated by black squares. The colour code and abbreviations are identical sequently prompting the tumour evaluation as unclassifiable. The to Fig. 1a. Numbers of this figure are summarized in Supplementary Data 4. effect of tumour cell purity on the classifier performance is likely Validation cohort (428 samples) Score >0.9 Score <0.9 classifier match no classifier match 322/428 (75%) 106/428 (25%) 59/428 Sarcoma subtypes CCS Chord not included Molecular review LS Concordance with Discrepancy with initial Discrepancy with initial Misleading diagnosis, molecular review diagnosis, lack of specific DFSP initial diagnosis profile SFT 263/428 (61%) supports classifier prediction mutations in respective 4/428 (1%) Sarcoma 29/428 (7%) entities, 26/428 (6%) NOS LMS GIST AS ES OS (HG) EWS RMS DSRCT MRT US Fig. 3 Validation of the sarcoma classifier. In total, 426 independent sarcoma samples were analysed. 75% matched to an established DNA methylation class with a classifier prediction cut-off score of ≥0.9. 25% reached a classifier prediction cut-off score of <0.9. Abbreviations are identical to Fig. 1a. 4 NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 ARTICLE Institutional Methylation Fig. 4 Comparison of pathological diagnosis and methylation class diagnosis class prediction prediction. Classifier validation using sarcoma cases enrolled in the Liposarcoma (myxoid) Liposarcoma (myxoid) MNP2.0, PTT2.0, INFORM or NCT MASTER trials. Institutional diagnosis Liposarcoma Well-/Dedifferentiated liposarcoma (left) and classifier prediction (right) of the 322 cases that received a Liposarcoma (dedifferentiated) methylation class prediction ≥0.9. The institutional diagnosis of 263 cases Liposarcoma (well differentiated) Dermatofibrosarcoma protuberans Lipoblastomatosis Solitary fibrous tumour matched the classifier prediction (concordant; grey bars). In 59 cases the Desmoid−type fibromatosis Dermatofibrosarcoma protuberans Inflammatory myofibroblastic tumour classifier prediction differed from institutional diagnosis, with 29 cases Solitary fibrous tumour Sclerosing epithelioid fibrosarcoma reclassified in favour of the methylation class prediction (discrepant— Infantil myofibromatosis Low−grade fibromyxoid sarcoma reclassified; blue bars), 26 cases where molecular validation analysis was Myxofibrosarcoma Inflammatory myofibroblastic tumour inconclusive (discrepant; light blue bars), and four cases with a misleading Leiomyosarcoma classifier result (discrepant – misleading; red bar). Leiomyosarcoma Rhabdomyosarcoma (embryonal) performance, it likely would reduce the number of discordant cases as suggested by a recent study pointing to a reclassification Rhabdomyosarcoma rate of 14% in sarcoma upon central review37. Rhabdomyosarcoma (alveolar) In summary, we introduce a tool based on DNA methylation Rhabdomyosarcoma (alveolar) data and on automated algorithm analysis using probability measures for sarcoma classification. We developed a webpage for Rhabdomyosarcoma (MYOD1) the scientific community listing characteristic features for the Rhabdomyosarcoma (embryonal) Angiosarcoma tumour methylation classes. This online platform also provides a Rhabdomyosarcoma (pleomorphic) Gastroinstestinal stromal tumour free upload service for locally generated methylation data, which Rhabdomyosarcoma, NOS Schwannoma Epithelioid haemangioendothelioma are analysed instantly and results are returned as molecular Malignant peripheral nerve sheath tumour Angiosarcoma Angiomatoid fibrous histiocytoma classifier report with a prediction confidence score (Supplemen- Angiomatoid fibrous histiocytoma tary Fig. 6). While the current version of the sarcoma classifier Synovial sarcoma Synovial sarcoma already includes some very rare entities, we acknowledge not to cover the entire spectrum. Analysis of additional sarcoma sam- Spindle cell carcinoma (transitional cell) Epithelioid sarcoma Epithelioid sarcoma ples, including uploaded data, subject to permission, will further Epithelioid sarcoma, proximal type Alveolar soft part sarcoma improve this tool by refining established and adding novel Alveolar soft part sarcoma Clear cell sarcoma of soft tissue methylation classes. The sarcoma classifier can be accessed at Clear cell sarcoma of soft tissue Extraskeletal myxoid chondrosarcoma www.molecularsarcomapathology.org. Desmoplastic small round cell tumour Desmoplastic small round cell tumour Rhabdoid tumour Malignant rhabdoid tumour Methods Endometrial stromal sarcoma Sample selection and quality control. All samples of the reference and validation Undifferentiated sarcoma set are from individual/different patients. All cases of the reference set had PEComa Undifferentiated sarcoma undergone rigorous morphological examination by pathologists specialized in Clear cell sarcoma of the kidney diagnosing sarcomas and also tumour-type specific molecular testing for identifi- Gastrointestinal stromal tumour Schwannoma cation of the relevant alterations, whenever possible. For each specimen, we aimed Malignant peripheral nerve sheath tumour Endometrial stromal sarcoma (low-grade) at a tumour cell content of ≥70%, with the caveat that microscopically estimated Sarcoma (RMS−like) Sarcoma (MPNST−like) tumour cell percentage is prone to being relatively imprecise. However, deter- Chondrosarcoma MCF Chondrosarcoma mining tumour cell content by random forest regression demonstrated that this Chondrosarcoma (mesenchymal) MCF Chondrosarcoma (IDH) goal was not reached for many samples38. Our usual approach was the identifi- Osteosarcoma, NOS Chondrosarcoma (mesenchymal) cation of a representative region on an H&E section followed by taking a 1.5 mm punch from the corresponding site in the formalin-fixed paraffin-embedded (FFPE) block. The validation set included sarcomas enrolled in the INFORM, Osteosarcoma (high−grade) Osteosarcoma (high−grade) NCT-MASTER, PPT and MNP2.0 studies28–30. Rare sarcoma entities have not been over-represented. However, availability determined inclusion resulting in