Accurate calling of KIAA1549‐BRAF fusions from DNA of human brain tumours using methylation array‐based copy number and gene panel sequencing data

KIAA1549‐BRAF fusions occur in certain brain tumours and provide druggable targets due to a constitutive activation of the MAP‐kinase pathway. We introduce workflows for calling the KIAA1549‐BRAF fusion from DNA methylation array‐derived copy number as well as DNA panel sequencing data.


INTRODUC TI ON
A fusion of the genes KIAA1549 and BRAF, generated through a focal tandem duplication on 7q34, causes constitutive activation of the mitogen-activated protein kinase (MAPK) pathway [1]. It initiates cell growth, migration, differentiation and survival, and plays a crucial role in tumour development [2]. Hence, MEK inhibitors may be considered as a therapeutic option in the presence of a KIAA1549-BRAF fusion [3,4]. The KIAA1549-BRAF fusion is the most common genetic alteration in pilocytic astrocytoma [1], but it also occurs in other central nervous system (CNS) tumours such as diffuse leptomeningeal glioneuronal tumour (DLGNT) [5,6] and high-grade astrocytoma with piloid features [7].
However, these techniques may miss some variants [12]. While RNA sequencing is currently becoming the gold standard for detection of gene fusions and is increasingly used in diagnostic settings [13][14][15], it is still not universally available and may be challenging from formalin-fixed paraffin-embedded samples (FFPE).
DNA methylation profiling is becoming more widely used in routine diagnostics of brain tumours [16]. Copy number profiles calculated from DNA methylation data may be used for the detection of a focal copy number gain on 7q34, resulting from the tandem duplication generating the KIAA1549-BRAF fusion. Visual calling of this gain has been done in some studies [6,7,16]. However, so far it has not been validated how reliably this gain can be detected, and how closely such an event called visually correlates with the presence of this fusion.
Furthermore, prior to a visual inspection, an automated algorithm to detect focal gains suggesting a KIAA1549-BRAF fusion might be helpful for a pre-evaluation for diagnostic purposes, or when screening larger cohorts as recently exemplified for YAP1 fusions [17].
In addition to DNA methylation profiling, DNA panel sequencing from FFPE tissue is also becoming more widely available in routine diagnostics. If intronic sequences of potential fusion partners such as KIAA1549 and BRAF are covered in a gene panel, it may be used for gene fusion detection. However, few data regarding how reliably the KIAA1549-BRAF fusion can be detected from DNA panel data and which algorithms are suitable for the analysis are available.

Data generation
DNA methylation analysis DNA was processed using the Illumina HumanMethylation450 or EPIC BeadChip array as previously described [18]. The data were analysed with the DNA methylation-based brain tumour classifier [18]. Samples were regarded classifiable to a DNA methylation class if the calibrated classifier score was ≥0.9. A detailed description of Methods: Copy number profiles were analysed by automated screening and visual verification of a tandem duplication on chromosome 7q34, indicative of the KIAA1549-BRAF fusion. Pilocytic astrocytomas of the ICGC cohort with known fusion status were used for validation. KIAA1549-BRAF fusions were called from DNA panel sequencing data using the fusion callers Manta, Arriba with modified filtering criteria and deFuse. We screened DNA methylation and panel sequencing data of 7790 specimens from brain tumour and sarcoma entities.

Results:
We identified the fusion in 337 brain tumours with both DNA methylation and panel sequencing data. Among these, we detected the fusion from copy number data in 84% and from DNA panel sequencing data in more than 90% using Arriba with modified filters. While in 74% the KIAA1549-BRAF fusion was detected from both methylation array-derived copy number and panel sequencing data, in 9% it was detected from copy number data only and in 16% from panel data only. The fusion was almost exclusively found in pilocytic astrocytomas, diffuse leptomeningeal glioneuronal tumours and highgrade astrocytomas with piloid features.

Conclusions:
The KIAA1549-BRAF fusion can be reliably detected from either DNA methylation array or DNA panel data. The use of both methods is recommended for the most sensitive detection of this diagnostically and therapeutically important marker.

