Medulloblastoma, the most common malignant pediatric brain tumour, arises in the cerebellum, and disseminates through the cerebrospinal fluid (CSF) in the leptomeningeal space to coat the brain and spinal cord 1 . Dissemination, a marker of poor prognosis, is found in up to 40% of children at diagnosis and most children at the time of recurrence. Therefore, affected children are treated with radiation to the entire developing brain and spinal cord followed by high dose chemotherapy with ensuing deleterious effects on the developing nervous system 2 . The mechanisms of CSF dissemination are poorly studied and medulloblastoma metastases have been assumed to be biologically similar to the primary tumour 3,4 . Here we show that in both mouse and human medulloblastoma, multiple metastases from a single patient are extremely similar to each other, but divergent from the matched primary tumour. Clonal genetic events in the metastases can be demonstrated in a restricted sub-clone of the primary tumour, suggesting that only rare cells within the primary tumour have the ability to metastasize. Failure to account for the bicompartmental nature of metastatic medulloblastoma could represent a major barrier to the development of effective targeted therapies. Thirty percent of Ptch +/− mice develop non-disseminated medulloblastoma by eight months of age 5 . Recently, the Sleeping Beauty (SB) transposon system was shown to be an effective tool for functional genomics studies of solid tumour initiation and progression 6,7 . We expressed the SB11-transposase in cerebellar progenitor cells in transgenic mice under the Math1 enhancer/promoter, but did not observe any tumours when bred to transgenic mice with a concatemer of the T2/Onc transposon (Fig. 1, Supplemental Fig. S1, S2) 8 . However, Ptch +/−/Math1-SB11/T2Onc mice showed increased penetrance of medulloblastoma (~100% (271 out of 279) compared to ~40% (54 out of 139) of controls, and decreased latency (8 months to 2.5 months) (Fig. 1, Supplemental Fig. S2). While Ptch +/− medulloblastomas are usually localized, the addition of SB transposition results in metastatic dissemination through the CSF pathways, identical to the pattern seen in human children (Fisher’s exact test, p=1.8e-07, odds ratio=5.2, Supplemental Table S1) (Fig. 1c, d, g, h, Supplemental Figure S2). As neither transposon, nor transposase alone had an effect on tumour incidence, latency, or dissemination, we conclude that SB-induced insertional mutagenesis drives medulloblastoma progression on the Ptch +/− background (Fig. 1i, Supplemental Fig. S2). Humans with germline mutations in TP53 have Li-Fraumeni syndrome and are at increased risk to develop medulloblastoma. While no medulloblastomas were found in Tp53mut (Tp53 +/− or Tp53 −/−) mice, 40% of Tp53mut /Math1-SB11/T2Onc mice developed disseminated medulloblastoma (Fig. 1e–h, j, Supplemental Fig. S2) 9 . Human medulloblastomas with TP53 mutations frequently have large cell/anaplastic histology. Tp53mut /Math1-SB11/T2Onc medulloblastomas exhibit large cells, nuclear atypia, and nuclear molding typical of large cell/anaplastic histology (Fig. 1f). We conclude that SB transposition can drive the initiation and progression of metastatic medulloblastoma on a Tp53mut background. We used linker-mediated PCR and Roche 454 sequencing to identify the site of T2/Onc insertions in Ptch +/−/Math1-SB11/T2Onc, and Tp53mut /Math1-SB11/T2Onc primary medulloblastomas and their matched metastases. Genes that contained insertions statistically more frequently than the background rate were identified as gene-centric commonly inserted sites (gCISes) 10 . We identified 359 gCISes from 139 primary tumours on the Ptch background and 26 gCISes from 36 primary medulloblastomas on the Tp53 background (Supplemental Tables S2–S7, Supplemental Figures S3–S5). A large number of gCISes targeted candidate medulloblastoma oncogenes/tumour suppressor genes (Supplemental Table S8) 11 . Insertions in candidate tumour suppressor genes including EHMT1, CBP, and MXI1 are predicted to be loss of function (Fig. 1k,l,m), while insertions in putative medulloblastoma oncogenes are largely gain of function, as exemplified by MYST3 (Fig. 1n). Many gCISes mapped to regions of amplification, focal hemizygous deletion, and homozygous deletion, which we recently reported in the genome of a large cohort of human medulloblastomas (Supplemental Table S8) 11 . There is a high level of overlap between gCISes and known cancer genes (COSMIC database) (Supplemental Table S9,10), suggesting that many gCISes are bona fide driver genes in medulloblastoma (Fisher’s exact test p=0.0012) 12 . Similarly, many mouse gCIS/ human amplified genes are over-expressed in human Shh medulloblastomas (Supplemental Fig. S6). Conversely, mouse gCISes deleted in human medulloblastomas were frequently expressed at a lower level in human medulloblastomas (Supplemental Fig. S6). Expression of 6/7 gCISes studied by immunohistochemistry on a human medulloblastoma tissue microarray were associated with a significantly worse overall and progression free survival in human medulloblastoma (Supplemental Table 11, Supplemental