over-representation of high-grade sarcomas in the validation set. To exclude low-quality samples from the cohort, the on-chip quality metrics of Ewing´s sarcoma all samples were checked and compared to a set of 7,500 pairs of IDAT-files. In addition, for each sample, an overall noise-level was computed using the R package Ewing´s sarcoma conumee version 1.6.0. Samples showing low quality values ranging in the 10th percentile for at least one of the sample controls (‘BC conversion I C1, C2, C3’, ‘BC Small cell tumour with BCOR alteration conversion I C4, C5, C6’ or ‘BC conversion II 1, 2, 3, 4’) and showing an overall Giant cell tumour of bone MCF Sarcoma with BCOR alteration noise level greater than 3, were excluded from this study. Chordoma Small blue round cell tumour (CIC) Giant cell tumour of bone Sarcoma, NOS Methylation array processing. All computational analyses were performed in R Chordoma version 3.4.4 (R Development Core Team, 2019). Raw signal intensities were Concordant (n = 263) Discrepant - reclassification (n = 29) obtained from IDAT-files using the minfi Bioconductor package version 1.24.0. Discrepant - not validated (n = 26) Discrepant - misleading (n = 4) Illumina EPIC and 450k samples were merged to a combined dataset by selecting the intersection of probes present on both arrays (combineArrays function, minfi). Each sample was individually normalized by performing a background correction to be dependent on the sarcoma subtype (Fig. 5). Future studies (shifting of the 5th percentile of negative control probe intensities to 0) and a dye- bias correction (scaling of the mean of normalization control probe intensities to with larger case numbers are required to elucidate the effect of 10,000) for both colour channels. Subsequently, a correction for the type of tumour purity on classifier performance. A possibility to over- material tissue (FFPE/frozen) and array (450k/EPIC) was performed by fitting come this problem might be to subtract methylation patterns univariate, linear models to the log2-transformed intensity values (removeBatch- typical for lymphocytes thereby accentuating patterns of the Effect function, limma package version 3.34.5). The methylated and unmethylated signals were corrected individually. Beta-values were calculated from the retrans- respective sarcoma entities. And lastly, our validation cohort did formed intensities using an offset of 100 (as recommended by Illumina). not receive a centralized pathological reference review. While Before further analysis was undertaken, the following filtering criteria were such centralized expert review would not affect the classifier applied: removal of probes targeting the X and Y chromosomes (n = 11,551), NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications 5 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 a ● Reference set ● Validation set classifiable ●●● ●●●● ●●● ●●●● ●● ●●●●● ●● ● ●● ●●●●● ●● ● ●●●● ●● ●●●● ●●●● ●● ●●●● ● ●●●● ●●●● ●●● ●●● ● ●●●● ●●●● ●●● ●●● ●●● ●● ●● ●● ●● ● ●●● ●● ●● ●●● ●● ●●●●● ●●●● ●● ● ●● ● ●●●● ● ●●●●● ●●●● ●● ● ●● ● ●●●● ● ● ● ●●●●●● ●●●●●● ●●●●● ● ●●●●● ● 100 ●● ●●● ●●● ●●●●●●●●●●●●●●● ●● ●●● ●● 100 ●● ●●● ●●● ●●●●●●●●●●●●●●● ●● ●●● ●● ●● ●● ●● ●● ●●● ●●●●● ●●● ●●●● ● ● ● ●●● ●●●● ●● ●●● ●●●●● ●●● ●●●● ● ● ● ●●● ●●●● ●● ●●● ● ●●● ● ● ● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●●● ● ●●● ●●●●●●●● ●● ●●● ● ●● ● ●●●●●●●● ●●● ●●●●●●●●●● ● ● ●●● ●● ●●●●●●●●●● ● ● ●●● ●●●● ● ●●●● ● ●●●●●● ● ●●●● ●● ●● ● ●●●●●● ●● ● ● ●● ●● ● ● ●●●●●● ● ●●●● ●● ●● ● ●●●●●● ●● ● ● ●● ●● ● ● ●●●●●●● ● ●●●●●● ●●● ● ●●● ●● ● ●●●●●●● ● ●●● ●●● ●●● ● ● ●● ●● ● ●● ● ●●●●●●● ●●●● ● ●● ● ●●●●●● ●●●● ● 50 ●● ● ●●● ● ●●● ●● ● ●●● ●● ●● ●● ●●● ●●● ●●●●● ● 50 ●● ● ●●● ● ●●● ●● ● ●●● ●● ●●●● ● ● ●●● ●●● ●●●●● ● ●●●● ● ● ●●● ● ●●●● ● ● ●● ● ●● ●● ● ● ● ● ●● ●● ● ● ● ● ●●● ● ● ●●● ● ●●●●●●●●● ● ●●●● ●●●● ●●●●●●●●● ● ●●●● ●●●● ●●●●● ● ● ● ●●●● ● ●●●● ● ● ● ● ●●●●● ●●●● ●●●●● ● ● ● ●●●● ● ●●● ● ● ● ● ● ●●●●● ●●●● ●●●●●●●● ●●●● ●●●●● ●● ● ●● ●● ● ●●● ●●●● ●●●● ●●● ●● ●● ●●●● ● ●●● ●●●● ●●●●●●● ●●●●● ●● ● ● ● ● ●●● ●●●● ●●●● ●●● ●● ●● ●●●● ● ●●● ●●●● ●●● ● ● ●●●●●●●●● ●●●●●●●●●● ● ●● ● ● ●●●●●●●●● ●●●● ●●●●●●● ●●●●●● ●●●●●● ●●●●● ●●●● ●●●●●●●●●● ●●●●●● ●●●●● t-SNE 2 t-SNE 2 ● ●●●● ● ●●● ● ● ●●●● ● ●●● ● ●●●●●● ●● ●●●●● ●●● ●●● ●● ● ● ●● ●● ●●●● ●●● ● ●● ●●●●●● ●●●●● ●● ●●●●●● ●● ● ● ●● ●● ●●●● ●●● ● ●● ●●●●●● ● ●●●●● ●● ●●●● ●● ●● ●●●●●●●●●●● ●● ● ●●● ●●● ● ●●●●●● ●● ●● ● ●●●●●● ●● ●●● ●●●●●● ● ●●●●● ●● ●●●● ●● ●● ●●●●●●●●●●● ●● ● ●●● ●●● ● ●●●●●● ●● ●● ● ●●●●●● ●● ●●● ● ●●●●● ●● ● ●●● ●●●● ● ●●●●● ●● ● ●●● ●●●● 0 ●●● ●●●●● ● ●●●●● ● ● ● ●●● ● ●●●●● ● ●●●● ●●●● ●●●●●●●●●●●● ●● ● ●● ● ●●●●●●●●● ●● ●● ● 0 ●●●●●●●●● ● ●●●●● ● ● ● ● ●● ● ● ●●● ●●●●●● ●●●● ●●●●●●●●●●●● ●● ● ●● ● ●●●●●●●●● ●● ●● ● ●●● ●● ● ● ●●● ●● ● ● ● ●● ● ● ●● ● ● ● ● ● ●● ● ● ●● ●● ●● ●●● ●●●● ●● ●●●●● ●● ●●●● ●● ●●● ●●● ●●●● ●● ●●●●● ● ● ●●●●●●● ● ● ●● ● ●● ● ●● ● ●●●● ● ●●●● ●● ●●● ● ●●●●●●● ● ● ●● ●●● ● ●● ● ●●●● ● ●●●● ●● ●●● ●●●● ●● ●● ● ●●●●● ● ● ●●●●● ●● ●●●● ●● ●● ● ●●●●● ● ● ●●●●● ●● ●● ●●● ●●● ●● ●●● ●●● ●●●●● ●●●●● ●● ●●●●●● ● ●●● ●● ●●●●● ●●●●● ●● ●●●●●● ● ●●● ●● ● ●●●●● ●●●● ●●●● ●● ● ●● ●● ●● ● ●●●●● ●●●● ●●●● ●● ●●● ●● ●● ● ● ●●●●●● ● ●●●●●● -50 ●● ● ● ●●●●● ●●● ●● ● ● ●● ● ●● ●● ●●●●●● ● ●●● ●● ●● ● ●●●● ●● ● -50 ●● ● ● ●●●●● ●●● ●● ● ● ●● ● ●● ●● ●●●●●● ● ●●● ●● ●● ● ●●●● ●● ● ● ●● ●●●●●● ●● ● ● ● ● ●● ●●●●●● ●● ● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ●● ●●● ●●● ●● ●●● ●●●●●● ●●●● ●● ●● ● ●● ●● ●● ●●● ● ● ● ● ● ● ●● ●●● ●● ● ● ● ●● ●●● ●●● ●● ●●● ●●●●●● ●●●● ●● ●● ● ●● ●● ●● ●●● ● ● ●● ●●●●● ●● ● ● ●● ●●●● ●● ●●●●● ●● ● ● ●● ●●●● ● ●●● ●● ●● ● ● ●●● ●● ●● ● ● ● ● ● ● ●● ●● ● ● ● ●● ●● ● ● ● -100 ●●●●●● ●● ●●● ● ●● ●● ● ●●●● ●● ●● ●●●● ●●●●● ● ● ●●● ●●●●● ● ●● -100 ●●●●●● ●● ●●● ● ●● ●● ● ●●●● ●● ●● ●●●● ●●●●● ● ● ●●● ●●●●● ● ●● ●●●●●●●●●●●● ● ●●● ●● ●● ● ●●● ●●●●●●●●●●●● ● ●●● ●● ●● ● ●●● ●●● ● ●●● ●●● ●●● ●●● ●●● ● ● ●●● ●●● ●●● ●●● ● ●●● ●●● ●●● ●●● ●● ●●● ●●●●● ●● ●●● ● ●● ● ● ●● ●● ●●● ●●●●● ●● ●●● ● ●● ● ● ●● ● ●●● ● ● ●●● ● ● ●● ●●● ● ●● ●●● ● ●●● ●● ●● ● ●●● ●● ●● -150 -150 -100 0 100 -100 0 100 t-SNE 1 t-SNE 1 ● Validation set non-classifiable ● Validation set misleading ●●● ●●●● ●●● ●●●● ●● ●●●●●●● ● ●● ●●●●●●● ● ●●●● ● ●●●●●● ●●●● ● ●●●●●● ●●●● ●●●●●●●●● ●● ●●●● ●●●●●●●●● ●● ●●● ●● ● ●●● ●● ●●● ●● ● ●●● ●● ●●●●● ● ●●● ●● ● ●● ● ●●●● ● ●●●●● ● ●●● ●● ● ●● ● ●●●● ● ● ● ●●●●●● ●●●●●● ●●●●● ● ●●●●● ● 100 ●●● ●●● ●●● ●●●●●●●●●●●● ●●● ●●● ● 100 ●●● ●●● ●●● ●●●●●●●●●●●● ● ●●● ●●● ●● ● ●●● ●● ●● ● ●●● ●● ●●● ●●● ●●● ●●●● ● ● ● ●●● ●●●● ●● ●●● ●●● ●●● ●●●● ● ● ● ●●● ●●●● ●● ●●● ● ●●● ● ● ● ●●●● ● ● ●● ●● ●●●● ● ● ●● ●● ●●●● ● ●● ● ●● ●●●●●●●● ●●●● ● ●● ● ●●●●●●●● ●●●●●●●●●● ● ● ●●● ●● ●●●●●●●●●● ● ● ●●● ●●●● ● ●●●● ● ●●●●●●● ●●●● ●● ●● ● ●●●●●● ●● ● ● ●● ●● ● ● ●●●●●●● ●●●● ●● ●● ● ●●●●●● ●● ● ● ●● ●● ● ● ●●●●●●● ● ●●●●●● ●●● ● ● ●● ● ● ●●●●●●● ● ●●● ●●● ●●● ● ● ●● ● ● ●● ● ●●●●●●●●●● ●●●●●●● ●● ● ●●●●●●●● ●●●●●●● 50 ●● ●● ●●● ●●● ●● ● ●●● ● ●●● ●● ● ●● ● ● ●● ●●● ●● ● 50 ●● ●● ●●● ●●● ●● ● ●●● ●● ●●●● ● ●● ● ● ● ● ●● ●●● ●● ● ● ●● ●● ● ●● ●● ● ●● ●● ● ● ● ●● ●● ● ● ● ●●● ●● ● ●●● ●● ●●●●●●●●● ● ●●●● ●●● ● ●●●●●●●●● ● ●●●● ●●● ● ●●●●● ● ● ● ●●●● ● ●●●● ● ● ● ● ●●●●● ●●●● ●●●●● ● ● ● ●●●● ● ●●● ● ● ● ● ● ●●●●● ●●●● ●●●●●●● ●●●●● ●● ● ●● ● ●●● ●● ● ●● ●●●● ●●● ●● ●● ●●●● ● ●●● ●●●● ●●●●●●● ●●●●● ●● ● ● ● ● ●●● ●● ● ●● ●●●● ●●● ●● ●● ●●●● ● ●●● ●●●● ●● ●●●●●●●●●● ● ●● ● ●●●●●●●●● ●● ●●●●●●●●●● ● ●● ● ●●●●●●●●● ●● ●●●●●●●●●● ●●●●●● ●●●●● ●● ●●●●●●●●●● ●●●●●● ●●●●● t-SNE 2 t-SNE 2 ● ●●●● ● ●●● ● ● ●●●● ● ●●● ● ●●●●●● ●●●●● ●● ●●●●●● ●● ● ● ●● ●● ●●●● ●●● ● ●● ●●●●●● ●●●●● ●● ●●●●●● ●● ● ● ●● ●● ●●●● ●●● ● ●● ●●●●●● ● ●●●●● ●● ●● ● ●● ●● ●●●●●●●●●●● ●● ● ●●● ●●●● ●●●●●● ●● ●● ● ●●●●●● ●● ●●● ●●●●●● ● ●●●●● ●● ●●● ●● ●● ●●●●●●●●●●● ●● ● ●●● ●●●● ●●●●●● ●● ●● ● ●●●●●● ●● ●●● ●● ●●●●● ●● ● ●●● ●●●● ●● ●●●●● ●● ● ●●● ●●●● 0 ●●●●● ● ●●●● ● ●●● ● ● ● ●●●● ●● ●●● ● ● ● ●● ●●●● ●●●●●●●●●●●● ●●● ● ●●●●●●●●● ●● ●● ● 0 ●●●●● ● ●●●● ● ●●● ● ● ● ●●●● ●● ●●● ● ● ● ●● ●●●● ●●●●●●●●●●●● ●●● ● ●●●●●●●●● ●● ●● ● ●●● ●● ●●● ● ●●● ●● ●●● ● ●● ●● ●● ●● ● ● ● ● ●● ●●●● ●● ●●● ●●●●● ●●●● ●● ●●● ●● ●● ●●●● ●● ●●● ●●●●● ●●●● ●● ●●●●● ● ● ●● ●● ●●● ● ● ●● ●● ● ●● ● ● ●●●●●● ● ● ●●●● ● ●●●● ● ●● ●● ● ● ●● ●● ●●● ● ● ●● ●● ● ●● ● ● ●●●●●● ● ● ●●●● ● ●●●● ● ●● ●● ●● ●● ●●● ●●● ●●●●● ●● ●● ●● ●●● ●●● ●●●●● ●● ●●● ●●● ●●●●● ●●●●● ●● ●●●●●●●●● ●● ● ●●●●● ●●●●● ●● ●●●●●● ● ●●● ●● ●●●●● ●●●● ●●●● ●● ● ● ● ●●●●● ●●●● ●●●● ●●● ● ● ●● ●● ●●●● ●● ●● ●● ●●●● ●● -50 ●● ● ● ●●●●● ●●● ●● ● ● ● ●●●● ● ●● ●● ●●●●●● ● ●●● ●● ●● ● ●●●● ●● ● -50 ●● ● ● ●●●●● ●●● ●● ● ● ● ●●●● ● ●● ●● ●●●●●● ● ●●● ●●●●●● ● ●●●● ● ● ●●●●● ●● ● ● ● ● ●●●●● ●● ● ● ● ●● ● ●● ● ● ● ● ● ●● ●●● ●● ● ● ● ●● ●●● ●●● ●● ●●● ●●●●●● ●● ●● ●● ● ●● ● ●● ●● ●● ●●● ● ● ●● ● ● ● ● ●● ●●● ●● ● ● ● ●● ●●● ●●● ●● ●●● ●●●●●● ●● ●● ●● ● ●● ● ●● ●● ●● ●●● ● ●●● ●●●● ●● ●● ● ●● ●● ●● ●●●● ●● ●● ● ●● ●●●● ● ●●● ●● ●● ● ● ●●● ●● ●● ● ● ● ● ● ●● ●● ● ● ● ●● ●● ● ● ● -100 ●● ●● ● ●● ● ●●●● ●● ●● ●●● ●●●●● ● ●● -100 ●● ●● ● ●● ● ●●●● ●● ●● ●●● ●●●●● ● ●● ●●●●●● ● ●●●●●●●●●●●● ●●●●●●● ●●●●● ● ●●●● ●● ●● ● ●●●●●● ● ●●●●●●●●●●●● ●●●●●●● ●●●●● ● ●●●● ●● ●● ● ●●● ●●● ●●● ● ●●● ●●● ●●● ●●● ●●● ● ● ●●● ●●● ●●● ●●● ● ●●● ●●● ●●● ●●● ●● ●●● ●●●●● ●● ●●● ● ●● ● ● ●● ●● ●●● ●●●●● ●● ●●● ● ●● ● ● ●● ● ●●● ● ● ●●● ● ● ●● ●●● ● ●● ●●● ● ●● ●●● ● ●● ● ●● ●●● ● ●● -150 -150 -100 0 100 -100 0 100 t-SNE 1 t-SNE 1 c 500 Reference set Cumulative [%] 100 Case number mean 47,5% 400 1 75 300 Calibrated Score 50 0.9 200 100 25 0.8 10 20 30 40 50 60 70 80 90 100 0.7 Purity [%] 0.6 200 Validation set classifiable 0.5 Cumulative [%] 100 Case number mean 51,3% 150 75 35 40 45 50 55 60 100 50 Purity [%] 50 25 10 20 30 40 50 60 70 80 90 100 Purity [%] 40 Validation set non-classifiable Cumulative [%] 100 Case number mean 47,4% 30 75 20 50 10 25 10 20 30 40 50 60 70 80 90 100 Purity [%] Fig. 5 Impact of tumour cell purity on classifier performance. a Unsupervised clustering of the combined reference (n = 1077) and diagnostic cohort (n = 428) using t-SNE dimensionality reduction. The reference set is indicated in the upper left plot. The diagnostic samples coded as classifiable (n = 318, grey dots; upper right plot), non-classifiable (n = 106, blue dots; lower left plot) and misleading (n = 4, red dots; lower right plot). The classifiable cases show high overlap with the reference cases. The non-classifiable cases frequently fall in the periphery of or are completely separate from the reference samples. b Tumour cell purity histogram plots of the reference set and the validation set subdivided into classifiable and non-classifiable cases. The mean value is indicated as dashed red line and provided as number [%]. c Tumour cell purity plotted against calibrated score for conventional osteosarcoma cases of the validation set. 6 NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 ARTICLE removal of probes containing a single-nucleotide polymorphism (dbSNP132 at the prespecified common threshold of 0.9. This allows us to establish a common Common) within five base pairs of and including the targeted CpG-site (n = 7998), threshold for all forthcoming updates of the proposed classifier, which facilitates probes not mapping uniquely to the human reference genome (hg19) allowing for the communication with clinicians. A scheme summarizing the classifier algorithm one mismatch (n = 3,965), and 450k array probes not included on the EPIC array. steps is provided in Supplementary Fig. 8. In total, 428,230 probes were kept for downstream analysis. Methylation class families. Misclassification errors mainly occurred within seven Unsupervised analysis groups of histologically and biologically closely related tumour methylation classes. t-SNE. To perform unsupervised non-linear dimension reduction, the 10,000 most Therefore, we defined three ‘methylation class families’ (MCF) encompassing these variable probes according to standard deviation were selected. The t-SNE plot was seven tumour groups. Calibrated MCF score were calculated by summing up the then computed via the R package Rtsne (version 0.13) using 3000 iterations and a