Recurrent somatic alterations of FGFR1 and NTRK2 in pilocytic astrocytoma

Abstract

Pilocytic astrocytoma, the most common childhood brain tumor¹, is typically associated with mitogen-activated protein kinase (MAPK) pathway alterations². Surgically inaccessible midline tumors are therapeutically challenging, showing sustained tendency for progression³ and often becoming a chronic disease with substantial morbidities⁴.

Here we describe whole-genome sequencing of 96 pilocytic astrocytomas, with matched RNA sequencing (n=73), conducted by the International Cancer Genome Consortium (ICGC) PedBrain Tumor Project. We identified recurrent activating mutations in FGFR1 and PTPN11 and novel NTRK2 fusion genes in non-cerebellar tumors. New BRAF activating changes were also observed. MAPK pathway alterations affected 100% of tumors analyzed, with no other significant mutations, indicating pilocytic astrocytoma as predominantly a single-pathway disease.

Notably, we identified the same FGFR1 mutations in a subset of H3F3A-mutated pediatric glioblastoma with additional alterations in NF1⁵. Our findings thus identify new potential therapeutic targets in distinct subsets of pilocytic astrocytoma and childhood glioblastoma.

Pilocytic astrocytoma is the most common central nervous system (CNS) neoplasm in childhood, accounting for ~20% of all pediatric brain tumors. Tumor locations in our cohort reflect the fact that pilocytic astrocytomas occur throughout the CNS, with about half arising outside the cerebellum (Supplementary Figure 1). Extra-cerebellar tumors are often surgically inaccessible, leading to a chronic disease with multiple recurrences, visual and neurological impairment, and/or side-effects of therapy^1,4. Genetic alterations within the MAPK signaling pathway are a hallmark of this tumor, with KIAA1549:BRAF fusion being the most frequent event^6–8. A smaller number harbor BRAF or KRAS point mutations, alternative BRAF/RAF1 fusions, or germline NF1 mutations². Pilocytic astrocytoma has therefore been hypothesized to represent a single-pathway disease². Previously, however, no MAPK pathway changes were identifiable in 15–20% of tumors (mostly non-cerebellar).

To investigate the full range of genetic alterations in pilocytic astrocytoma, we performed whole-genome sequencing of tumor and blood DNA from 96 patients (Supplementary Table 1). Corresponding RNA sequencing data and larger insert mate-pair sequencing (for enhanced structural rearrangement detection) were generated for 73 and 68 samples, respectively. The average somatic mutation rate was extremely low (<0.1/Mb), with a mean of 1.6 non-synonymous single nucleotide variants (SNVs) per tumor (range 0–9; Supplementary Table 1) – similar to the rate described in NF1-associated pilocytic astrocytomas⁹. This is markedly lower than recently reported for the malignant pediatric brain tumor medulloblastoma^10–12, and several other pediatric solid tumors¹³. The average number of small insertions/deletions (InDels) affecting coding sequences was <1 per case. All coding somatic SNVs/InDels are listed in Supplementary Table 2.

In line with other tumor types^10,14,15, genome-wide mutation rates positively correlated with patient age (r = 0.42, P = 2.3 × 10⁻⁵, Pearson’s product-moment correlation; Supplementary Figure 2a). The observed mutations were predominantly C-to-T transitions at CpG sites (likely deamination of methylated cytosines), suggesting that the age-dependent increase may largely be due to background processes occurring in progenitor cells prior to tumorigenesis, as recently reported in leukemia¹⁵ (Supplementary Figure 2b).

Most of the known MAPK pathway activating events were also found in this series, including KIAA1549:BRAF fusion variants (70 cases), a FAM131B:BRAF fusion¹⁶, four BRAF^V600E mutations and one BRAF^ins599T (Supplementary Table 1). Three tumors were associated with Neurofibromatosis Type 1. This is fewer than would be expected for prospective cohorts (5–10%), since material for biological studies from these typically optic pathway-associated tumors is limited. NF1 has been reported to follow a classical tumor suppressor model in pilocytic astrocytoma, with a somatic second hit in addition to the germline alteration⁹. This also held true in our series (Supplementary Table 1).

