Concordance of copy number abnormality detection using SNP arrays and Multiplex Ligation-dependent Probe Amplification (MLPA) in acute lymphoblastic leukaemia

In acute lymphoblastic leukaemia, MLPA has been used in research studies to identify clinically relevant copy number abnormality (CNA) profiles. However, in diagnostic settings other techniques are often employed. We assess whether equivalent CNA profiles are called using SNP arrays, ensuring platform independence. We demonstrate concordance between SNP6.0 and MLPA CNA calling on 143 leukaemia samples from two UK trials; comparing 1,287 calls within eight genes and a region. The techniques are 99% concordant using manually augmented calling, and 98% concordant using an automated pipeline. We classify these discordant calls and examine reasons for discordance. In nine cases the circular binary segmentation (CBS) algorithm failed to detect focal abnormalities or those flanking gaps in IKZF1 probe coverage. Eight cases were discordant due to probe design differences, with focal abnormalities detectable using one technique not observable by the other. Risk classification using manually augmented array calling resulted in four out of 143 patients being assigned to a different CNA risk group and eight patients using the automated pipeline. We conclude that MLPA defined CNA profiles can be accurately mirrored by SNP6.0 or similar array platforms. Automated calling using the CBS algorithm proved successful, except for IKZF1 which should be manually inspected.

IKZF1 plus 7 were developed using copy number data generated by Multiplex Ligation-Depend Probe Amplification (MLPA). MLPA is a multiplexed PCR assay that uses probes specific for exons within genes of interest, enabling the relative copy number of each covered exon to be reported as a ratio against other probe(s) within reference regions. In addition, MLPA has been used to assist the molecular diagnosis of a wide variety of diseases 8 . SNP arrays are also used for detection of CNAs with several advantages: detection of CNAs involving all genes rather than a defined set, other clinically relevant events such as aneuploidy, ploidy changes or partial chromosomal imbalances can also be detected. From a clinical diagnostic laboratory perspective, SNP arrays provide a useful real-time technique, they are genome wide and not disease-specific such as most MLPA kits. Thus, many diagnostic laboratories frequently have SNP array platforms in place for other cancer work or constitutional genetics. While previous studies have compared MLPA and SNP array copy number concordance, they have either focussed on single genes or have been applied to the analysis of germline DNA; neither of which are relevant to validation of the UKALL-CNA classifier. In B-ALL specifically, Dörge et al. investigated concordance among 25 ALL patients for IKZF1 abnormalities. They found 23 known IKZF1 deletions detected by SNP6.0 array also to be detectable by MLPA, with minor differences at exon level due to probe positioning 9 . In addition, Zanardo et al. investigated a cohort of 93 patients with developmental and congenital abnormalities, using MLPA and SNP arrays and showed a 98% concordance between the two techniques 10 . To our knowledge, there has been no large-scale systematic study investigating the comparative performance of CNA detection using SNP array and MLPA techniques for an established set of key prognostic genes and genomic regions in leukaemia.
It is important to validate biomarkers using different techniques as well as different patient cohorts. The UKALL-CNA and the IKZF1 plus profiles have been validated across different patient populations. In this study, we validate the use of SNP arrays for detection of key CNAs, which underpin these two copy number profiles in ALL. In addition, we sought to compare the performance of the two techniques in order to assess both the concordance of SNP6.0 arrays and MLPA, as well as the ability of automated SNP array analysis to call CNAs.

Methods
Cohort and classification criteria. Patients were diagnosed with B-cell precursor ALL by standard flow-cytometric criteria and were treated on Medical Research Council (MRC) ALL97/99 (1997)(1998)(1999)(2000)(2001)(2002) or United Kingdom UKALL2003 (2003)(2004)(2005)(2006)(2007)(2008)(2009)(2010)(2011). Full details of these treatment regimens have been previously published 11,12 . Local ethical committee approval was obtained for ALL97 by individual treatment centres, whereas approval for UKALL2003 was obtained from the Scottish Multi-Centre Research Ethics Committee. Informed consent was given by parents and patients in accordance with the Declaration of Helsinki. Briefly, in ALL99 and UKALL2003, patients were assigned to regimen A or B based on whether they were National Cancer Institute (NCI) standard (<10 years old and WCC < 50 × 10 9 /L) or high risk (≥10 years old or WCC ≥ 50 × 10 9 /L), respectively.