calibrated class scores within one MCF. perplexity value of 30. In addition, to assess the stability of the resulting projection, we repeated the t-SNE 500 times for subsamples of 90% of the data, sampled Estimating tumour purity from DNA methylation data. The estimated tumour without replacement. purity for all reference cases was computed using the R package RF_Purify as described38. For the illustrations, the predictions obtained with the method Hierarchical clustering. Unsupervised hierarchical clustering was performed using ‘ABSOLUTE’ were used. the 20,000 most variably methylated CpG sites across the dataset according to median absolute deviation, Euclidean distance and Ward’s linkage method. Copy number profiling. Copy number alterations of genomic segments were inferred from the methylation array data based on the R-package conumee after Classifier development. Similar to the development of the brain tumour classifier14 additional baseline correction (https://github.com/dstichel/conumee). Summary the Random Forest27 algorithm (R package randomForest version 4.6-12) was copy number profiles were created by summarizing these data in the respective set applied to generate 10,000 binary decision trees, incorporating genome-wide infor- of reference cases for each methylation class. mation from all 1077 reference samples of the 65 methylation classes. We used the 10.000 CpGs with highest variable importance. In addition, to address unequal class Validation analysis. Cases enrolled in INFORM and NCT MASTER were sub- size we performed downsampling as described14. The distribution of these CpGs jected to total RNA and whole-exome sequencing; cases enrolled in MNP2.0 and position within the gene region and their regulatory feature group are indicated PTT2.0 were subjected to a customized gene panel NGS42 and total RNA (Supplementary Fig. 7). Each binary decision tree assigns a given sample to one of the sequencing from FFPE material43, whenever necessary. 65 classes, resulting in aggregate raw scores. To enable the comparison of classifier results between classes, these are transformed to a probability that measures the confidence in the class assignment (the calibrated score) by a L2-penalized multi- Reporting summary. Further information on research design is available in the Nature nomial logistic regression calibration model (R package glmnet version 2.0-18). Research Reporting Summary linked to this article. Cross-validation of the Random Forest classifier resulted in an estimated error rate of 1.95% for raw scores and 0.65% for calibrated scores and a multi-class area under receiver operating characteristic curve39 of 0.99 and a Brier score40 of 0.05. This Data availability indicates a high discriminating power. To be able to classify samples from biologically Methylation data required for building the sarcoma classifier (reference set) were closely related tumour classes, we introduced methylation class families. In those the deposited at the public repository Gene Expression Omnibus under the accession calibrated scores were added to one score for the methylation class family14. number GSE140686. Supplementary Data 1 indicates the IDAT file names for each case. The remaining data are available within the Article, Supplementary Information or available from the authors upon request. Classifier calibration. To obtain classifier scores that are comparable between classes and that are improved estimates of the certainty of individual predictions, we performed a classification score recalibration by mapping the original scores to Received: 23 March 2020; Accepted: 8 December 2020; more accurate class probabilities15. To find such a mapping, a L2-penalized, multinomial, logistic regression model was fitted, which takes the methylation class as the response variable and the Random Forest scores as explanatory variables. The R package glmnet41 was used to fit this model. In addition, the model was fitted by incorporating a small ridge-penalty (L2) on the likelihood to prevent overfitting, as well as to stabilize estimation in situations in which classes are References perfectly separable. Independent Random Forest scores are needed to fit this 1. Fletcher, C. D. M., Bridge, J. A., Hogendoorn, P. C. W. & Mertens, F. WHO model, that is, the scores need to be generated by a Random Forest classifier that Classification of Tumours of Soft Tissue and Bone (IARC Press, Lyon, 2013). was not trained using the same samples, otherwise the Random Forest scores would 2. Gatta, G. et al. Rare cancers are not so rare: the rare cancer burden in Europe. be systematically biased and not comparable to scores of unseen cases. As such, Eur. J. Cancer 47, 2493–2511 (2011). Random Forest scores generated by the threefold cross-validation are used. To validate the class predictions generated by using the recalibrated scores of the 3. Ray-Coquard, I. et al. Sarcoma: concordance between initial diagnosis and calibration model, a nested threefold cross-validation loop is incorporated into the centralized expert review in a population-based study within three European main threefold cross-validation that validates the Random Forest classifier15. regions. Ann. Oncol. 23, 2442–2449 (2012). Within each cross-validation run this nested threefold cross-validation is applied to 4. Italiano, A. et al. Clinical effect of molecular methods in sarcoma diagnosis generate independent Random Forest scores, which are then used to train a cali- (GENSARC): a prospective, multicentre, observational study. Lancet Oncol. bration model. The predicted Random Forest scores resulting from predicting the 17, 532–538 (2016). one-third test data of the outer cross-validation loop are then recalibrated by 5. Topol, E. J. High-performance medicine: the convergence of human and applying the calibration model that was fitted on the Random Forest scores gen- artificial intelligence. Nat. Med. 25, 44–56 (2019). erated during the nested cross-validation. 