Analysis of copy number and structural alterations using DNA and RNA sequencing revealed four novel BRAF fusions (Figure 1, Supplementary Figure 3). As expected, all variants resulted in loss of the amino-terminal regulatory region of BRAF. An RNF130:BRAF fusion derived from a reciprocal t(5;7)(q35;q34) translocation was seen in two cases (Figure 1a), with single examples of CLCN6:BRAF, MKRN1:BRAF and GNAI1:BRAF fusions (Supplementary Figure 3a–c). Thus, non-KIAA1549:BRAF fusions comprise a significant minority of activating events, with BRAF apparently being a promiscuous fusion partner.

Figure 1. Novel BRAF alterations in pilocytic astrocytoma.

a, Schematic representation of the RNF130:BRAF fusion gene in ICGC_PA112 resulting from a translocation between chromosomes 5 and 7. A similar fusion was observed in ICGC_PA96. The cDNA sequence at the fusion breakpoint and resulting exon and protein structures are indicated. A reciprocal fusion between RUN and FYVE domain containing 1 (RUFY1) and transmembrane protein 178B (TMEM178B) on the derivative chromosome 5 in ICGC_PA112 was also found to be expressed by RNA-seq analysis (not shown). RPM, reads per million; KD, kinase domain.

b, Computational modeling of two BRAF monomers (light/dark grey) with a VLR insertion (blue/magenta) between Arg506 and Lys507 (green), as identified in ICGC_PA65. The PDB structure 4E26 was used as a template. Val600, a mutational hotspot, is shown in yellow. A novel dimer interface is formed between the protomers, with hydrogen bonds from the new arginine side chains (dashed lines) and a hydrophobic interaction between the leucine side chains (magenta).

c, Western blot analysis of NIH3T3 cells transfected with empty vector (EV), BRAF^WT, BRAF^V600E and BRAF^insVLR. The newly identified insVLR mutant results in elevation of ERK1/2 phosphorylation (pERK1/2) at a similar level to that seen with the known oncogenic V600E form.

d, Co-immunoprecipation assay with pull-down of HA-tagged BRAF^insVLR, demonstrating that this novel mutant forms homodimers with co-transfected AU1-tagged insVLR, but does not appear to form a strong heterodimer with wild-type BRAF.

Another new BRAF alteration was identified in ICGC_PA65, resulting in a three amino acid insertion (p.R506_insVLR) in the interdomain cleft of BRAF - a structural region linked to its activity¹⁷ and homodimerization¹⁸. Protein modeling predicted that these residues stabilize a dimeric form of BRAF (known to be active independent of RAS stimulation¹⁹) (Figure 1b). Homodimerization was confirmed by immunoprecipitation, and the BRAF^insVLR mutant increased ERK phosphorylation as effectively as BRAF^V600E (Figure 1c,d).

Novel alterations in KRAS were also observed. ICGC_PA117 and ICGC_PA142 both showed two distinct mutations (E63K/R73M and L19F/Q22K, respectively). DNA and RNA-seq data confirmed that both alterations affected the same allele (Supplementary Figure 4). Whilst there are reports of double KRAS mutations in entities such as colon cancer²⁰, these typically involve at least one hotspot residue (codons 12, 13 or 61), and often represent heterogeneous tumor subclones rather than two hits in one allele (although this has also been described, e.g. ²¹). The alterations seen in our tumors do not encompass classical hotspots, suggesting that further characterization of downstream effects is warranted.

All but one of the cerebellar tumors in our series harbored a BRAF fusion, with the other displaying a KRAS alteration. 9/48 (19%) of the non-cerebellar tumors, however, lacked any of the above alterations. Further assessment of structural rearrangements revealed two novel gene fusions in a total of 3 samples, involving the kinase domain of NTRK2 (also known as TrkB) - an oncogene implicated in the tumorigenesis of neuroblastoma, amongst others^22,23. The related NTRK1 and NTRK3 genes have previously been shown to be activated by fusion events (e.g. TPM3:NTRK1 in papillary thyroid cancer²⁴ and ETV6:NTRK3 in multiple tumors²⁵). The QKI:NTRK2 and NACC2:NTRK2 fusions identified here were verified by PCR (Figure 2, Supplementary Figure 3d). Both 5′ partners contain dimerization domains, and are therefore predicted to induce ligand-independent dimerization. Interestingly, N-terminal TrkB truncation has recently been shown to induce transformation of neural crest cells²⁶. The downstream effects of TrkB activation are mediated at least in part via MAPK pathway activation²⁷.