Cytogenetic and FISH testing was performed on pre-treatment bone marrow samples by member laboratories of the UK Cancer Cytogenetics Group or centrally by the Leukaemia Research Cytogenetics Group, and results were reported using established nomenclature and definitions 3 . Patients with absence of the following chromosomal abnormalities: ETV6-RUNX1, high hyperdiploidy (51-65 chromosomes), BCR-ABL1, KMT2A/MLL rearrangements, near haploidy (<30 chromosomes), low hypodiploidy (30-39 chromosomes), intrachromosomal amplification of chromosome 21 (iAMP21) or TCF3-HLF; were classified as B-other-ALL and were included in this study. MLPA. MLPA was performed on DNA extracted directly from pre-treatment bone marrow samples, all samples had >90% blasts, sample handling and processing of cells was as previously described in Schwab et al. 19 . The same DNA source was used for MLPA and SNP6.0. The SALSA MLPA kit P335 (MRC Holland, Amsterdam, The Netherlands), which includes probes for IKZF1, CDKN2A/B, PAX5, EBF1, ETV6, BTG1, RB1, and the PAR1 region, was used to identify CNAs involving these genes 20 . Previous studies have demonstrated that MLPA can accurately detect deletions within each of these genes when present in 20% or more cells 19,20 . A detailed description and breakdown of each CNA and the correlation with specific chromosomal abnormalities for all the patients within these two cohorts has been previously published 20 . MLPA data analysis methodology and peak ratio values used to make MLPA calls are as previously described in Schwab et al. 19 .     Table 1. i = no SNP array probes covering the abnormality -meaning that due to the array design it is impossible to make a call. ii = MLPA result was informed by a single MLPA probe. iii = The CBS algorithm failed to pick up a small focal abnormality, which could be called visually by inspection of the LRR data such that they would be concordant with the MLPA calls. iv = The two techniques disagree despite sufficient probe coverage. In one of these six cases of disagreement FISH data was available and validated the normal calls made by SNP6.0 this is indicated by †. v = SNP6.0 detected an abnormality which could not be detected by MLPA due to lack of probe coverage. www.nature.com/scientificreports www.nature.com/scientificreports/ Copy number call comparison. Discordant calls between MLPA and the automated array pipeline were extracted from both datasets, each discordant call was manually inspected by three independent reviewers (MB, RH, AE) by visualising the LRR data for the gene of interest. This approach also permitted manual calling of focal events that were discernible by eye in the LRR plots but missed using the CBS algorithm. Joint manual review was undertaken to determine the nature and likely cause of discordance.
For the purpose of this study, we have considered losses/deletions and gains/amplifications, as the same event owing to the use of different copy number thresholds and sensitivities between SNP6.0 and MLPA. Additionally, as no formal definition or threshold for copy number gain versus amplification and similarly loss versus deletion exist, we have chosen to aggregate these two pairs of events in to either gain/amplification or loss/deletion. Details of MLPA probe location determination methodology and representativeness of our cohort is given in the on-line supplemental methods section.

Results
Concordance of copy number abnormality calls. We compared the copy number calls from MLPA and our SNP array pipeline for 143 diagnostic samples from B-other-ALL patients treated on ALL97-99 or UKALL2003. The demographics of these patients are detailed in Supplementary Tables S1 and S2. Specifically, we compared the genes included in the P335-MLPA kit used in the UKALL-CNA classifier: IKZF1, CDKN2A/B, PAX5, EBF1, ETV6, BTG1, RB1, and the PAR1 region (CRLF2/CSF2RA/IL3RA) 4 .
Using the value of copy returned by the CBS algorithm, we found 98% of the calls to be concordant between the SNP6.0 pipeline and MLPA, with 25 calls discordant out of a total of 1,287 (Table 1). Manual event calling by visual inspection of the plotted LRR data for cases which the CBS algorithm missed (nine cases) improved concordance to 99% reducing the number of discordant calls to 16 (Table 1). Discordant calls were classified into one of five categories: (i) no SNP array probes covering the abnormality -meaning that due to the array design it was impossible to make a call (n = 5). (ii) MLPA result was informed by a single MLPA probe (n = 2). (iii) The CBS algorithm failed to pick up a small focal abnormality, which could be called visually by inspection of the LRR data, such that they would be concordant with the MLPA calls (n = 9). (iv) The two techniques disagreed despite sufficient probe coverage (n = 6). (v) SNP6.0 detected an abnormality which could not be detected by MLPA due to lack of probe coverage (n = 3). Table 2 shows the discordant calls on a case-by-case basis, using both the manually augmented and fully automated CBS calling methods. Figure 1 shows examples of typical concordant and discordant profiles for each of the five categories.
