Complex structural variation is prevalent and highly pathogenic in pediatric solid tumors

Summary

In pediatric cancer, structural variants (SVs) and copy-number alterations contribute to cancer initiation as well as progression, thereby aiding diagnosis and treatment stratification. Although suggested to be of importance, the prevalence and biological relevance of complex genomic rearrangements (CGRs) across pediatric solid tumors is largely unexplored. In a cohort of 120 primary tumors, we systematically characterized patterns of extrachromosomal DNA, chromoplexy, and chromothripsis across five pediatric solid cancer types. CGRs were identified in 56 tumors (47%), and in 42 of these tumors, CGRs affect cancer driver genes or result in unfavorable chromosomal alterations. This demonstrates that CGRs are prevalent and pathogenic in pediatric solid tumors and suggests that selection likely contributes to the structural variation landscape. Moreover, carrying CGRs is associated with more adverse clinical events. Our study highlights the potential for CGRs to be incorporated in risk stratification or exploited for targeted treatments.

Graphical abstract

Highlights

• •WGS data from 120 tumor-normal pairs across 5 pediatric solid tumor types
• •Complex genomic rearrangements are observed in 47% of tumors
• •Most complex rearrangements affect cancer drivers or lead to unfavorable alterations
• •Carrying complex genome rearrangements is associated with adverse clinical events

Van Belzen et al. show through systematic analysis of pediatric solid tumors that despite the overall low tumor mutation burden, complex structural variants are common. Moreover, events such as extrachromosomal DNA, chromoplexy, and chromothripsis are likely pathogenic and shape the structural variation landscape of these tumors.

Introduction

Structural variants (SVs) and copy-number (CN) alterations occur frequently in pediatric cancer and can be leveraged for diagnosis and treatment stratification as they often contribute to cancer initiation and progression. Well-known examples include the EWSR1::FLI1 fusion gene in Ewing sarcoma (EWS),¹ the PAX3/7::FOXO1 fusion gene in fusion-positive rhabdomyosarcoma (FP-RMS),² and MYCN amplification in neuroblastoma (NBL).³ Recent studies have highlighted the potential clinical utility of considering the underlying mutational mechanisms of these driver alterations, as they may indicate differences in tumor aggressiveness. In EWS, patients with fusion genes arising from complex rearrangements have a worse prognosis than those with fusions resulting from simple reciprocal translocations.¹ Similarly, NBL patients with more complex structures of the MYCN focal amplification (amplicon) and co-amplification of oncogenes seem to form a higher risk group.⁴ However, a systematic characterization of complex structural variation patterns across pediatric solid tumors is lacking.

Complex genomic rearrangements (CGRs) are characterized by clusters of breakpoints reflecting the repair of multiple double-stranded DNA breaks that likely occurred at the same time. In recent years, many different CGR types have been distinguished based on patterns in sequencing data and hypotheses of how the DNA damage has occurred and was repaired. Examples include chromothripsis, chromoplexy, breakage-fusion bridge (BFB), local jumps, pyrgo, rigma, and tyfona.^5,6 However, the criteria of CGR types have been inconsistently used across studies, especially between cancer, developmental disease, and germline.⁷ Recent mechanistic studies using single-cell sequencing have shown that many different complex rearrangement patterns can arise from straightforward events during a single cell division, such as chromatin-bridge breaking after telomere fusion resulting in focal chromothripsis-like patterns.^8,9 Furthermore, the observed patterns of gains and losses could be explained by the unequal division of genomic material and do not require complicated molecular mechanisms involving DNA synthesis.^8,9 Considering these aspects, Zhang and Pellman (2022) conclude it might be difficult to distinguish the underlying mechanisms, such as BFB and chromothripsis, based on CN patterns alone.⁸