6. Lokk, K. et al. DNA methylome profiling of human tissues identifies global and tissue-specific methylation patterns. Genome Biol. 15, r54 (2014). 7. Hovestadt, V. et al. Decoding the regulatory landscape of medulloblastoma Calibration model parameter tuning. To determine the optimal amount of L2- using DNA methylation sequencing. Nature 510, 537–541 (2014). penalization for a calibration model a parameter tuning is performed using a 8. Sturm, D. et al. Hotspot mutations in H3F3A and IDH1 define distinct epigenetic resampling approach. To this end, each time a calibration model is fitted using raw and biological subgroups of glioblastoma. Cancer Cell 22, 425–437 (2012). RF scores from training data to calibrate raw RF scores from test data, 500 random 9. Pajtler, K. W. et al. Molecular classification of ependymal tumors across All data sets are generated by sampling 70% of the raw scores training data without CNS compartments, histopathological grades, and age groups. Cancer Cell 27, replacement. For each of these random data sets, L2-multinomial logistic regres- 728–743 (2015). sion models were fitted applying a range of reasonable penalization parameters 10. Sturm, D. et al. New brain tumor entities emerge from molecular classification lambda. The remaining 30% scores were then calibrated by these models and maximum of the calibrated scores over all methylation classes was used to generate of CNS-PNETs. Cell 164, 1060–1072 (2016). class predictions. Then a new binary class was defined, that is, predictions in 11. Sahm, F. et al. DNA methylation-based classification and grading system for agreement with the actual true class were considered ‘classifiable’ and predictions meningioma: a multicentre, retrospective analysis. Lancet Oncol. 18, 682–694 not in agreement were labelled ‘non-classifiable’. This new binary variable and the (2017). accompanying maximum score over all class scores was then analysed by a receiver 12. Reinhardt, A. et al. Anaplastic astrocytoma with piloid features, a novel operator characteristics (ROC), i.e. calculating the Youden index (Specificity + molecular class of IDH wildtype glioma with recurrent MAPK pathway, Sensitivity − 1) for all possible thresholds. The final lambda was then determined CDKN2A/B and ATRX alterations. Acta Neuropathol. 136, 273–291 (2018). such that the average Youden index over all resampling iterations at the pre- 13. Koelsche, C. et al. Primary intracranial spindle cell sarcoma with specified cut-off threshold of 0.9 is maximal15. By tuning the calibration model in rhabdomyosarcoma-like features share a highly distinct methylation profile this way we can regulate the amount of calibration so that the scores perform well and DICER1 mutations. Acta Neuropathol. 136, 327–337 (2018). NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications 7 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 14. Capper, D. et al. DNA methylation-based classification of central nervous and therapeutically relevant gene fusions. Acta Neuropathol. https://doi.org/ system tumours. Nature 555, 469–474 (2018). 10.1007/s00401-019-02039-3 (2019). 15. Maros, M. E. et al. Machine learning workflows to estimate class probabilities for precision cancer diagnostics on DNA methylation microarray data. Nat. Protoc. 15, 479–512 (2020). Acknowledgements 16. Koelsche, C. et al. Array-based DNA-methylation profiling in sarcomas with This work was funded by Deutsche Krebshilfe grant 70112499, the NCT Heidelberg and small blue round cell histology provides valuable diagnostic information. Mod. an Illumina Medical Research Grant. Part of this work was funded by the National Pathol. 31, 1246–1256 (2018). Institute of Health Research (to S.B. and Z.J.) and to UCLH Biomedical research centre 17. Renner, M. et al. Integrative DNA methylation and gene expression analysis in (BRC399/NS/RB/101410). Human tissues were obtained from University College Lon- high-grade soft tissue sarcomas. Genome Biol. 14, r137 (2013). don NHS Foundation Trust as part of the UK Brain Archive Information Network 18. Seki, M. et al. Integrated genetic and epigenetic analysis defines novel (BRAIN UK, Ref: 18/004) which is funded by the Medical Research Council and Brain molecular subgroups in rhabdomyosarcoma. Nat. Commun. 6, 7557 (2015). Tumour Research UK. The methylation profiling at NYU is supported by a grant from 19. Rohrich, M. et al. Methylation-based classification of benign and malignant the Friedberg Charitable Foundation (to M.Sn.). M.Mi. would like to thank the Lux- peripheral nerve sheath tumors. Acta Neuropathol. 131, 877–887 (2016). embourg National Research Fond (FNR) for the support (FNR PEARL P16/BM/ 20. Wu, S. P. et al. DNA methylation-based classifier for accurate molecular 11192868 grant). diagnosis of bone sarcomas. JCO Precis. Oncol. https://doi.org/10.1200/ PO.17.00031 (2017). Author contributions 21. Weidema, M. E. et al. DNA methylation profiling identifies distinct clusters in C.K. and A.v.D. conceived and supervised the project. C.K., D.Sc., D.St., M.Si., V.Ho., angiosarcomas. Clin. Cancer Res. 26, 93–100 (2020). M.Sc., M.B.H., D.C., D.T.W.J., S.M.P. and A.v.D. performed DNA methylation data 22. Kommoss, F. K. F. et al. DNA methylation-based profiling of uterine analysis and interpretation. C.K., D.Sc., D.St., M.Si., F.Sa., D.E.R., M.Bl., B.Wo., C.He., neoplasms: a novel tool to improve gynecologic cancer diagnostics. J. Cancer K.Be., P.Ho., S.Kr., E.Pf., S.St., P.Jo., F.Se., J.Ec., D.St., A.Re., A.K.W., P.Si., A.Eb., A.Su., Res. Clin. Oncol. 146, 97–104 (2020). F.F.K., B.Ca., A.Ko., F.K.F.K., M.Kr., M.Ko., A.St., B.B., H.G., C.Hei., W.H., D.T.W.J., S. 23. van der Maaten, L. & Hinton, G. Visualizing Data using t-SNE. J. Mach. F., S.M.P. and A.v.D. performed validation data analysis and interpretation. C.K., T.Mi., Learn. Res. 9, 2579–2605 (2008). K.W.P., O.Wi., A.Ku., L.R.P., T.G.P.G., T.Ki., W.Wi., M.Pl., A.Un., M.Uh., A.Ab., J.De., 24. Idbaih, A. et al. Myxoid malignant fibrous histiocytoma and pleomorphic B.Le., C.Th., M.Ha., W.Pa., C.Ha., O.St., M.Pr., J.He., S.Fr., Y.M.H.V., M.E.W., T.Me., liposarcoma share very similar genomic imbalances. Lab. Invest. 