K E Y W O R D S
Arriba, DNA methylation, DNA panel sequencing, gene fusion, KIAA1549-BRAF, pilocytic astrocytoma the methylation classes is outlined under https://www.molec ularn europ athol ogy.org. The same analysis was done using the sarcoma classifier (https://www.molec ulars arcom apath ology.org; Koelsche C et al., accepted for publication).

DNA panel sequencing
Gene panel sequencing from FFPE samples was performed and data were processed as previously described. [19] The applied brain tumour gene panel covers intronic regions of the genes KIAA1549 and BRAF (see Table S1 for details on all covered regions on chromosome 7).
RNA sequencing RNA sequencing of formalin-fixed paraffin-embedded tissue was performed as previously described [14]. These data were used as additional validation. 16 -BRAF ex. 9 and KIAA1549 ex. 16 -BRAF ex. 11 were then done as described previously [9].

Calling of the KIAA1549-BRAF fusion
Calling of the KIAA1549-BRAF fusion by visual and automated inspection of copy number profiles calculated from DNA methylation array data Copy number profiles were computed from the methylation data using the R package conumee [20]. Visual inspection indicated a KIAA1549-BRAF fusion in the copy number profiles if a narrow gain of the 7q34 region, representing the duplication, was present. In the automated analysis, evidence of the KIAA1549-BRAF fusion in the copy number profile was given if the median intensity of the probes on 7q34, which are involved in the duplication region (range: 7:138.500.000-7:140.500.000, GRCh37), was 0.1 higher than the median intensity of all probes on the whole chromosome 7q, as well as 0.07 higher than the median intensity of all probes on 7q33 and 0.07 higher than the median intensity of all probes on 7q35. To avoid false positives, the fusion was not called automatically if one arm of chromosome 7 was split into 10 or more segments by the circular binary segmentation algorithm implemented in the R package DNAcopy [21]. This may occur in cases with low DNA quality or complex rearrangements such as chromothripsis on chromosome 7 [22].
Calling of the KIAA1549-BRAF fusion from DNA panel sequencing data To call KIAA1549-BRAF fusions from fastq files derived from DNA panel sequencing, we used the independent tools Manta [23] (version 1.6.0; with parameter --rna, all calls with "PASS" in the column "filter" kept), deFuse [24] with standard settings and Arriba (https://github.com/suhri g/Arriba, version 1.2.0 with STAR aligner [25] version 2.6.1e), initially with standard settings. In addition, we searched for KIAA1549-BRAF fusions in the 'discarded' output of the Arriba analysis for all cases for which no KIAA1549-BRAF fusion was indicated by Arriba in the main output file. The standard settings of Arriba were developed for analysis of RNA sequencing data. Thus, all fusions with no supporting reads overlapping an exon are automatically discarded. As the breakpoints on the DNA level are almost exclusively in intronic sequences, we retrieved all KIAA1549-BRAF fusions from the original discarded output of Arriba that had been discarded by the filters 'intronic', 'mismatches' or 'mismappers', provided that at least two supporting reads were present.
If more than one alignment variant of the KIAA1549-BRAF fusion was found in similar positions in the discarded output, we chose the one with the highest number of split reads.

Sample selection
The cohort was compiled from archival tissue for which ethical approval for research use was granted by local regulations.

Further data analysis
All further analyses were performed in R, [26] version 3.4.4. Plots of Figure S3 were created using the script 'draw_fusions.R' in R, available at https://github.com/suhri g/Arriba.

