In first instance, we generated tumor-normal 10X-LR datasets for the 13 samples that were part of the Northcott et al. study, and analyzed them using our in-house 10X-LR pipeline (see Methods) to identify previously undiscovered structural variants and provide a comprehensive list of somatic structural rearrangements. In addition, we re-analyzed the existing WGS data using an enhanced SV detection pipeline which uses 6 different SV callers in order to improve sensitivity (see Methods) (29). Combining the results from our WGS and 10X linked-read pipelines, we detected 265 somatic SV (which were manually confirmed using Loupe) in the 13 samples previously characterized by WGS ( Additional Table 2 , contains breakpoint coordinates for each technology and caller). Of these, 74 were detected by both 10X-LR and WGS, 173 were found by 10X only, and 18 were called by WGS only but visually confirmed in the 10X data. Our findings include mutations in recurrently mutated genes such as a SNCAIP duplication ( Figure 1A ), an inversion and an amplification in GFI1B ( Figures 1B, C ), an intrachromosomal rearrangement in GFI1 ( Figure 1D ), and 2 large amplifications which include the CDK6 locus (MDT-AP-2878, chr7: 86,891,518- 95,624,126, Figure 1E , MDT-AP-1209, chr7:90,074,594-93,426,132, Figure 1F ) all of which were previously detected by Northcott et al. and validated by 10X-LR (12). We also detected a novel complex rearrangement on chromosome 8 in MDT-AP-2130 (validated by WGS, Figure 2A ). Additionally, we detected a complex event in MDT-AP-2878 involving chromosomes 2 and 16 with a breakpoint downstream of IDH1 that had not been previously characterized but was validated in the re-analyzed WGS data ( Supplemental Figures 1A, B ).

Next, we applied 10x Genomics linked-read sequencing in 12 uncharacterized samples: 8 new G4 medulloblastomas, two G3 and two SHH medulloblastomas (see Findings in non-G4 medulloblastomas below). In the 8 G4 medulloblastomas that had not been characterized previously, we detected 147 somatic SVs that were manually confirmed using Loupe ( Additional Table 2 ). MDT-AP-2940 was found to have chromoplexy involving chromosomes 3, 5, 6, 11, 12, 13, 15 and 17 ( Figure 2B ) as well as a complex event on chr5 leading to the amplification of TERT ( Supplemental Figure 1C ). Additionally, we detected chromothripsis in one sample involving chromosome 8 co-occurring with loss of 17p which contains TP53 (MDT-AP-3743, Figures 2C, D ). Loss of TP53 is thought to be required for chromothripsis and although 17p loss in common in G4s, chromothripsis is rare (30).

MYCN was found amplified in MDT-AP-3670 and further analysis showed that this was part of a much larger complex SV and amplification with a breakpoint connecting it to a region 27.4Mb downstream on chromosome 2 ( Figures 3A–D ). Interestingly, both of these events were shown to share barcodes across the genome which suggests that there are many copies of these regions within the nucleus that are being caught within the emulsion created by the 10X-LR protocol ( Figures 3C, D ). We hypothesized that this patterning indicated ecDNA which we then validated using Hi-C data from the same sample ( Supplementary Figure 2 ). Hi-C has previously been shown to be able to detect ecDNAs in a wide-range of tumors and cell lines (31–34). Additionally, copy number calls generated from 10X-LR using TitanCNA indicated approximately 75 copies of chromosome 2 from 14.6-16.3Mb and 41.7-41.9Mb as well as even higher amplification (~125 copies) of chromosome 2 from 15.5-15.74Mb and 15.75-15.96Mb which contains additional rearrangements ( Figure 3E ). As far as we are aware, this represent the first time ecDNA has been identified using 10X linked-reads.

In terms of point mutations, SNVs described as functional SNVs and indels by Northcott et al. were manually validated in the linked-read data and as well as in the RNA-Seq where applicable ( Additional Table 3 ) (12). These include a TERT promotor mutation in MDT-AP-2130 (C228T, Supplemental Figure 1D ) as well as germline mutations in BRCA2 (p.Tyr3225IlefsTer30, Pathogenic/Likely pathogenic in ClinVar), RAD51D (p.Asp98ValfsTer25, likely pathogenic in ClinVar) and ATM (p.Arg2136Ter, Pathogenic/Likely pathogenic in ClinVar), all of which are associated with the double-stranded break repair pathway and cancer predisposition syndromes (35). Additionally, analysis of the linked-red data allowed us to detect two mutations in KDM6A, a frameshift variants in KMT2D [known to be recurrently mutated in G4-MB (10)], a mutation in CREBBP annotated as likely pathogenic in ClinVar, and a second TERT promoter mutation (C228T).KDM6A is a lysine demethylase recurrently mutated in both G3 and G4 medulloblastomas (36). The mutations were a missense mutation (p.R1255W, MDT-AP-2151, validated in WGS) previously detected in carcinomas of the pancreas, endometrium, prostate, breast, and skin as well as a truncating mutation annotated as likely pathogenic in ClinVar (R1331fs, MDT-AP-2857) found in the germline of three patients with Kabuki syndrome 2 (https://www.ncbi.nlm.nih.gov/clinvar/variation/216950/).