85, 176–181 K.Gr., E.d.A., J.D.M., M.A.I.G., K.T.C., S.Y.Y.L., A.Cu., M.Mi., M.My., S.Ru., U.Sc., V.F. (2005). M., J.Sc., J.Se., M.Sn., R.B., T.Kl., R.Bu., M.G., P.W., W.N.M.D., S.B., Z.J., Ly.I., P.S., O.M. 25. Barretina, J. et al. Subtype-specific genomic alterations define new targets for T., M.Sa., J.M., J.A., X.G.M., S.M., M.E., J.K.B., M.L., E.W., C.A., A.F., U.D., P.H., D.B., soft-tissue sarcoma therapy. Nat. Genet. 42, 715–721 (2010). C.V., G.M., U.F., I.P., S.F., S.M.P. and A.v.D. provided cases and meta data. C.K., D.Sc., 26. Cancer Genome Atlas Research Network. Electronic address: elizabeth. D.St. and M.Si. created the figures. C.K., D.Sc., D.St., M.Si. and A.v.D. wrote the demicco@sinaihealthsystem.ca; Cancer Genome Atlas Research Network manuscript. The manuscript underwent an internal collaboration-wide review process. Comprehensive and integrated genomic characterization of adult soft tissue sarcomas. Cell 171, 950–965 e928 (2017). 27. Breiman, L. Random forests. Mach. Learn 45, 5–32 (2001). Funding Open Access funding enabled and organized by Projekt DEAL. 28. Worst, B. C. et al. Next-generation personalised medicine for high-risk paediatric cancer patients—The INFORM pilot study. Eur. J. Cancer 65, 91–101 (2016). 29. Selt, F. et al. Pediatric targeted therapy: clinical feasibility of personalized Competing interests diagnostics in children with relapsed and progressive tumors. Brain Pathol. A patent for a DNA methylation-based method for classifying tumour species of the 26, 506–516 (2016). brain has been applied for by the Deutsches Krebsforschungszentrum Stiftung des 30. Horak, P. et al. Precision oncology based on omics data: The NCT Heidelberg öffentlichen Rechts and Ruprecht-Karls-Universität Heidelberg (EP 3067432 A1) with S. experience. Int. J. Cancer 141, 877–886 (2017). M.P., A.v.D., D.T.W.J., D.C., V.Ho., M.Si., M.B.H. and M.Sc. as inventors. The other 31. Dickson, B. C. et al. NUTM1 Gene fusions characterize a subset of undifferentiated authors declare no competing interests. soft tissue and visceral tumors. Am. J. Surg. Pathol. 42, 636–645 (2018). 32. Watson, S. et al. Transcriptomic definition of molecular subgroups of small round cell sarcomas. J. Pathol. 245, 29–40 (2018). Additional information 33. Khan, J. et al. Classification and diagnostic prediction of cancers using gene Supplementary information is available for this paper at https://doi.org/10.1038/s41467- expression profiling and artificial neural networks. Nat. Med. 7, 673–679 (2001). 020-20603-4. 34. Sheffield, N. C. et al. DNA methylation heterogeneity defines a disease spectrum in Ewing sarcoma. Nat. Med. 23, 386–395 (2017). Correspondence and requests for materials should be addressed to A.v.D. 35. Hovestadt, V. et al. Robust molecular subgrouping and copy-number profiling of medulloblastoma from small amounts of archival tumour material using high- Peer review information Nature Communications thanks Rosandra Kaplan and the density DNA methylation arrays. Acta Neuropathol. 125, 913–916 (2013). other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer 36. Capper, D. et al. Practical implementation of DNA methylation and copy- reviewer reports are available. number-based CNS tumor diagnostics: the Heidelberg experience. Acta Neuropathol. 136, 181–210 (2018). Reprints and permission information is available at http://www.nature.com/reprints 37. Perrier, L. et al. The cost-saving effect of centralized histological reviews with soft tissue and visceral sarcomas, GIST, and desmoid tumors: The experiences of the Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in pathologists of the French Sarcoma Group. PLoS ONE 13, e0193330 (2018). published maps and institutional affiliations. 38. Johann, P. D., Jäger, N. & Pfister, S. M. RF_Purify: a novel tool for comprehensive analysis of tumor-purity in methylation array data based on random forest regression. BMC Bioinformatics https://doi.org/10.1186/ Open Access This article is licensed under a Creative Commons s12859-019-3014-z (2019). Attribution 4.0 International License, which permits use, sharing, 39. Hand, D. J. & Till, R. J. A simple generalisation of the area under the roc curve adaptation, distribution and reproduction in any medium or format, as long as you give for multiple class classification problems. Mach. Learn. 45, 171–186 (2001). appropriate credit to the original author(s) and the source, provide a link to the Creative 40. Brier, G. W. Verification of forecasts expressed in terms of probability. Commons license, and indicate if changes were made. The images or other third party Monthly weather review 78, https://doi.org/10.1175/1520-0493(1950)0782.0. material in this article are included in the article’s Creative Commons license, unless CO;2 (1950). indicated otherwise in a credit line to the material. If material is not included in the 41. Friedman, J., Hastie, T. & Tibshirani, R. Regularization paths for generalized article’s Creative Commons license and your intended use is not permitted by statutory linear models via coordinate descent. J. Stat. Softw. 33, 1–22 (2010). regulation or exceeds the permitted use, you will need to obtain permission directly from 42. Sahm, F. et al. Next-generation sequencing in routine brain tumor diagnostics the copyright holder. To view a copy of this license, visit http://creativecommons.org/ enables an integrated diagnosis and identifies actionable targets. Acta licenses/by/4.0/. Neuropathol. 131, 903–910 (2016). 