MLPA and SNP6.0 probe locations. For the genes represented in the P335 MLPA kit, we investigated the precise location of MLPA probes in comparison to the SNP6.0 probes in order to determine whether their positioning may explain any of the observed differences in calls. The top scoring BLAT 21 alignment for each probe uniquely located each probe sequence (provided by MRC Holland, Amsterdam, The Netherlands) to a single fully aligned region of the genome without gaps. The co-ordinates of the aligned MLPA probe locations are given in Supplementary Table S3 and plotted in Figs. 2 and 3. Most of the genes targeted by the MLPA kit were well covered by probes on the SNP6.0 array. However, a few genes had poor probe coverage on the SNP6.0 array: CDKN2A lacked probes internal to the gene; while the genes CRLF2/CSF2RA/IL3RA within the PAR1 region were covered by a single probe. The MLPA kit had been designed specifically to detect copy number events within these genes, hence the probes were well positioned within the exons. However, not every exon of every gene was covered, as is the case in RB1, EBF1 and PAX5. The partial coverage of EBF1 and PAX5 resulted in MLPA missing three small deletions, which were detected by SNP6.0. These three cases are illustrated in Fig. 1(B) (EBF1) and Fig. 4 (PAX5). CDKN2B was also only covered by a single MLPA probe covering exon 1. Issues arising from these differences in coverage are discussed in the next section.
Examination of discordant calls. Of the 25 discordant calls made by the SNP6.0 pipeline using the CBS algorithm, subsequent manual examination of the LRR data plots produced by DNACopy for these calls revealed that nine of the abnormalities were visible by eye but had evaded detection by the CBS algorithm (disagreement category iii). Manual inspection of LRR plots in these nine cases allowed for concordant deletion of these otherwise missed abnormalities, eight in IKZF1 and one in PAX5. For the IKZF1 gene, 21/29 deletions were called, with eight deletions missed by the CBS algorithm but visible in the LRR data plots, in seven out of these eight cases, IKZF1 deletions spanned exons 4 to 7/8. In these cases, the CBS algorithm failed to call focal abnormalities which flanked a region of the IKZF1 gene, where there was no probe coverage within intron 3 (between exons 3 and 4, ~42Kb in length) on the SNP6.0 platform. It should be noted that all array platforms lack coverage of this region of IKZF1, due to the repetitive nature of the reference genome sequence of this region and the inability of array platforms to target probes complementary to this region in a sequence-specific way. In these cases, visual inspection allowed for manual calling of the abnormality; an example of an event downstream of the "IKZF1 hole" missed by the CBS algorithm is shown in Fig. 1(C). Here the black line indicates that segmentation is normal, but LRR data points reveal a deletion concordant with MLPA data. A further two cases of exon 4-5 and exon 4-8 deletions were successfully detected by CBS, so whilst it was possible to detect these deletions, the presence of a gap in coverage spanning exons 3-4 clearly presents problem for the CBS algorithm when dealing with events The grey dotted box shows the extents of CDKN2A in which no SNP6.0 LRR data points can be found. (E) An example of a case in which the two techniques disagree (discordance category iv). MLPA shows a deletion in two MLPA probes within exon 1 of ETV6, no concordant deflection of the SNP6.0 LRR scatter around the copy number normal log 2 value of zero is observed. (F) Observation from a single MLPA probe which covers CDKN2B and indicates a loss, no concordant shifts can be seen in the LRR values of many SNP6.0 probes. www.nature.com/scientificreports www.nature.com/scientificreports/ immediately upstream of the gap. This issue was easily surmounted by manually inspecting plotted LRR data for the IKZF1 gene.
A single gain within PAX5 was missed by the CBS algorithm. It was highly focal and evidenced by a single probe, which showed a clearly distinct LRR value of ~1.5. Given the gain was only identified by one probe, it would be impossible and unwise for an automated method to call. We employed a minimum number of markers cut-off of five for the CBS algorithm in our study, meaning that for an event to be detected at least five probes must be involved with a statistically significant deviation. We also tested using a minimum cut-off of four, three and two markers; none of these options resulted in any of the nine missed events being called. These missed events were not due to the CNA thresholds that we applied over the top of the segmentation data, used to enable categorical www.nature.com/scientificreports www.nature.com/scientificreports/ labelling of gains, amplifications, losses, and deletions (as illustrated in Supplementary Fig. S3). Simply, these missed calls occurred because the CBS algorithm did not segment the events.