Alternatively, the umbrella term “chromoanagenesis” can be used to encompass a spectrum of complex rearrangements that result in rearranged derivative chromosomes. Long-read and Hi-C data of germline genomes also support this diversity of rearrangements and categorization into chromothripsis-like and chromoplexy-like patterns.¹⁰ Chromothripsis is characterized by DNA shattering and haphazard repair in which genomic fragments are randomly joined, resulting in many interleaved SVs and an oscillating CN pattern with losses.¹¹ However, chromoplexy is a CN balanced complex rearrangement characterized by translocations between multiple chromosomes. It can result in fusion genes such as EWSR1::FLI1 and is thought to arise when DNA damage occurs during co-localization of chromosomes in transcription hubs.¹ Both chromothripsis and chromoplexy can result in derivative chromosomes, differing mainly in the presence or absence of large CN changes.^10,12

In contrast, extrachromosomal DNA (ecDNA) fragments can result in high-level amplifications of oncogenes. A well-studied example includes MYCN amplification in NBL: an initiating event excises the oncogene, after which it is amplified as a circular ecDNA fragment.^4,13 The ecDNA can undergo further rearrangements and either remain an extrachromosomal fragment or integrate back into the genome as a homogeneous staining region.¹⁴ The initiating event usually has characteristics of chromothripsis or BFB and can also include additional genomic loci such as cancer genes or distal enhancers.^4,13,14 This can give rise to a great diversity of ecDNA amplicon structures, and research into associations with treatment resistance, tumor aggressiveness, and patient outcome is ongoing.^14,15,16

Here, we systematically characterized complex structural variation patterns across pediatric solid tumors and identified candidate pathogenic CGRs that likely contribute to tumorigenesis. To study CGRs agnostic to type, we first detected clusters of SV breakpoints representing CGRs and subsequently categorized them into chromothripsis, chromoplexy, or ecDNA/amplicons. We found that CGRs occurred in approximately half of our cohort (56 of 120 tumors) and that the same CGR types occur across multiple cancer types, indicating similarities in mutational mechanisms. Furthermore, recurrently altered genomic regions often overlap with cancer genes and are mostly cancer-type specific, which suggests that selection is involved in shaping the observed complex structural variation landscape. In 75% of tumors carrying a CGR (42 tumors), we identified candidate pathogenic CGRs that affect cancer driver genes or result in unfavorable chromosomal alterations associated with poor prognosis. Furthermore, patients whose tumor carried a CGR involving a cancer driver gene experienced a clinical event twice as often compared to patients without CGRs affecting cancer driver genes. In conclusion, our results indicate that CGRs are highly pathogenic in pediatric solid tumors and that the clinical implications of these rearrangements warrant further study.

Results

Patterns of complex structural variation across pediatric solid tumors

To investigate pan-cancer patterns of chromothripsis, chromoplexy, and ecDNA, we analyzed somatic SVs in a cohort of 120 patients across 5 types of pediatric solid tumors (Figure S1). Paired tumor-normal whole-genome sequencing (WGS) data were generated from primary tumors of patients diagnosed with NBL (n = 39), EWS (n = 22), Wilms tumor (WT, n = 34), hepatoblastoma (HBL, n = 7) and RMS (n = 18) (Table 1). Overall, these tumor genomes have an anticipated low tumor mutation burden of SNVs and indels (median 0.32/Mb) and a fraction of genome altered (FGA) by CN alterations (median 12.6%) (Figure 1A), compared to adult cancers.¹⁷ Additionally, we identified a median of 14 somatic SVs with >0.1 tumor allele fraction (Figure 1A). To infer CGRs, we identified clusters of SV breakpoints within a 5-Mbp interval and categorized them into different types of CGRs based on a combination of SV and CN features (Figure 1B). To compare CGR patterns across cancer types, we further focused on three established types (chromothripsis, chromoplexy, and ecDNA).

Table 1.

Cohort demographics

Cancer_type	n	Female	Male	Age, y [0,1]	Age, y [1,3]	Age, y [3,8]	Age, y [8,17]
Overall	120	59	61	39	23	28	30
Alveolar rhabdomyosarcoma	5	2	3	1	0	3	1
Embryonal rhabdomyosarcoma	13	5	8	3	2	3	5
Ewing sarcoma	22	11	11	0	2	2	18
Hepatoblastoma	7	5	2	5	1	1	0
Nephroblastoma	34	17	17	14	8	10	2
Neuroblastoma	39	19	20	16	10	9	4

Breakdown of the cohort with regards to age and sex. See also Table S6.