Figure 2. NTRK2 is a new gene fusion target in pilocytic astrocytoma.

Schematic representation of the QKI:NTRK2 fusion gene in ICGC_PA159 resulting from a translocation between chromosomes 6 and 9. A similar fusion was observed in ICGC_PA82. The cDNA sequence at the fusion breakpoint and resulting exon and protein structures are indicated. RPM, reads per million; KD, kinase domain; QUA1, Qua1 dimerization domain.

A second new recurrent alteration, namely mutation of two hotspot residues (N546 & K656) in the kinase domain of FGFR1, was seen in five tumors (Supplementary Table 3 and Figure 3a). FGFR1 is more commonly activated through amplification, in tumors such as breast²⁸ and lung^29,30 cancer. Occasional FGFR1 mutations have been observed in adult glioblastoma (GBM)^31,32, a highly malignant astrocytoma, as have FGFR1:TACC1 or FGFR3:TACC3 fusion genes³³. Mutations in homologous codons in FGFR2 and FGFR3 are commonly found in other tumor types, particularly bladder, skin and endometrial cancers (see the COSMIC database³⁴). Both mutations result in midbrain hyperproliferation in developing chick embryos³⁵. The N546K mutation alters FGFR1 auto-phosphorylation, resulting in increased kinase activity and transforming potential³⁶, while the K656E variant is also transforming in vitro³⁷. Interestingly, the latter study suggested that FGF2 (bFGF) ligand was necessary in addition to FGFR1 mutation to maintain neurosphere formation in vitro. Gene expression array data of 118 pilocytic astrocytomas, including 66 from the present series, revealed significantly increased FGF2 expression in pilocytic astrocytoma compared with 158 other astrocytic tumors^38,39, or normal tissues⁴⁰. This increase was not restricted to only FGFR1-mutant or wild-type tumors, suggesting that ligand-mediated pathway activation may play a general role in tumorigenesis (Figure 3b). Immunohistochemical detection of phospho-FGFR1 revealed strong, diffuse positivity in all seven pilocytic astrocytomas harboring an FGFR1 mutation for which material was available. No positivity was observed in eleven FGFR1^WT tumors. All samples displayed strong phospho-ERK staining (Supplementary Figure 5). Interestingly, ICGC_PA89 harbored an alternative alteration in FGFR1, consisting of a ~4.5kb internal tandem duplication (ITD) of the portion of the gene encoding the kinase domain, reminiscent of the activating internal tandem duplications of the FLT3 kinase observed in acute myeloid leukemia⁴¹ (Figure 3c).

Figure 3. FGF pathway signaling molecules are recurrently altered in pilocytic astrocytoma.

a, Schematic of the domain structure of FGFR1, indicating the position and frequency of the hotspot mutations in pilocytic astrocytomas sequenced in the present study (including replication cases).

b, Gene expression data for FGF2 indicating significantly elevated expression in pilocytic astrocytomas (red) compared with other astrocytic tumors (blue), normal cerebellum (black) or other normal tissues (green); P <0.001, two-sided t-test. The pilocytic astrocytomas with expression data which harbor FGFR1 alterations (four mutants plus FGFR1-ITD) are highlighted.

c, Schematic of an additional alteration in FGFR1 identified in ICGC_PA89, comprising an internal tandem duplication (ITD) of part of intron 10, exons 11–17 and part of exon 18. The duplicated amino acids are aa478–820 (numbered according to the alpha A1 isoform), with an additional 40 amino acid linker sequence encoded by part of intron 10. The whole kinase domain is therefore duplicated in the resulting predicted protein (TK1′ and TK2′).

d, Schematic of the structure of PTPN11 (SHP-2), indicating the position and frequency of mutations in pilocytic astrocytomas sequenced in the present study.

e, Gene expression data for PTPN11 indicating significantly elevated expression in pilocytic astrocytomas compared with other groups, as in (b); P <0.001, two-sided t-test. The pilocytic astrocytomas with expression data which harbor FGFR1 alterations (four mutants plus FGFR1-ITD) are highlighted.

Ig, immunoglobulin-like domain; TM, transmembrane domain; TK, tyrosine kinase domain; SH2, src homology 2 domain; PTP, protein tyrosine phosphatase domain; DA, WHO Grade II diffuse astrocytoma; AA, anaplastic astrocytoma; K27/G34/IDH1 GBM, glioblastoma carrying a mutation at K27 or G34 in H3F3A or in IDH1; ped., pediatric; CBM, cerebellum, CNS, central nervous system.