In the remaining 16 discordant calls, we found that eight calls resulted from design differences in the two platforms, such that abnormalities detected in one could not be observed in the other owing to a lack of probe coverage; these cases belonged to disagreement categories (i) SNP6.0 lacking probes, and (v) MLPA lacking probes. For category (i) Fig. 1(B) shows an example of a focal deletion within CDKN2A, which was not detectable by SNP6.0 because no probes lie within the gene itself (Fig. 2(E)); this situation occurred in three cases. We also observed this type of discordance in two cases involving the PAR1 region, where the abnormalities occurred in a region within CSF2RA/IL3RA, as there are no probes on the array in this area (Fig. 3(C)). Importantly, none of these cases (n = 8) reflected a failure of the CBS algorithm to detect the events, they were simply not observable using this particular array. The MLPA kit also failed to detect events due to the absence of probes in the appropriate region (discordance category v); we observed three cases of this type. The SNP array was able to detect a focal loss in EBF1 (Fig. 1(B)). There were also two cases of focal deletions in PAX5 observed on the SNP array that were not detected by the MLPA kit. These both occurred in exon 7 of PAX5, which has no coverage by MLPA probes (Fig. 3(B)).
It is also worth highlighting that even when abnormalities called by both methods were concordant, the SNP6.0 pipeline was able to reveal extra details and events not observable by MLPA. Figure 5 highlights three cases of concordant PAX5 calls, which revealed extra levels of detail regarding copy number events by the SNP6.0 array using our pipeline that were not observable using the MLPA kit. www.nature.com/scientificreports www.nature.com/scientificreports/ There were two cases where MLPA detected a focal deletion of CDKN2B with just a single probe (type ii). Both of these discordant cases occurred within CDKN2B, in which normal values of copy were observed from the many SNP6.0 probes in this region. Here we consider the many repeated observations from the array to be the most reliable compared to the single observation from the MLPA kit; an example is given in Fig. 1(F), and relative probe locations are shown in Fig. 2(B). These two cases illustrate the potential limitation in measuring the copy number of a gene using a single probe in the MLPA kit, which is error prone and unreliable (in part due to the inability of MLPA probes to tolerate SNPs) versus the multiple observations taken from the SNP6.0 array. For this reason, normally ≥2 adjacent MLPA probes are required for calling a copy number alteration 19 .
We observed only six calls in total where the two techniques were in disagreement and the supporting evidence was not concordant despite sufficient probe coverage, (disagreement category iii). In all these cases, MLPA detected an abnormality but there was no obvious or robust deflection of LLR data points by SNP array. In one of these cases, FISH data was available for the gene in question (PAX5), which further corroborated the SNP6.0 call (highlighted in Table 2). In the remaining discordant cases, it was possible that differences in sensitivity of the two techniques to sub-clonal events means that events detected by one are not observable by the other. We also checked whether any patient samples were repeatedly discordant indicating a sample swap. All discordant cases, came from different samples, with the exception of a single CDKN2A and CDKN2B deletion which occurred within the same patient (22572). As these two genes are adjacently located, it was likely the same event detected by MLPA.
Effect of using SNP6.0 array data with the UKALL-CNA classifier. Previously, we designed 4 and validated 22 a copy number based ALL risk classifier (UKALL-CNA classifier), which segregates patients based on the copy number states of the eight genes and the PAR1 region included in the MLPA P335 kit. This classifier assigns patients into three groups: good risk (GR), intermediate risk (IR), and poor risk (PR), with significantly different outcomes. In this study, we have grouped together the IR and PR group as the major prognostic findings from our discovery and validation studies was that B-other-ALL patients with a GR CNA profile had an excellent outcome 4 and B-other-ALL formed the bulk of this cohort (121/143 cases). We used the same algorithm to segregate 143 samples by UKALL-CNA classifier using both the MLPA CNA calls and our SNP6.0 pipeline with the fully automated CBS or manually augmented calls. Using manually augmented calling, 97% (n = 139) of patients would have been assigned to the same CNA-risk profile (Table 3). Here four patients with different profiles, were assigned to GR using the SNP calls in place of an IR/PR assignment using the original MLPA calls. Using the fully automated SNP CBS calls an additional four cases, eight in total were moved from the IR/PR group to GR (Table 3). Table 2 shows the changes in UKALL-CNA classifier result for each discordant case.