In the G4 medulloblastomas, SVs affecting GFI1 and GFI1B and the recurrent tandem duplication of SNCAIP are known to cause overexpression of GFI1, GFI1B and PRDM6, respectively, by putting them under the control of super-enhancer regions (termed enhancer hijacking, EH). We used the RNA-Seq data, which was available for 13 samples, to validate these EH events. Expression levels supported enhancer hijacking of GFI1B in MDT-AP-1206 and MDT-AP-2673 as well as PRDM6 in MDT-AP-0074 ( Figures 4A, B ). Interestingly, MDT-AP-2151 was also shown to have overexpression of PRDM6 despite no evidence of a tandem duplication of SNCAIP by either Northcott et al. or us (12) ( Figure 4C ). However, copy number data from TitanCNA suggests a small duplication over PRDM6 which explains the increased expression and suggests that tandem duplication of SNCAIP may not be the only mechanism leading to overexpression of PRDM6 ( Figure 4D ).

Group 4 medulloblastomas are also known to be tetraploid, with 11/21 samples in this study having a ploidy of 4 (52%, Supplemental Table 1 ). Group 4 MBs are also characterized by extensive copy number variants, two of the most characteristic being gain of chromosome 17q (14/21, 66%) with or without loss of chromosome 17p (13/21, 62%, contains TP53), as well as a gain of chromosome 7 (11/21, 52%) and loss of chromosome 8 (10/21, 47%) (14) ( Supplemental Figure 3 ).

We detected 11 manually confirmed somatic variants in the previously uncharacterized non-G4 medulloblastomas (2 in 2 SHH-MBs and 9 in 2 G3-MBs, Additional Table 2 ). One SHH tumor was found to have a TERT promoter mutation (C228T, MDT-AP-3724) as well as a previously described CREBBP mutation (p.R1446L c.4337G>T), a single-base deletion in exon 34 of lysine-specific methyltransferase 2D (KMT2D), an interchromosomal translocation between chromosomes 3 and 14 ( Supplemental Figure 4A ), 4 copies of 3q and loss of 14q24.1-q32.33 ( Supplemental Figure 4B ). The other was characterized by an interchromosomal translocation between chromosomes 7 and 18 (MDT-AP-3862, Supplemental Figure 4C ), a gain of 7q31.2-36.3, loss of 20 and loss of homozygosity on 10q which contains SUFU ( Supplemental Figure 4D ).

The group 3 medulloblastomas were mainly characterized by copy number changes and structural variants although none were recurrent between the two samples ( Additional Table 2 ). Of note, one G3-MB was found to have a germline interchromosomal translocation between chromosomes 2 and 5 occurring near 2 protocadherin genes and a protocadherin gene cluster (MDT-AP-4037, Supplemental Figures 4E, F ). Previous work suggests that protocadherins may play a role in tumorigenesis in medulloblastomas (37, 38).

We generated PacBio data from 5 G4 MB tumor-normal pairs where additional DNA material was available (7-19X coverage, Supplemental Table 1 ). In addition, we were able to generate paired tumor-normal Nanopore data from two of these G4 tumor samples at 15-30X coverage plus deep sequencing data from the tumor of MDT-AP-2673 (53x coverage). All samples with long-read data also had WGS and RNA-Seq data available ( Supplemental Table 1 ). Analysis of the long-read data allowed the detection of 16 somatic SVs that had been confirmed as somatic and included the focal events around GFI1B and SNCAIP leading to enhancer hijacking ( Additional Table 2 ). Additionally, we detected 4 SVs found uniquely by long-reads which we validated as somatic ( Additional Table 2 ).