Figure 1.

Occurrence of CGRs in pediatric solid tumors

(A) Number of SVs per tumor colored by CGR type (top), number of nonsynonymous SNVs/indels (center), and FGA by CN gain or loss (bottom) across pediatric solid tumors. From left to right: Ewing sarcoma (EWS), neuroblastoma (NBL), fusion-positive rhabdomyosarcoma (FP-RMS), fusion-negative rhabdomyosarcoma (FN-RMS), Wilms tumor (WT), and hepatoblastoma (HBL). Symbols denote genomic instability mutations: TP53 disruption (triangle), MDM2 amplification (circle). ∗M002AAB has a germline TP53 alteration.

(B) Characteristics of CGR types and circos plots of examples. From left to right: ecDNA/amplicon-type (patient M721AAC), chromoplexy (patient M135AAD), and chromothripsis (patient M050AAB). CGRs are categorized based on the following core characteristics: ecDNA/amplicon, SVs with breakpoints within 1 kbp of amplicons; chromoplexy, SVs form a closed cycle and connect multiple chromosomes via CN-balanced interchromosomal breakpoints; chromothripsis, footprints with oscillating CN segments and at least 10 overlapping SVs of mixed types, indicating randomly joined fragments. CGRs not complying with these criteria were categorized as complex other. See STAR Methods for more details. The circos plots contain SV breakpoints (links) involved in the CGR and depict gains (red) and losses (blue) of genomic material of the affected chromosomes.

Applying this SV clustering and categorization approach, we identified CGRs in 56 tumors (47%) across all 5 cancer types (Figure 1; Table S1). Most of these tumors (n = 31) carry a single CGR, but those with mutations in TP53 or MDM2 carry multiple CGR events (median 3.5) and have a higher tumor mutation burden (median 0.80 vs. 0.32, p < 0.01) and FGA (41% vs. 12%, p < 0.01). However, not all tumors with a CGR have a mutation in TP53 or MDM2, illustrating that it is not a prerequisite for complex rearrangements to occur or to be tolerated (Figure 1A). Each CGR type was identified in at least two of the five solid cancer types. In total, chromoplexy was identified in 16 tumors (EWS, WT, RMS, and NBL), ecDNA/amplicon in 16 tumors (NBL and RMS), and chromothripsis in 8 tumors (NBL, WT, HBL, and RMS) (Figure 1A; Table S1). This demonstrates that CGRs occur in many pediatric solid tumors.

Chromoplexy was the most prevalent in EWS; it was detected in seven tumors as the underlying mechanism generating the pathognomonic driver fusions with EWSR1 (Table S1). In four other EWSs, we also identified CGRs underlying the driver fusions. However, these were categorized as “other” as they either did not form a closed cycle or included an unbalanced translocation and therefore did not pass all criteria for canonical chromoplexy (Table S1). Chromoplexy was also detected in RMS (n = 2), NBL (n = 3), and WT (n = 4) (Figure 1A), making it the most widespread CGR type. Many of these events will remain undetected by exome sequencing given their CN balanced nature, stressing the importance of performing WGS as part of the molecular diagnostic process.

Amplicon-overlapping CGRs were identified in NBL and RMS (Figure 1A; Tables S1 and S2). In seven of nine NBLs with MYCN amplification, we detected overlapping CGRs that likely reflect ecDNAs and typically consist of interleaved SVs of mixed types, sizes, and variant allele fractions (Table S1), as expected.^4,13 In addition, we identified breakpoints connecting multiple amplified loci that indicate these are part of the same ecDNA construct (Table S2). Apart from MYCN amplification in NBLs, we also identified CGRs overlapping amplicons in nine RMS tumors (Table S1). Although their genomic locations differ, the underlying rearrangements of these complex events resemble those of the NBLs, such as amplified loci on different chromosomes that are physically linked (Table S2). This shows that ecDNA is a more widespread phenomenon than MYCN amplifications in NBL alone.