43. Stichel, D. et al. Routine RNA sequencing of formalin-fixed paraffin- embedded specimens in neuropathology diagnostics identifies diagnostically © The Author(s) 2021 8 NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 ARTICLE Christian Koelsche 1,2,3,90, Daniel Schrimpf1,2,90, Damian Stichel2,90, Martin Sill4,5,90, Felix Sahm 1,2, David E. Reuss1,2, Mirjam Blattner4,6, Barbara Worst4,6,7, Christoph E. Heilig8, Katja Beck8,9, Peter Horak 8, Simon Kreutzfeldt 8, Elke Paff4,6,7, Sebastian Stark4,6,7, Pascal Johann4,6,7, Florian Selt4,7,10, Jonas Ecker4,7,10, Dominik Sturm 4,6,7, Kristian W. Pajtler4,5,7, Annekathrin Reinhardt1,2, Annika K. Wefers1,2, Philipp Sievers1,2, Azadeh Ebrahimi2, Abigail Suwala1,2, Francisco Fernández-Klett1,2, Belén Casalini2, Andrey Korshunov1,2, Volker Hovestadt 11,12, Felix K. F. Kommoss 3, Mark Kriegsmann3, Matthias Schick13, Melanie Bewerunge-Hudler 13, Till Milde4,7,10, Olaf Witt4,7,10, Andreas E. Kulozik 4,7, Marcel Kool4,5, Laura Romero-Pérez 14, Thomas G. P. Grünewald 14, Thomas Kirchner15, Wolfgang Wick 16,17, Michael Platten 18,19, Andreas Unterberg20, Matthias Uhl21,22, Amir Abdollahi21,22,23,24, Jürgen Debus21,22,23,24, Burkhard Lehner25, Christian Thomas 26, Martin Hasselblatt 26, Werner Paulus26, Christian Hartmann27, Ori Staszewski28,29, Marco Prinz 28,30,31, Jürgen Hench32, Stephan Frank 32, Yvonne M. H. Versleijen-Jonkers 33, Marije E. Weidema33, Thomas Mentzel34, Klaus Griewank 35, Enrique de Álava36,37, Juan Díaz Martín 36, Miguel A. Idoate Gastearena38, Kenneth Tou-En Chang 39, Sharon Yin Yee Low40, Adrian Cuevas-Bourdier41, Michel Mittelbronn41,42,43,44, Martin Mynarek 45, Stefan Rutkowski45, Ulrich Schüller 45,46,47, Viktor F. Mautner48, Jens Schittenhelm 49, Jonathan Serrano50, Matija Snuderl50, Reinhard Büttner 51, Thomas Klingebiel 52, Rolf Buslei53, Manfred Gessler 54, Pieter Wesseling55,56, Winand N. M. Dinjens57, Sebastian Brandner 58,59, Zane Jaunmuktane59,60, Iben Lyskjær61, Peter Schirmacher3, Albrecht Stenzinger3, Benedikt Brors62, Hanno Glimm63,64,65,66, Christoph Heining64,65,66, Oscar M. Tirado67, Miguel Sáinz-Jaspeado67, Jaume Mora68, Javier Alonso 69, Xavier Garcia del Muro 70, Sebastian Moran 71, Manel Esteller 72,73,74,75, Jamal K. Benhamida 76, Marc Ladanyi76, Eva Wardelmann 77, Cristina Antonescu 76, Adrienne Flanagan 78,79, Uta Dirksen 80,81, Peter Hohenberger 82, Daniel Baumhoer 83, Wolfgang Hartmann 84, Christian Vokuhl85, Uta Flucke86, Iver Petersen87,88, Gunhild Mechtersheimer3, David Capper 89, David T. W. Jones4,6, Stefan Fröhling 8, Stefan M. Pfister4,5,7 & Andreas von Deimling 1,2 ✉ 1 Department of Neuropathology, Institute of Pathology, Heidelberg University Hospital, Heidelberg, Germany. 2Clinical Cooperation Unit Neuropathology, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 3Department of General Pathology, Institute of Pathology, Heidelberg University Hospital, Heidelberg, Germany. 4Hopp Children’s Cancer Center Heidelberg (KiTZ), Heidelberg, Germany. 5Division of Pediatric Neurooncology, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 6Pediatric Glioma Research Group, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 7Department of Pediatric Oncology, Hematology and Immunology, Heidelberg University Hospital, Heidelberg, Germany. 8 Division of Translational Medical Oncology, National Center for Tumor Diseases (NCT) Heidelberg and German Cancer Research Center (DKFZ), German Cancer Consortium (DKTK), Heidelberg, Germany. 9Heidelberg Center for Personalized Oncology (HIPO), German Cancer Research Center (DKFZ), Heidelberg, Germany. 10Clinical Cooperation Unit Pediatric Oncology, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 11Broad Institute of MIT and Harvard, Cambridge, MA, USA. 12Department of Pathology and Center for Cancer Research, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA. 13Genomics and Proteomics Core Facility, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 14Max-Eder Research Group for Pediatric Sarcoma Biology, Institute of Pathology, Faculty of Medicine, LMU Munich, Munich, Germany. 15Institute of Pathology, Faculty of Medicine, LMU Munich, Munich, Germany. 16Neurology Clinic and National Center for Tumor Diseases, University Hospital Heidelberg, Heidelberg, Germany. 17Clinical Cooperation Unit Neurooncology, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 18Department of Neurology, Mannheim University Medical Center, University of Heidelberg, Mannheim, Germany. 19Clinical Cooperation Unit Neuroimmunology and Brain Tumor Immunology, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 20Department of Neurosurgery, Heidelberg University Hospital, Heidelberg, Germany. 21Department of Radiation Oncology, Heidelberg University Hospital, Heidelberg, Germany. 22Heidelberg Institute of Radiation Oncology (HIRO), National Center for Radiation Research in Oncology (NCRO), Heidelberg, Germany. 23Heidelberg Ion-Beam Therapy Center (HIT), Heidelberg, Germany. 24 Translational Radiation Oncology, German Cancer Consortium (DKTK), National Center for Tumor Diseases (NCT), German Cancer Research Center (DKFZ), Heidelberg, Germany. 25Department of Orthopaedics, Trauma Surgery and Paraplegiology, Heidelberg University Hospital, Heidelberg, Germany. 26Institute of Neuropathology, University Hospital Münster, Münster, Germany. 27Department of Neuropathology, Institute of Pathology, Hannover Medical School (MHH), Hannover, Germany. 28Institute of Neuropathology, Faculty of Medicine, University of Freiburg, Freiburg, Germany. 29Berta-Ottenstein-Programme for Clinician Scientists, Faculty of Medicine, University of Freiburg, Freiburg, Germany. 30 Signalling Research Centers BIOSS and CIBSS, University of Freiburg, Freiburg, Germany. 31Center for Basics in NeuroModulation (NeuroModulBasics), Faculty of Medicine, University of Freiburg, Freiburg, Germany. 32Department of Neuropathology, Institute of Pathology, Basel University Hospital, Basel, Switzerland. 33Department of Medical Oncology, Radboud University Medical Center, Nijmegen, The Netherlands. 34 Dermatopathology Bodensee, Friedrichshafen, Germany. 35Department of Dermatology, University Hospital Essen, West German Cancer Center, NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications 9 ARTICLE NATURE COMMUNICATIONS | https://doi.org/10.1038/s41467-020-20603-4 University Duisburg-Essen, Essen, Germany. 36Department of Pathology, Institute of Biomedicine of Sevilla (IBiS), Virgen del Rocio University Hospital, CSIC/University of Sevilla/CIBERONC, Seville, Spain. 