Since 10x Genomics has discontinued their linked-read kit, we decided to test the TELL-Seq library kit by generating data for 4 tumor samples in order to compare the SV calls. We chose samples which had somatic SVs in GFI1B (MDT-AP-1206 and MDT-AP-2673), in TERT (MDT-AP-2940) and GFI1 (MDT-AP-2878) that had previously been detected by 10X-LR.QC metrics for both technologies were generated using the LongRanger pipeline. On average, the 10X-LR samples had longer mean molecules lengths compared to TELL-Seq although this is likely due to sample degradation over time since the same HMW DNA extractions we used to generate both tumor linked-read datasets about 2 years apart ( Supplemental Table 2 ). As a result, the 10X-LR data out-performed TELL-Seq in terms of the number of phased SNPs and longest phase block. Both technologies had similar numbers of large SV and short deletions calls made by LongRanger ( Supplemental Table 2 and Additional Table 1 ); however, the TELL-Seq data had much more even coverage compared to the 10X-LR data ( Supplemental Figure 5 ).

Nearly all high-quality somatic calls (detect by at least 2 callers and >10kb) made in the 10X-LR data were validated by TELL-Seq. 84-125 calls were made by both technologies, 21-44 calls were made by 10X only, and 3-6 were made by TELL-Seq only ( Figure 5A and Additional Table 4 ). 63 somatic calls were made across the 4 samples of which 49 were called by at least 2 callers in the 10X-LR dataset (the rest where either detected by WGS or a single caller in the 10X-LR datasets). Of these 49 calls, 32 were detected in both the 10X-LR and TELL-Seq datasets from the same patient ( Figure 5B ). Of the 22 somatic SVs which occurred in a gene of interest or are part of a complex genomic event such as chromoplexy, 14 were detected in both the 10X-LR and TELL-Seq datasets ( Additional Table 4 ) and included both enhancer hijacking events in GFI1B ( Figures 5C, D ) and the amplification around TERT ( Figure 5E ). Eight somatic SVs were only detected in the 10X-LR, however, manual inspection of the TELL-Seq data in Loupe allowed us to confirm visually the presence of the SVs not called by TELL-Seq including the SV affecting GFI1 in MDT-AP-2878 ( Figure 5F ).

10x Genomics linked-read libraries were generated for 25 tumors and corresponding normal samples. HMW DNA was extracted from tumors using phenol chloroform extractions or the Nanobind tissue kit (PacBio, Menlo Park, California, United States, cat# SKU 102-302-100) while matching blood samples were extracted using the QiaAmpBlood Kit (Qiagen, Hilden, Germany) or the Nanobind tissue kit (PacBio, Menlo Park, California, United States, cat# SKU 102-302-100) ( Additional Table 1 ). Size selection was done with the SRE and SRE XS kits (PacBio, Menlo Park, California, United States, cat# SKU 102-208-200 and SKU 102-208-300) as needed and dependent on the availability of DNA ( Additional Table 1 ). Concentration was assessed by Qubit™ dsDNA BR Assay Kit (ThermoFisher Scientific, cat# Q32853) and size distribution was profiled using the Femto Pulse (Genomic DNA 165 kb Kit, 3 hours run, Agilent Technologies, Inc., Santa Clara, California, United States, cat# FP-1002-0275). Samples and library preparation were done following the Chromium™ Genome Reagent Kits v2 User Guide (10x Genomics, Pleasanton, California, United States). Libraries concentration was assessed by qPCR (Roche, Basel, Switzerland, KAPA Library Quantification Kits - Complete kit (Universal), cat# 07960140001) and the size distribution of the libraries was evaluated using the Caliper LabChip (DNA High Sensitivity assay, PerkinElmer, Inc., Waltham, Massachusetts, United States). Libraries were sequenced using 150PE Illumina reads, either on a single lane of HiSeqX or pooled on a NovaSeq S4 flowcell. Average molecule length, calculated by LongRanger, ranged from 19kb-85kb for tumor samples and 42kb-104kb for normal samples ( Additional Table 1 ).

Data was analyzed as detailed in Zwaig et al. (42). In brief, alignment and variant calling was done using 10x Genomics’ LongRanger pipeline followed by additional SV calling with GROC-SV (43), NAIBR (44) and LinkedSV (45) and SvABA (46) and copy number calling with TitanCNA 10x workflow (23). A custom R script was used to find calls made by multiple callers and we manually confirmed all SV calls detected by at least 2 callers and over 10kb in Loupe which are listed with breakpoint information and gene annotation in Additional Table 2 . SVs labeled as “variants of interest” in Additional Table 2 include all SVs occurring in a genes known to be recurrently mutated in G4 medulloblastomas (CDK6, GFI1, GFI1B, MYCN, SNCAIP/PRDM6), those occurring in or near genes known to be recurrently mutated in other cancers types (IDH1 and TERT), those near suspected medulloblastoma genes (procadherin genes), and complex somatic variants such as chromoplexy and chromothripsis). These variants of interest are discussed in more detail within the results section. Only 3 other genes were mutated in more than one patient; these include ARFGEF1 and KB-1047C11.2 which contain breakpoints associated with chromoplexy and chromothripsis on chromosome 8 in MDT-AP-2130 and MDT-AP-3743, respectively, and STEAP2-AS1 which is found near CDK6 and contains breakpoints in both samples with CDK6 amplifications.