Automated and visual calling of a gain of 7q34, indicative of the KIAA1549-BRAF fusion, from copy number profiles calculated from DNA methylation data show a high correlation
To investigate whether an automated analysis of copy number profiles from DNA methylation data can be reliably used for a screening of a 7q34 duplication, we defined criteria for an automated analysis (see methods). We screened DNA methylation data from 19,532 samples (13,617 FFPE, 5,915 fresh frozen), classifiable with the brain tumour classifier, [18] for a gain of 7q34. A gain of 7q34 was detected in 732/19,532 (4%; 545 FFPE, 187 fresh frozen) with the automated analysis. An example of a copy number profile showing the typical gain is depicted in Figure 1. All cases with evidence of a fusion were independently re-analysed visually by an experienced neuropathologist. Comparison to the visual analysis showed accordance in the vast majority of cases (673/732, 92%; Figure 2A). The majority of discrepancies in cases with a gain in the automated analysis related to low DNA quality resulting in noisy copy number profiles (33/732 cases (5%) scored 'not evaluable', 26/732 cases (4%) 'possible gain' in the visual analysis).
After establishing a protocol for analysis of DNA panel data for the KIAA1549-BRAF fusion (see below), we repeated this analysis for all cases with FFPE material for which DNA panel data were available, and for which the KIAA1549-BRAF fusion was detected either by visual calling from copy number plots derived from DNA methylation data, independent of the classifier scores, or from DNA or RNA sequencing data, that is, cases in which the KIAA1549-BRAF fusion was detected with at least one method used (n = 354, Table   S2; Figure S1). Again, there was concordance in the majority of cases F I G U R E 1 Exemplary copy number profile from a pilocytic astrocytoma WHO grade I with a KIAA1549 ex. 16-BRAF ex. 9 fusion showing a 7q34 gain (arrow) caused by the tandem duplication of KIAA1549 and BRAF on the chromosome 7q34 locus   100% and a sensitivity of 96% in this cohort as compared to WGS data ( Figure 2B, Table S3). The results of the automated and visual analysis differed in nine samples (only one clearly discrepant; two "possible" and six "not evaluable" in the visual analysis).
In summary, we found that a gain of 7q34, indicative of a KIAA1549-BRAF fusion, can be detected from DNA methylation data with high sensitivity and specificity. The automated calling of a gain of 7q34 shows a high agreement with a visual analysis.

A gain of 7q34 is almost exclusively detected in a few glioma entities known to harbour KIAA1549-BRAF fusions
To determine in which tumour entities a gain of 7q34, indicative of a KIAA1549-BRAF fusion, occurs, we analysed to which methylation classes (MC) of the brain tumour classifier the 673 tumours belonged for which a gain was detected both with the automated and visual analysis (Table S4) In summary, all cases but one (1/673, 0.1%) with a gain of 7q34 were assigned to methylation classes of entities known to harbour the KIAA1549-BRAF fusion, that is, mainly pilocytic astrocytoma but also DLGNT and HGAP. With the automated analysis alone, a small number of additional cases with hints of a KIAA1549-BRAF fusion assigned to different methylation classes were detected.
However, many of these were scored "not evaluable" with a visual analysis ( Figure 2A; Table S4).
Additionally, we analysed methylation data of n = 3311 tumours that were classifiable with the sarcoma classifier (Koelsche C et al., accepted for publication) but not the brain tumour classifier, as the KIAA1549-BRAF fusion has not been described in sarcomas. The automated analysis predicted a gain of 7q34 in only 0.4% of cases (n = 14/3311; Table S5). Most of these cases were related to an increased signal-to-noise ratio and scored 'not evaluable' (n = 9/14, 64%) or 'possible' (n = 3/14, 21%) in the visual analysis. Only in two cases was a gain confirmed in the visual analysis. However, one of these cases was a pilocytic astrocytoma that was misclassified by the sarcoma classifier due to a high content of inflammatory cells.
The other one was a chondrosarcoma with a 7q34 gain that was not entirely prototypical and, as expected, no KIAA1549-BRAF fusion was detected by RNA sequencing.
Thus, we conclude that a gain of 7q34, indicative of a KIAA1549-BRAF fusion, in copy number profiles calculated from DNA methylation data is almost exclusively found in the few gliomas/glioneuronal tumours known to harbour KIAA1549-BRAF fusions.  Figure 2D).  Figure 3; Table S2) while Manta detected it in 72% of cases (n = 249/348). We achieved the highest detection rate of 90% with Arriba with modified filter settings (n = 314/348; Figure 3; Table S2). Since only in few cases the fusion was detected with Manta and/or deFuse but not with Arriba, use of all three fusion callers combined only slightly increased the overall detection rate from 90% with Arriba only to a total of 94% (n = 328/348; Figure 3). Thus, in this setting, Arriba with modified filter settings is well suited to detect the KIAA1549-BRAF fusion. Figure S3 illustrates an exemplary fusion.