Chromothripsis-like CGRs were identified across multiple cancer types in eight tumors (Figure 1A). Full-chromosome chromothripsis was rare in our cohort. We detected it in only two NBLs: M050AAB has an additional chromothriptic copy of chr17 and M575AAC has subclonal chromothripsis of chr15. However, focal chromothripsis was identified more often, namely in HBL (n = 3), WT (n = 1), and RMS (n = 2), and resulted in either loss of the remainder of the chromosome arm (n = 4) or novel derivative chromosomes (n = 2) consistent with chromatin bridge-breaking events (Table S1).⁹ Furthermore, the presence of chromothripsis did not require TP53 or MDM2 to be altered. One WT with focal chromothripsis had biallelic TP53 disruption, but we did not identify alterations in TP53 or MDM2 in the other seven tumors with chromothripsis (Figure 1A). This substantiates that chromothripsis can result from a single event and does not necessarily indicate ongoing genomic instability due to inactivation of the P53 pathway.

Overall, the characteristic patterns of ecDNA/amplicons, chromoplexy, and chromothripsis were commonly observed across multiple solid cancer types. In line with previous work, we identified chromoplexy underlying fusion genes in EWS and ecDNA underlying MYCN amplifications in NBL. However, we also identified similar CGRs in other cancer types (Figure 1A). Therefore, we further investigated pan-cancer patterns of CGRs and assessed their potential contribution to tumorigenesis.

Hotspots of CGRs point to potentially pathogenic events and not genome fragility

To analyze whether CGRs affect the same genomic regions across different tumors, we conducted a systematic genome-wide analysis and identified genomic regions that have SV breakpoints in more than 2% of our cohort (n ≥ 3, Figure 2; Table S3). In total, 13 hotspot regions were affected by CGRs in ≥3 tumors (Table S3). Among these hotspots were regions mostly comprising SVs underlying known driver alterations, namely with FOXO1, EWSR1, FLI1, MYCN, and ALK. In addition, novel CGR hotspots outside of the known driver events were identified on chromosomes 1, 11, 12, 16, and 17. These include candidate regions of interest that harbor cancer genes, such as WT1 and MDM2, or where chromosomal alterations have previously been associated with poor outcome, such as chr1p.^3,18,19 Furthermore, the breakpoints that make up the CGR hotspots often originate from the same cancer type (Figure 2; Table S3).

Figure 2.

Hotspots of CGRs overlap with cancer driver genes

Genome-wide overview of recurrently altered genomic regions (Table S3) with SV breakpoints from three or more tumors (gray bars). CGR hotspots containing SV breakpoints that are part of a single CGR from three or more tumors are highlighted (circles, thick lines), colored according to their cancer type and annotated with overlapping cancer driver genes.

To assess whether the hotspots reflect genome sequence propensity for rearrangement, we analyzed breakpoint regions for enrichment in repetitive elements, such as short and long interspersed nuclear elements (SINEs/LINEs) (see STAR Methods). While a fraction of all SVs is likely repeat mediated, with both start and end breakpoints near repeat elements of the same class, these are less often part of CGRs and instead occur as simple SVs (7.6% vs. 27%, p < 0.01). When only considering CGR breakpoints located in hotspots, an even lower proportion is possibly repeat mediated (4%). Consistent with this, the CGR hotspots are depleted of repeat-mediated SVs (6% vs. 18%, p < 0.01) compared to the remainder of the genome, regardless whether they are simple or complex, suggesting that the observed recurrence is not likely due to genome fragility.

Alternatively, we hypothesized that selection plays a role in the formation of CGR hotspots. Large-scale SVs are highly dysregulating and therefore unlikely to be selectively neutral; hence, the observed recurrence can indicate that these alterations provide a competitive advantage to the tumor cells.²⁰ CGR hotspots are significantly enriched for SVs with breakpoints inside (46% vs. 5.9%, p < 0.01) or nearby (93% vs. 34%, p < 0.01) known pediatric cancer genes. This does not solely arise from SVs being either complex or simple. Comparing complex to simple SVs shows that they have only slightly more breakpoints near pediatric cancer genes (43% and 36%, p < 0.05) and similar fractions inside the gene body (9.4% and 9.3%, not significant). However, there are marked differences between CGR types: 26% of chromoplexy breakpoints are located inside gene bodies compared to just 5.7% and 4.5% for ecDNA/amplicons and chromothripsis, respectively. This fits with the hypothesis that the mutational mechanism underlying chromoplexy involves co-localization of highly expressed genes in transcription hubs. For ecDNA/amplicons, the SV breakpoints tend to flank the genes instead of residing in gene bodies, while the amplified region spans the gene. Moreover, amplicons often hit known cancer genes (27 of 34; Table S2), also indicating specificity. In contrast, chromothripsis events tend to cause more widespread disruptions and have more breakpoints overall, some of which can intersect genes by chance. Since chromothripsis is expected to arise from random breakage, recurrent events are of particular interest. Almost all hotspots either reflect cancer driver events known to occur in these pediatric cancers or they harbor relevant cancer genes or chromosomal alterations (Table S3), suggesting that these recurrent CGRs are pathogenic events.