WGS data was available through Northcott et al. (12) and processed using the GenPipes Tumor-Pair pipeline for SV calling (-t sv) and SNP calling (-t ensemble) (29). We also ran SvABA on the WGS data (46).

Two tumors and their matching normal samples (MDT‐AP‐1367 and MDT‐AP‐1405) were sequenced on the MinIon (Oxford Nanopore Technologies Limited, Oxford, United Kingdom). MDT-AP-2673 was sequencing on 2 PromethION flowcells (Oxford Nanopore Technologies Limited, Oxford, United Kingdom). Both the MinIon and PromethIon libraries used 2µg of HMW DNA as input. Nanopore data was aligned to genome build b37 with minimap2 (version 2.17) using parameter -ax map-ont (47). Structural variants were called SVIM (48) (parameters –min_mapq 7 –min_sv_size 50 –max_sv_size 100000), NanoVar (49) (version 1.3.9, parameters –data_type ont –mincov 2 –minlen 50), CuteSV (50) (version 1.0.11, parameters –min_size 50 –max_size 100000 –min_support 2 –min_mapq 7 –max_cluster_bias_INS 100 –diff_ratio_merging_INS 0.3 –max_cluster_bias_DEL 100 –diff_ratio_merging_DEL 0.3), and Sniffles2 (51) (version 2.0.6, parameters –minsupport 2 –mapq 7 –minsvlen 50 –non-germline).

PacBio data was available for 5 tumors and their matching normal samples. Samples were normalized to a concentration of 125pM and sequenced with 4-hour movies. Data was aligned to genome build b37 using NGMLR (51) (version 0.2.7) and SVs were called using Sniffles (51) (version 1.0.10, parameters –min_support 2 –min_length 30).

TELL-Seq libraries were generated using the same HMW DNA aliquots as the 10X-LR libraries. 5ng of HMW DNA was used per library (as recommended by the UST TELL-Seq™ WGS Library Prep User Guide V8) and quantified by Qubit™ dsDNA HS Assay Kit (ThermoFisher Scientific, cat# Q32854). Final libraries were assessed by qPCR (KAPA Library Quantification Kits) and Caliper LabChip. Libraries were sequenced using 150PE Illumina reads on 1 lane of NovaSeq SP. QC and barcodes correction was done using the TELL-Read pipeline, and SNP calling was done using the TELL-Sort pipeline. We used the ust10x tools to convert the TELL-Seq data to 10X-LR format and ran our in-house pipeline detailed above with the exception of GROC-SV which did not run to completion on the TELL-Seq data.

Bulk RNA-Seq data was generated for 13 samples and analyzed using the GenPipes RNA-Seq pipeline (29). Overexpression of genes affected by enhancer hijacking was measured using the reads per kilobase of transcript, per million mapped reads (RPKM) calculated by the pipeline.

Hi-C data was available for MDT-AP-3670 (unpublished work). Detailed description of the library preparation protocol and analysis workflow can be found in (42).

A custom R script was used to find SVs detected by multiple technologies (i.e. were both the start and end breakpoints fell within ±1000bp of each other) and we manually assessed all SV calls made by 2 or more callers and larger than 10kb in size using Loupe. All manually confirmed somatic structural variant calls can be found in Additional Table 2 .

Evenness of coverage was compared using BVAtools depthofcoverage (parameters, –gc–minMappingQuality 15 –minBaseQuality 15 –ommitN –maxDepth 1000 –binsize 50000–summaryCoverageThresholds 0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40) and plotted using the karyoploteR package in R. A custom R script was used to compare SVs made by both 10X-LR and TELL-Seq (i.e. were both the start and end breakpoints fell within ±1000bp of each other). All SV calls made by 2 or more callers and large than 10kb in size were manually validated using Loupe. All structural variant calls across both linked-read technologies and all callers can be found in Additional Table 4 .