Further recurrent alterations were found in a RAS/MAPK-related adaptor protein, the phosphatase PTPN11 (also called SHP-2, Figure 3c). Both alterations (E69K and E76A) were previously reported in juvenile monomyelocytic leukemia, which is frequently associated with SHP-2 activation^42,43. Interestingly, both were found in FGFR1-mutant tumors (ICGC_PA84 & ICGC_PA166), suggesting a cooperative role in tumorigenesis (Supplementary Table 3). Overexpression of mutant SHP-2 alone did not elevate pERK levels in vitro, while the two FGFR1 mutants, either alone or in combination with mutant SHP-2, upregulated pERK (Supplementary Figure 6). This would support the hypothesis that PTPN11 mutation alone is insufficient for pilocytic astrocytoma development, but that it may play a modifying role in FGFR1-mutant tumors. Of note, PTPN11/SHP-2 expression was increased in pilocytic astrocytomas when compared with other astrocytomas or normal tissues (Figure 3d), suggesting a broader role in the biology of this entity. An additional cohort of 45 non-cerebellar pilocytic astrocytomas, negative for KIAA1549:BRAF, was screened for FGFR1 (exons 12 & 14) and PTPN11 (exon 3) mutations. Nine cases harbored FGFR1 mutations at N546 or K656, and one additionally carried a PTPN11/SHP-2 E69K change (Supplementary Table 3), confirming our whole-genome sequencing findings. Germline PTPN11 mutations are one of the causes of the hereditary developmental disorders Noonan syndrome (NS)⁴⁴ and LEOPARD syndrome (LS)⁴⁵. A few case reports have described pilocytic astrocytomas occurring in NS and LS patients^46–49. Together with NF1, this makes three known ‘RASopathies’, characterized by germline MAPK pathway mutations⁵⁰, linked with pilocytic astrocytoma tumorigenesis. In our germline sequencing data, however, NF1 was the only RASopathy gene disrupted at a higher frequency than in the 1000 Genomes Project (URL below).

Strikingly, all of the pilocytic astrocytomas in our cohort harbored a MAPK pathway alteration. BRAF, FGFR1, KRAS and NF1 were the only genes found to be significantly mutated using the Genome MuSiC algorithm (URL below, Supplementary Table 4). With the exception of FGFR1 and PTPN11, each case typically harbored only one pathway alteration (P < 0.0001, permutation test, Figure 4). Together with the finding that BRAF kinase activation alone is sufficient to induce pilocytic astrocytomas in mice^51,52, these data strongly support the concept of pilocytic astrocytoma as a prototypic single-pathway disease driven by a limited number of oncogenic hits (possibly only one in some cases; Supplementary Figure 7).

Figure 4. Summary of MAPK pathway alterations in pilocytic astrocytoma.

An overview of MAPK pathway alterations identified in the 96 whole-genome sequencing cases included in the present study, indicating the mutual exclusivity of the majority of these hits (with the exception of FGFR1 and PTPN11); P <0.0001, permutation test on 10,000 iterations. Each column represents one tumor sample. Red filled boxes indicate that a given alteration is present in this sample. The blue filled box represents FGFR1-ITD rather than a point mutation; the green box indicates a BRAF^E451D mutation in a case with a KIAA1549:BRAF fusion (unknown functional significance, but included in the exclusivity testing); the black/red boxes indicate that the NF1 alterations comprise one germline and one somatic hit per case.

Interestingly, one of the FGFR1-mutant tumors (ICGC_PA69) also displayed an H3F3A K27M mutation and somatic alterations of NF1 – both of which are more commonly encountered in pediatric glioblastoma (GBM)⁵. Three experienced neuropathologists (AK, GR and JF) reported pilocytic astrocytoma histology for this case, although a diagnosis of GBM cannot be conclusively excluded due to limited material. By examining previous exome sequencing data for pediatric GBM⁵, we identified 3/48 samples (6%) with an FGFR1 mutation. Strikingly, all three harbored the same constellation of H3F3A K27M, somatic NF1 alteration, and FGFR1 activation (Supplementary Table 3). They were also wild-type for TP53, which is mutated in most K27M-mutant GBM or diffuse intrinsic pontine gliomas^5,53. One tumor reported in a targeted sequencing cohort of medulloblastoma¹⁰ had a similar triplet, with H3F3A K27M, NF1 mutation and an FGFR2 K659E mutation (homologous to FGFR1 K656E), making a total of 5 cases with this triple-mutation combination. Gene expression analysis indicated that this is likely a GBM previously mis-classified as medulloblastoma. It is not currently clear why these mutations occur in concert, and additional work will be required to assess their roles. One possibility is that NF1 mutation may mimic elevated PTPN11 expression, since activation of PTPN11/SHP-2 inhibits the recruitment of Ras GTPase-activating proteins (RasGAPs, including NF1) to the plasma membrane⁵⁴.