Discussion
In this study, we have validated the use of SNP arrays to call MLPA defined UKALL-CNA profiles, which have been previously validated across several patient cohorts 22 . We investigated concordance between two different techniques: SNP6.0 array and MLPA for the detection of copy number alterations across key genes somatically altered in B-ALL. Our aim was to investigate how well SNP arrays perform for risk classification of B-other-ALL using the UKALL-CNA classifier 4,22 in comparison to the MLPA kit 19 . These results are important because they provide good evidence that the UKALL-CNA profile can be reliably called using a different platform. Hence, future protocols, which decide to adopt the UKALL-CNA classifier into their risk stratification algorithm, can be confident that diagnostic laboratories using SNP arrays will be calling the same profile identified by the research studies using MLPA. Although we have focussed on validating our own CNA profile using a single MLPA kit (P335), this research has implications for other MLPA derived profiles; for example the IKZF1 plus profiles recently (A) Shows a concordant PAX5 call in this case both techniques reported a deletion, however, the SNP6.0 CBS segmentation show a nested deletion event in exons 6-7 which was not detectable using the MLPA kit because these exons have no coverage (Fig. 3(B)). (B) Similarly shows a nested deletion of exon 9 in PAX5. (C) Here both techniques show a gain, the gain here taking precedence in SNP6.0 as the reported event as it is higher than the lowest nested deletion event, (CBS copy number values of 3.6 vs 1.1 respectively). The nested loss here was not detectable via MLPA as no probe covers this region, additionally this copy number pattern is characteristic of a PAX5 rearrangement which is of diagnostic significance.  www.nature.com/scientificreports www.nature.com/scientificreports/ defined and validated by the BFM and AEIOP study groups 7 . SNP arrays are increasingly being used in a diagnostic setting to detect a wide range of losses, gains, and loss of heterozygosity in cancer. This is evidenced by the fact that the American College of Medical Genetics and Genomics (ACMG) and the Cancer Genomics Consortium (CGC) have issued guidelines for diagnostic laboratories analysing SNP arrays 23 . Moreover, chromosomal ploidy (for example, near-haploidy and low hypodiploidy) and arm-level events normally derived from conventional karyotyping are also detectable using SNP arrays which are key in ALL 24 . As SNP arrays are genome-wide, their routine application means that new risk associated CNA events in other genes can be immediately investigated. Additionally, our SNP6.0 pipeline employs automated copy number calling, using the CBS algorithm 16 . Automated calling via the CBS algorithm has advantages over manual calling, being much quicker, and scaling systematically beyond a handful of genes to the whole-genome in an unbiased way, while returning only statistically significant copy number events 17 . Our pipeline uses the popular open source packages PennCNV 13,14 and DNACopy 16,17 and can be easily replicated elsewhere, additionally, data exported from either Affymetrix Genotyping Console / Chromosome Analysis Suite and Illumina Genome Studio can be analysed with these tools to yield high quality segmentation via the CBS algorithm. The freely available Nexus Copy Number Plugin for Genome Studio also has an implementation of the CBS algorithm, which is usable with Illumina SNP array data. The CBS algorithm is also available for CNA calling of high-throughput sequencing data in methods such as CNVKit 25 .
Using our SNP6.0 analysis pipeline, we found 98% concordance between SNP6.0 and MLPA across 1,287 individual calls and 99% concordance if we allowed for manual inspection of IKZF1 and PAX5 LRR plots. Of these 16 remaining discordant calls, only five cases (4%) were undetectable by SNP6.0 owing to design limitations in the location of probes.
The UKALL-CNA risk classifier validated in this study is a useful tool to help segregate patients with differential relapse risk for treatment stratification. Here we have assigned UKALL-CNA profile based risk classification using both manual and automated SNP copy number calls for each patient in comparison to the original MLPA based classification. We observed a small proportion of patients who would have been assigned to different risk groups (3% for manual and 6% for automated calls). Most of the inconsistencies (four out of eight) arose from the lack of information due to automatically calling the copy number states of IKZF1. Among those cases where a lack of coverage of probes on the array led to three and two events being missed in CDKN2A and PAR1 respectively, (disagreement category i in Tables 1 and 2), the resultant differences in UKALL-CNA profile risk classification over MLPA informed classification would potentially lead to two patients (included in the four above) being assigned to a different CNA risk group.
Our observations have indicated that the SNP6.0 pipeline, employing the CBS algorithm and manually augmented calling for IKZF1, could be used as an alternative to MLPA based copy number detection systems in ALL CNA profile based risk classification. It has also highlighted the need to improve CBS based callers in order to cope with missing information, such as the "IKZF1 hole" in the SNP6.0 and other array platforms, which is also problematic for genomic alignments of high-throughput sequencing data. Finally, we highlight that other array-based platforms will require similar scrutiny with respect to probe location and coverage before being employed to call profiles derived using targeted copy number analysis techniques like MLPA.