CGRs provide insights into underlying genomic rearrangements of driver alterations

To identify potentially pathogenic CGRs, we selected events that affect cancer-type-specific driver genes or result in unfavorable chromosomal alterations that have been associated with poor outcome or high risk (Figure 3; Table S1). In total, we identified 48 candidate CGRs in 42 tumors (75% of tumors with CGRs, 45% of all CGRs), half of which are drivers known to arise from complex rearrangements (n = 21), namely EWSR1::FLI fusion genes, MYCN amplifications, and PAX3/7::FOXO1 fusion genes. When these alterations are identified, they are considered to be the main driver events for these cancer types. Consistent with this, these events were the only CGR found in most of these tumors (n = 19 of 21 tumors; Figure 3). Furthermore, even though these CGRs give rise to recurrent driver events, their underlying genomic rearrangements show a large variation among the tumors (Table S1).

Figure 3.

CGRs affect known cancer driver genes and chromosomal alterations

Number of complex events (top) per tumor colored by CGR type and filled with patterns indicating their effect: cancer driver gene (cross) or unfavorable chromosomal alteration (lines). Patients are annotated by whether the tumors carry a clinically relevant driver alteration (circle) or an alteration in TP53 or MDM2 (triangle), as well as whether they experienced a clinical event (progression, relapse or death, red circle). ∗M002AAB has a germline TP53 alteration. Mutation status per tumor (bottom) across selected genes and genomic regions that are relevant in at least one of the included cancer types. Alterations are colored by type: complex (red), SV (green), SNV/indel (purple), or CN alteration (CNA, beige). Alterations of known relevance in that cancer type are highlighted by a black border. See also Figures S1–S10.

All of the SVs underlying the EWSR1::FLI fusion genes have breakpoints in the EWSR1 and FLI1 hotspots, whether they are simple (n = 11) or complex (n = 11). Although some SVs from non-EWS tumors also occur in this region, they do not have breakpoints that map inside the genes and therefore do not have the same functional effect. While the simple reciprocal translocations between EWSR1 and FLI1 only affect these specific genes and result in the balanced t(11; 22), the complex rearrangements can additionally affect other cancer genes and result in different derivative chromosomes (Table S1).

The structure of the MYCN amplicons differs across NBL tumors, sometimes involving additional enhancers or cancer genes in the resulting ecDNA (Table S2). For example, some tumors display amplification of a broad consecutive region likely including the downstream e4 enhancer as it extends beyond 16.4 Mbp.⁴ In contrast, other ecDNAs comprise multiple separate regions, such as upstream loci or full cancer genes. For example, we identified co-amplification of MYCN and ALK and of FBXO8 and CEP44 (Table S2), both of which have been associated with poor outcomes.^4,21