All FGFR1-mutant tumors were extra-cerebellar, mostly in midline locations (Supplementary Table 3), suggesting a link between cell of origin and/or microenvironment with FGFR1-driven tumorigenesis. The H3F3A K27M mutation is also associated with midline GBM³⁹. Notably, FGFR1 plays a role in neural stem cell self-renewal⁵⁵, and is essential for midline glial cell development⁵⁶. This spatial clustering may also reflect differential sensitivity of distinct neural precursors to activating stimuli, particularly NF1 loss^57,58. The type and timing of second hits (H3F3A/NF1 mutation) and/or the differentiation stage of the cell of origin may contribute to determining a fate of oncogene-induced senescence and slow growth (pilocytic astrocytoma)^59,60 versus aggressive malignancy with poor outcome (GBM).

In summary, this study has identified new insights into the tumorigenesis of pilocytic astrocytoma. Each tumor harbored very few mutations, in keeping with a generally benign behavior. Importantly, our findings confirm the concept that pilocytic astrocytomas are predominantly driven by aberrant activation of the MAPK pathway. Most strikingly, however, we report novel recurrent alterations in NTRK2, FGFR1 and PTPN11, which were mutually exclusive with other RAF/RAS alterations. Combined with the observation of FGF2 and PTPN11 overexpression, these results indicate upstream contributors to MAPK pathway activation in this entity. Emerging preclinical data suggest that BRAF inhibitors may trigger paradoxical activation in tumors harboring KIAA1549:BRAF, i.e. the majority of pilocytic astrocytomas⁶¹. Single-drug or combination therapy with FGFR, NTRK2 and/or MEK inhibitors, several of which are currently in preclinical and clinical trials^62–64, may therefore represent rational treatment options. BRAF^V600E-specific agents may also be a logical choice for ~5% of patients. Finally, the identification of recurrent FGFR1 mutations in a subset of pediatric GBM provide an opportunity for therapeutic targeting of FGFR signaling in these clinically challenging brain tumors.

Methods

Sample collection

Informed consent and an ethical vote (Ethics Committee of the Medical Faculty of Heidelberg) were obtained according to ICGC guidelines (http://www.icgc.org). Tumor tissues were subjected to neuropathological review to confirm histology and tumor cell content.

DNA Sequencing

Paired-end library preparation was conducted using Illumina, Inc. v2 protocols. 1–5μg of genomic DNA were fragmented to ~300 bp insert-size with a Covaris device, followed by size selection through agarose gel excision. Mate-pair (long-range paired-end mapping) DNA library preparation was carried out using the v2 protocol from Illumina. 10ug of genomic DNA were fragmented to 4.5kb with a Hydroshear device, followed by size selection. Deep sequencing was carried out with Illumina HiSeq2000 instruments.

RNA Sequencing

Twenty-three RNAseq libraries were prepared with purified polyA+ RNA fractions using strand-specific methods, following dUTP-based protocols as described in⁶⁵, featuring cDNA fragmentation after mRNA priming with random hexamers (dN)6 and oligo(dT) primers. Six libraries were constructed with a modified protocol whereby the polyA+ fraction was fragmented using RNA fragmentation reagents (Ambion, Cat. #AM8740); first strand synthesis was then performed with random hexamers only (cDNA fragmentation was omitted). Fifty RNAseq libraries were prepared with a ribosomal RNA-depleted fraction. In brief, 0.2 μg of total RNA was prepared using the RiboZero^™ Gold kit (Epicentre). The resulting RNA was further processed following the library preparation protocol described in ⁶⁶, starting at the fragmentation step (2^nd step). Sequencing (2×51bp) was carried out on the HiSeq2000.