The KIAA1549-BRAF fusion can be reliably detected from DNA panel sequencing data
To check whether these results are specific, we assigned all classifiable cases with DNA methylation data to methylation classes using the brain tumour classifier. The fusion was again only detected in different methylation classes of pilocytic astrocytoma, DLGNT, HGAP and control tissue, suggestive of a low tumour cell content (integrated diagnoses in Table S2). Thus, we conclude that the algorithm is specific.
In summary, the analysis of DNA panel sequencing data is well suited to detect the KIAA1549-BRAF fusion in a diagnostic setting.

The KIAA1549-BRAF fusion is detected in most cases by both copy number analyses from DNA methylation data and by fusion calling from DNA panel data
To evaluate whether calling of the KIAA1549-BRAF fusion from DNA methylation data or from DNA panel sequencing data is more sensitive, we selected cases for which both DNA methylation data and data from DNA panel sequencing were available, and in which the KIAA1549-BRAF fusion was detected with at least one of the methods used. Among these, we detected the fusion from DNA methylation data in 84% of the cases by visual analysis (Fig. S2). From DNA panel sequencing data, the fusion was again detected in more than For an additional validation, we selected all cases from our database for which RNA sequencing data from FFPE tissue was available, and for which the KIAA1549-BRAF fusion was detected either by visual calling from copy number plots derived from DNA methylation data or from any type of sequencing data (n = 30; Table S2).
In total, we screened RNA sequencing data from 1462 samples of various brain tumour and sarcoma entities with deFuse and Arriba for a KIAA1549-BRAF fusion. For one of the selected samples, only RNA sequencing data were available while for all other cases, DNA methylation data and for 23 cases also DNA panel data were available.
We then compared the detection of the KIAA1549-BRAF fu-

DISCUSS ION
In this study, we compared the performance of different approaches DNA methylation arrays, however, are becoming more and more widely used and can be used for computing a classifier diagnosis and a copy number profile at the same time [16,18]. Visual analysis of the copy number profile has a high sensitivity for the detection of this fusion as validated with the ICGC dataset in our study.
Moreover, the frequency of a gain of 7q34 and the age distribution in different methylation classes of pilocytic astrocytoma, DLGNT and HGAP were in line with the frequencies of KIAA1549-BRAF fusions reported earlier [6,7,9,12]. The fusion was almost exclusively detected in these tumours, confirming the specificity of the method.
Interestingly, in our large cohort, we never found a KIAA1549-BRAF fusion in gangliogliomas, for which the fusion has been described in single cases [29,30], or dysembryoplastic neuroepithelial tumours.
For use in a clinical setting, we recommend confirming a due to a poor signal-to-noise ratio, and both copy number analyses and panel sequencing may miss the fusion due to a low tumour cell content or poor DNA quality. Thus, in cases without evidence of a In conclusion, we show that the KIAA1549-BRAF fusion can be reliably detected both from copy number profiles generated from DNA methylation data and from DNA panel sequencing data in a diagnostic setting, and that this fusion is restricted to a handful of distinct molecular brain tumour classes.

Ethic s approval and consent to par ticipate
Tissue collection and processing as well as data collection were in compliance with local ethics regulations and approval.

ACK N OWLED G EM ENTS
We

PEER R E V I E W
The peer review history for this article is available at https://publo ns.com/publo n/10.1111/nan.12683.

DATA AVA I L A B I L I T Y S TAT E M E N T
All processed data are included in this published article and it's supporting information files. Raw data are available upon request for collaborative research projects.