Figures S7, S8) 13 . We conclude that our SB-driven leptomeningeal disseminated medulloblastoma model resembles the human disease anatomically, pathologically and genetically and thus represents an accurate model of the human disease that can be used to identify candidate driver events and understand the pathogenesis of human medulloblastoma. We compared the gCISes identified from Ptch +/−/Math1-SB11/T2Onc, and Tp53mut /Math1-SB11/T2Onc primary medulloblastomas and matched metastases (Supplemental Table S2). Strikingly, the overlap between primary tumour gCISes (pri-gCISes) from Ptch +/−/Math1-SB11/T2Onc tumours and those from the metastases (met-gCISes) from the same animals was only 9.3% of gCISes (Figure 2a). Similarly, the overlap in pri-gCISes from primary Tp53mut /Math1-SB11/T2Onc gCISes and the matching met-gCISes was only 8.9% (Figure 2b). Leptomeningeal metastases and the matched primary tumour share identical, highly clonal insertion sites (Fig. 2c). The chances of two (or three) unrelated tumours having SB insertions in exactly the same TA dinucleotide are extremely low. We conclude that leptomeningeal metastases and matched primary tumour arise from a common transformed progenitor cell, and have subsequently undergone genetic divergence. Sequencing also identified insertions that are highly clonal in the metastases, but not seen in the matched primary tumour (not shown). Endpoint PCR for these insertions in the matched primary/metastatic tumours show that the insertion is highly clonal in the metastase(s), and present in a very small subclone of the primary tumour (Fig. 2d, Supplemental Figure S9). These data are consistent with a model in which metastatic disease arises from a minor restricted subclone of the primary tumour. Dissemination could occur repeatedly from the same subclone of the primary tumour, which seeds the rest of the CNS, or could occur once followed by reseeding of the rest of the leptomeningeal space by the initial metastasis. Insertions that are restricted to a minor subclone of the primary tumour, but which are clonal in the metastases, could correspond to the ‘metastasis virulence’ genes, described previously, that offer a genetic advantage during dissemination, but not to the primary tumour 14 . Another explanation of our data could be reseeding of the primary tumour by a metastatic clone that had acquired additional genetic events in the periphery. This latter hypothesis is mitigated by the presence in the same animal of highly clonal insertions in the metastasis that are completely absent from the primary tumour 15 . As reseeding should be accompanied by contamination of the primary tumour with events found in the metastases, absence of these events in the matched primary tumour makes reseeding much less likely (Fig. 2e). We hypothesize that events found only in one metastasis represent progression events acquired post-metastasis, and which could lead to localized progression of metastatic disease as is sometimes seen in human children. We observed highly clonal insertions in the primary tumour, including known medulloblastoma oncogenes such as Notch2, or Tert, which are not found in the matching metastases (Fig. 2f). This pattern could be explained through remobilization of the SB transposon in the metastatic tumour; however, no signs of the DNA footprint left after SB remobilization at these loci were observed (Supplemental Fig. S10) 16 . We suggest that these events, which may constitute driver events in the primary tumour, have arisen in the primary tumour after the metastases have disseminated (post-dispersion events). Although these known oncogenes represent attractive targets for therapy, their utility as targets for therapy may be limited if the target is not also found in the leptomeningeal compartment of the disease. Our data from two separate mouse lines supports a model in which medulloblastoma disseminates early from a restricted subclone of the primary tumour, and where the primary tumour and the matched metastases then undergo differential clonal selection and evolution. Failure to account for the differences between the primary and leptomeningeal compartments could lead to the failure of targeted therapies. Failure to study the leptomeningeal disease could result in systematically overlooking critical targets for therapy in this compartment (Fig. 2e). Examination of met-gCIS genes using GSEA demonstrates differences between the primary and metastatic disease, which importantly include enrichment for genes involved in the cytoskeleton among the metastases (Supplemental Table S12). Targets that are present in both compartments, and which are maintenance genes, will be optimal targets for therapy of both the primary and metastatic compartments, as exemplified by Pdgfra (Fig. 2c, Supplemental Tables S7, S9). Pten, Akt2, Igf2, and Pik3r1 are all met-gCISes, implicating the PI3-kinase pathway in medulloblastoma progression. We injected the cerebella of Nestin-TVA mice 17 with either Shh virus alone, or Shh + Akt virus. Cerebellar injection of Shh alone resulted in medulloblastoma in 6/41 animals, compared to 20/42 animals injected with