Mapping and analysis

Sequencing reads were mapped and aligned to the hg19 reference assembly as previously described¹⁰, using BWA⁶⁷ (version 0.5), and processed with samtools⁶⁸ (version 0.1.17) and Picard tools (version 1.61).

An in-house analysis pipeline based on samtools mpileup and bcftools⁶⁸ was used to detect single nucleotide variants (SNV) and small insertions or deletions (InDels). In addition to previously described filters to remove artefacts¹⁰, we excluded variants located in regions of low mappability or overlapping with the hiSeqDepthTop10Pct, Encode DAC Blacklisted Regions, or Duke Excluded Regions tracks from UCSC genome browser. High confidence somatic SNVs were not allowed to overlap with any two of the following features: tandem repeats, simple repeats, low complexity, satellite repeats, or segmental duplications. In addition, following heuristic criteria were applied: (1) at least 5 tumor reads at the position; (2) more than one variant read per strand, or at least 5 variant reads in total and variant allele fraction > 0.1; (3) at least 12 reads at the position in the matching control; (4) less than 1/30 of the control reads supporting the variant; (5) less than 300 reads at the corresponding position in the control; (6) no non-reference, non-variant bases at the corresponding position in the control.

InDels were called with samtools mpileup and bcftools on reads with mapping quality >20 and scored in a similar way as SNVs. Overlap with tandem or simple repeats, however, was not penalized, since these elements are prone to InDels due to polymerase slippage. Since InDel alignments in the matched control can be slightly shifted in comparison to the tumor, or mismatches preferred over gaps, germline events can be falsely called as somatic. We therefore required not more than one mismatch or InDel in the control within 20bp of the tumor InDel.

Tumor and matched control samples were also analyzed with Pindel (version 0.2.4h)⁶⁹. Events in the tumor were only considered when supported by at least five reads and if the number of supporting reads divided by the maximum of the read depth at the left and right breakpoints was >0.05. The matched control sample was also analyzed by samtools mpileup at tumor InDel positions and 10bp up- and downstream. Variants were classified as somatic if both Pindel and samtools mpileup did not call a multibase variant in the control sample. Only additional InDels in “RASopathy” genes were reported from the Pindel analysis (due to a high false positive rate), all other InDels were called with samtools as described above.

SNVs and InDels were functionally annotated with RefSeq gene annotations using Annovar⁷⁰ and Oncotator (http://www.broadinstitute.org/oncotator/). For the identification of significantly mutated genes we used the high confidence somatic SNVs and InDels as input for Genome MuSiC⁷¹ (version 0.3), setting max-fdr to 0.05 in the genome music smg module. Substitution patterns of SNVs were evaluated in a sequence context of all 96 possible trinucleotides to assess mutational signatures⁷².

Integration of SNVs and InDels with RNAseq data

Gene expression levels were calculated per-exon according to reads per kilobase of exon model per million mapped reads (RPKM) using BedTools⁷³ and custom Perl scripts. Where available, candidate DNA variant positions were annotated with RNA information by generating a pileup of the DNA variant position in the RNA BAM file (Supplementary Table 2).

Structural rearrangement detection & verification

Rearrangements based on paired-end data were identified using read-depth analysis, CREST⁷⁴, DELLY⁷⁵, and manual inspection of sequencing reads. Rearrangements based on mate-pair data were identified using DELLY and manual inspection of sequencing reads, as previously described¹⁰.

Structural variants were verified by PCR (conditions and primer sequences available upon request). PCR products were sent to GATC Biotech (Germany) for Sanger sequencing to confirm breakpoints.

Fusion transcript detection & verification

Fastq files from transcriptome sequencing were used for de novo annotation of fusion transcripts using the TopHat-Fusion⁷⁶ and deFuse⁷⁷ algorithms with standard parameters. Where neither algorithm detected fusions, but whole-genome sequencing supported a fusion, we extracted corresponding transcriptome reads matching the theoretical sequence surrounding the predicted fusion border and then counted as fusion reads those where the entire 51bp sequences were derived from the predicted fused exons.

Primers for amplification of neighboring exons in normal (unfused) transcripts were tested in RT-PCR using total RNA from normal cerebellum (BioChain, lot number B307003). Validated primers were used to amplify the normal transcripts (control) and fusion transcripts from tumor RNA using the SuperScript III One-Step RT-PCR System with Platinum Taq DNA Polymerase (Invitrogen). PCR conditions and primer sequences are available upon request.