37Department of Normal and Pathological Cytology and Histology, School of Medicine. University of Seville, Seville, Spain. 38Department of Pathological Anatomy, Clínica Universidad de Navarra, University of Navarra, Pamplona, Spain. 39Department of Pathology and Laboratory Medicine, KK Women’s and Children’s Hospital, Singapore, Republic of Singapore. 40 Department of Neurosurgery, National Neuroscience Institute, Singapore, Republic of Singapore. 41National Center of Pathology (NCP), Laboratoire National de Santé (LNS), Dudelange, Luxembourg. 42Luxembourg Center of Neuropathology (LCNP), Luxembourg, Luxembourg. 43 Luxembourg Centre for Systems Biomedicine (LCSB), University of Luxembourg, Luxembourg, Luxembourg. 44Department of Oncology (DONC), Luxembourg Institute of Health (LIH), Luxembourg, Luxembourg. 45Department of Pediatric Hematology and Oncology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany. 46Institute of Neuropathology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany. 47Research Institute Children’s Cancer Center Hamburg, Hamburg, Germany. 48Department of Neurology, University Medical Center Hamburg-Eppendorf, Hamburg, Germany. 49Department of Neuropathology, Institute of Pathology and Neuropathology, University Hospital of Tübingen, Tübingen, Germany. 50Department of Pathology, New York University School of Medicine, New York, NY, USA. 51Institute of Pathology, Cologne University Hospital, Cologne, Germany. 52Department of Pediatric Hematology and Oncology, University Children’s Hospital, Frankfurt/Main, Germany. 53Institute of Pathology, Sozialstiftung Bamberg, Klinikum am Bruderwald, Bamberg, Germany. 54Theodor-Boveri- Institute/Biocenter, Developmental Biochemistry, Würzburg University, Würzburg, Germany. 55Princess Máxima Center for Pediatric Oncology, Utrecht, The Netherlands. 56Department of Pathology, Amsterdam University Medical Centers/VUmc, Amsterdam, The Netherlands. 57 Department of Pathology, Erasmus MC Cancer Institute, Rotterdam, The Netherlands. 58Department of Neurodegeneration, Institute of Neurology, University College London, London, UK. 59Division of Neuropathology, The National Hospital for Neurology and Neurosurgery, University College London Hospitals NHS Foundation Trust, London, UK. 60Department of Molecular Neuroscience, UCL Queen Square Institute of Neurology, University College London, London, UK. 61Research Department of Pathology, University College London, London, UK. 62Division of Applied Bioinformatics, German Cancer Consortium (DKTK), German Cancer Research Center (DKFZ), Heidelberg, Germany. 63Translational and Functional Cancer Genomics, National Center for Tumor Diseases (NCT) and German Cancer Research Center (DKFZ), Heidelberg, Germany. 64 Department of Translational Medical Oncology, National Center for Tumor Diseases (NCT) Dresden and German Cancer Research Center (DKFZ), Dresden, Germany. 65Center for Personalized Oncology, National Center for Tumor Diseases (NCT) Dresden and University Hospital Carl Gustav Carus Dresden at TU Dresden, Dresden, Germany. 66German Cancer Consortium (DKTK), Dresden, Germany. 67Sarcoma Research Group, Oncobell Program, Bellvitge Biomedical Research Institute (IDIBELL), CIBERONC, Barcelona, Catalonia, Spain. 68Department of Pediatric Onco- Hematology and Developmental Tumor Biology Laboratory, Hospital Sant Joan de Déu, Barcelona, Catalonia, Spain. 69Pediatric Solid Tumor Laboratory, Human Genetic Department, Research Institute of Rare Diseases, Instituto de Salud Carlos III (ISCIII), Madrid, Spain. 70Medical Oncology Service, Catalan Institute of Oncology (ICO), Bellvitge Biomedical Research Institute (IDIBELL), University of Barcelona, Barcelona, Catalonia, Spain. 71Cancer Epigenetics and Biology Program (PEBC), Bellvitge Biomedical Research Institute (IDIBELL), Barcelona, Catalonia, Spain. 72 Josep Carreras Leukaemia Research Institute (IJC), Badalona, Barcelona, Catalonia, Spain. 73Centro de Investigacion Biomedica en Red Cancer (CIBERONC), Madrid, Spain. 74Institucio Catalana de Recerca i Estudis Avançats (ICREA), Barcelona, Catalonia, Spain. 75Physiological Sciences Department, School of Medicine and Health Sciences, University of Barcelona (UB), Barcelona, Catalonia, Spain. 76Department of Pathology, Memorial Sloan Kettering Cancer Center, New York, NY, USA. 77Gerhard Domagk Institute of Pathology, University Hospital Münster, Münster, Germany. 78Department of Histopathology, Royal National Orthopaedic Hospital NHS Trust, Stanmore, Middlesex, UK. 79University College London Cancer Institute, London, UK. 80Pediatrics III Pediatric Hematology, Oncology, Immunology, Cardiology, Pulmonology, West German Cancer Center, University Hospital Essen, Essen, Germany. 81International Ewing Sarcoma Study Group, German Cancer Consortium (DKTK), West German Cancer Center (WTZ), University Duisburg-Essen, Essen, Germany. 82Division of Surgical Oncology and Thoracic Surgery, Mannheim University Medical Center, University of Heidelberg, Mannheim, Germany. 83Bone Tumour Reference Centre at the Institute of Pathology, University Hospital Basel and University of Basel, Basel, Switzerland. 84Division of Translational Pathology, Gerhard Domagk Institute of Pathology, University Hospital Münster, Münster, Germany. 85Department of Pediatric Pathology, University Hospital of Schleswig-Holstein, Kiel, Germany. 86Department of Pathology, Radboud University Medical Center, Nijmegen, The Netherlands. 87Institute of Pathology, SRH Poliklinik Gera GmbH, Gera, Germany. 88Institute of Pathology, Jena University Hospital Jena, Germany. 89Department of Neuropathology, Charité Universitätsmedizin Berlin, Berlin, Germany. 90These authors contributed equally: Christian Koelsche, Daniel Schrimpf, Damian Stichel, Martin Sill. ✉email: andreas.vondeimling@med.uni-heidelberg.de 10 NATURE COMMUNICATIONS | (2021)12:498 | https://doi.org/10.1038/s41467-020-20603-4 | www.nature.com/naturecommunications