Shh + Akt (p=0.0018). Poignantly, while metastases were never seen with Shh virus alone (0/41), medulloblastoma metastases were seen in 9/42 animals injected with Shh + Akt (p=0.0024) (Supplemental Fig. S11). In vivo modeling validates PI3-kinase signaling and suggests that it can contribute to leptomeningeal dissemination of medulloblastoma. Prior publications and clinical approaches to human medulloblastoma have largely assumed that the primary tumour and its matched metastases are highly similar 3,4 . To test this assertion we formally reviewed all cases of medulloblastoma from the last decade at The Hospital for Sick Children, and identified 19 patients who had both bulk residual primary tumour post-surgery, and MRI visible metastases, both of which could be followed for response to treatment in the two compartments (Supplemental Fig. S12 and Supplemental Table S13). While it is possible that metastases might have received reduced radiotherapy than the primary tumor in a subset of patients, in 58% of overall cases (11/19) we observed a disparate response to therapy between the primary tumor and matched metastases (binomial test, p 90% of the coding bases of the exome defined by the consensus coding sequence (CCDS) project were covered by at least 10 reads. Adaptor sequences and quality trimmed reads were removed by using the Fastx toolkit (http://hannonlab.cshl.edu/fastx_toolkit/) and then a custom script was used to ensure that only read pairs with both mates present were subsequently used. Reads were aligned to hg19 with BWA1, and duplicate reads were marked using Picard (http://picard.sourceforge.net/) and excluded from downstream analyses. Single nucleotide variants (SNVs) and short insertions and deletions (indels) were called using samtools (http://samtools.sourceforge.net/) pileup and varFilter2 with the base alignment quality (BAQ) adjustment disabled, and quality filtered to require at least 20% of reads supporting the variant call. Variants were annotated using both Annovar3 and custom scripts to identify whether they affected protein coding sequence, and whether they had previously been seen in dbSNP131, the 1000 genomes pilot release (Nov. 2010), or in approximately 160 exomes previously sequenced at our center. SNV analysis of whole Exome sequencing data For clustering analysis, a SNV frequency matrix was constructed by calculating frequencies from the read counts of the reference and the alternative nucleotide. The matrix was not standardized (i.e. converted to z-scores) prior to clustering, since the absolute SNV frequencies were of interest. For Venn analysis, the samples were grouped into primary-metastasis sets, and the filtered SNVs were used to identify SNVs that are enriched in one sample compared to all other samples of the same set, as determined by the hypergeometric test (p-value threshold = 0.05). For sets consisting of three or more samples (A, B, and C), a SNV was considered to be enriched in samples A and B if the SNV was enriched in A as compared to C alone and also enriched in B as compared to C alone. SNVs that were not enriched in any sample or subsets of samples were considered to be common SNVs. Many of these common SNVs likely represented germline SNVs specific to the patient. Analysis of CpG promoter methylation data The similarities between patient-matched metastatic and primary tumour samples and between patient-matched metastatic tumour samples were determined by using Pearson correlation analysis. As Pearson’s r values are not normally distributed, they were standardized by Fisher’s z transformation. Subsequently, the correlations between metastatic samples and the matched primary samples were compared to the correlations between the patient-matched metastatic samples, using the paired heteroscedastic Student’s t test. Clustering analysis was performed as described above. The methylation matrix was not standardized prior to clustering, as doing so would discard critical information regarding the differences in overall methylation profiles among samples or average methylation among CpG promoters. The stability of the CpG hypermethylation profile clusters was assessed using three methods. First, the clustering analysis was run for different numbers of CpG hypermethylation sites that vary most widely among samples. The partitions generated by each clustering run is compared to the reference partitions generated by original clustering based on the 1000 most variable hypermethylated CpG islands using the Jaccard similarity index. The same analysis was applied to a set of 100 background hypermethylation data matrices in which the sites are permuted within and independently for each sample. Second, the clustering analysis was performed for random sub-samples of 1000 sites, for 1000 repeat runs. In each run, the resulting cluster was compared to the original cluster using the Jaccard index. Analysis on the original data matrix was compared to a set of 100 background matrices, permuted as described above. Third, the cluster stability was further assessed by bootstrap re-sampling of the samples using the Pvclust R package (v1.2). Supplementary Material 1 2