Computational protein modeling

A dimeric model of mutant BRAF^insVLR was produced with Modeller⁷⁸ using the PDB structure 4E26 as a template. Ten models were produced, with the one having the lowest DOPE score shown in main Figure 1b.

Expression Array Analysis

Affymetrix U133 Plus2.0 expression array data for genes of interest were extracted from publicly-available datasets via the R2 software tool (http://r2.amc.nl), and for additional cases on an early-access basis through collaboration with the Microarray Department of the University of Amsterdam, the Netherlands. The MAS5.0 algorithm of the GCOS program (Affymetrix Inc) was used for normalization and assignment of detection p-values. Array quality was ensured by inspection of beta-actin and GAPDH 5′-3′ ratios as well as the percentage of present calls.

Verification of SNVs/InDels

All SNVs/InDels in FGFR1, PTPN11, BRAF, KRAS, NF1 and H3F3A were verified by PCR followed by capillary (Sanger) sequencing. Additional variants were also verified in this way, as listed in Supplementary Table 2. Verification rates for SNVs was >98%, and for InDels was >70%. Alterations determined to be false are not included in Supplementary Table 2.

In vitro and protein assays

Coding sequences of BRAF, PTPN11 and FGFR1 were cloned from normal brain cDNA (Stratagene) into a pcDNA3.1 vector with HA, FLAG or AU1 epitope tags. Site-directed mutagenesis (QuikChangeII XL, Agilent Technologies) was used to generate BRAF^V600E, BRAF^insVLR, PTPN11^E69K, PTPN11^E76A, FGFR1^N546K and FGFR1^K656E.

NIH3T3 murine fibroblasts (Leibniz Institute DSMZ, Germany; mycoplasma tested) were cultured in DMEM (Life Technologies), 10% FBS (Life Technologies) and penicillin/streptomycin at 37°C in 5% CO₂. Cells were transfected using Lipofectamine® 2000 diluted in Opti-MEM® (Invitrogen), and switched to serum-free DMEM 24h after transfection. After a further 24h, cells were lysed in either RIPA buffer or RLT buffer (Qiagen). Protein electrophoresis was performed using 4–12% gradient NuPAGE Bis-Tris Precast Gels (Life Technologies) with transfer to a PVDF membrane. Antibodies for Western blotting, with detection using ECL (GE Healthcare Life Sciences), were as follows: anti-ERK1/2 (rabbit polyclonal, Cell Signaling Technology #9102), anti-phospho-ERK1/2 (rabbit monoclonal 20G11, Cell Signaling Technology #4376), anti-HA (rabbit monoclonal C29F4, Cell Signaling Technology #3724), anti-AU1 (rabbit polyclonal, Abcam ab3401), anti-DYKDDDDK (FLAG®) (rabbit polyclonal, Cell Signaling Technology #2368), HRP-conjugated goat anti-Rabbit IgG (Santa Cruz Biotechnology, sc-2004), HRP-conjugated goat anti-mouse Ig (Santa Cruz Biotechnology, sc-2005).

For co-immunoprecipitation (Co-IP) to assess BRAF dimerization, cells were washed twice in ice-cold PBS, scraped, pelleted, and then lysed on ice in 5 pellet-volumes of lysis buffer (50mM HEPES, pH7.5; 250mM NaCl; 1mM EDTA; 2,5 mM EGTA; 10% Glycerol; 0.1% Triton X-100; protease inhibitors) with regular vortexing. Lysates were run through a QiaShredder column (Qiagen), centrifuged, and transferred to a new tube. 8μl of anti-HA agarose slurry (Thermo Scientific) were washed 3 times then resuspended in 20μl lysis buffer. 50μg protein extract was added to the anti-HA slurry and rotated overnight at 4°C. Finally, beads were pelleted, washed seven times in lysis buffer, and resuspended in SDS sample buffer for Western blotting.

Immunohistochemistry (IHC)

IHC was performed with an automated stainer (Benchmark XT; Ventana). Phospho-FGFR (Y653/Y654) was assessed using antibody PA5-12594 (Thermo Scientific), diluted 1:50. Phospho-ERK1/2 (T202/Y204) was assessed using antibody #9101 (Cell Signaling Technology), diluted 1:100.