For the five fusion-positive RMS tumors, we found profound differences in the underlying genomic alterations of their driver fusion genes (Figure S2). Despite recurrent breakpoints in specific locations of the PAX3/7 and FOXO1 genes, the underlying structural rearrangements are different (Table S1). Two tumors carry PAX7::FOXO1 fusion genes (M157AAB and M947AAA), and in both cases, we identified ecDNA/amplicon-type CGRs resulting in amplification of the driver fusion. In addition, tumor M157AAB carries an MYCN amplification, and SV patterns indicate that it is likely part of the same ecDNA as its PAX7::FOXO1 fusion gene (Table S2). In contrast, for two of three tumors with PAX3::FOXO1, the underlying rearrangements are unbalanced reciprocal translocations, and for one, it originated through chromoplexy. Fluorescence in situ hybridization using FOXO1 break-apart probes supported the fact that M157AAB and M947AAA have an amplification (>7 copies), and all five fusion-positive RMS have different underlying rearrangements (Figure S3). Furthermore, for the patients in whom we identified PAX fusions arising from ecDNA in the primary tumor, we verified the presence of the same fusion breakpoints at relapse or in organoid culture (Figure S4),²² indicating that they are maintained during tumor evolution. Overall, the individual differences we observed in driver alterations between tumors provide opportunities for applying precision medicine approaches when treating high-risk cancers.

Most CGRs are advantageous for the tumor

In addition to drivers known to arise from complex rearrangements, we identified candidate CGRs in 21 tumors to investigate further (Table S1). This includes CGRs affecting driver genes usually mutated by SNVs/indels, as well as complex rearrangements resulting in gains or losses relevant to the cancer type. A subset of these candidate CGRs has breakpoints in the CGR hotspots on chromosomes 1, 11, and 12, and they reflect cancer-type specific recurrent events. No other driver alterations have been identified in six tumors carrying candidate CGRs (Figure 3; Table S4), increasing the likelihood that these CGRs are pathogenic. The variant allele fractions and CNs indicate that many CGRs are clonally present (Tables S1 and S5). In tumors with known complex drivers, we observed few additional CGRs. Similarly, in 10 of 21 tumors with a candidate CGR, this CGR was the only complex rearrangement that was identified.

Loss of the tumor suppressor gene WT1 is an important cancer driver event in WT.²³ In three WTs, we identified chromoplexy breakpoints that reside inside WT1 and result in rearranged chromosomes and substantial disruption of the gene (Figure S5). The CGR was fully CN balanced for tumor M062AAB, making it impossible to detect with exome sequencing or targeted assays. For the remaining two tumors, the CGR is accompanied by focal deletions at the breakpoints, but the resulting derivative chromosomes go undetected without the use of WGS. Furthermore, these chromoplexy breakpoints are located in the CGR hotspot on chr11, which also harbors breakpoints from simple SVs affecting WT1 in two other WTs (Table S3). In all five of these WTs, the SVs are clonally present, and the tumors do not carry SNVs or indels in WT1 (Figure 3). This further stresses the importance of analyzing SVs in WTs to detect all pathogenic events.

Focal chromothripsis resulting in the loss of chr1p was detected in three of the seven HBLs. All of them have breakpoints located in a hotspot on chr1 and loss of the adjacent region of chr1p (Figure 4; Table S5), indicating that this focal chromothripsis represents a recurrent event in HBL. We identified a shared lost region on chr1p (chr1:1-34,847,815), and in two tumors, the CGR also resulted in either chr1q or chr2q gain. All three of these chromosomal alterations have been associated with tumor aggressiveness and/or poor prognosis.^18,24 The remaining fourth CGR that we identified in an HBL tumor also results in 1p loss, 1q gain, and 2q gain (Figure S6). Moreover, these instances of CGRs are the only complex events in these tumors, and SV/CN patterns indicate that they are likely to be clonally present (Tables S1 and S5). In conclusion, we identified four complex events in HBLs that result in unfavorable chromosomal alterations, indicating that they may provide proliferative advantages.

Figure 4.

Focal chromothripsis in HBL

Recurrent focal chromothripsis was identified in three HBLs (from left to right): tumors from patients M103AAA, M651AAB, and M333AAB. All three examples have breakpoints in the CGR hotspot on chromosome 1 (chr1:33517560-35493723) (Table S3) and a recurrently lost region on chr1p: chr1:1-34847815 (Table S5). The circos plots contain the SV breakpoints (links) of the CGR and the gains (red) and losses (blue) of genomic material of the affected chromosomes.

In five NBLs, we identified CGRs resulting in chromosomal alterations that have been associated with poor outcome, such as chr1p loss, chr11q loss, and chr17q gain^3,25,26 (Figure S7). These candidate CGRs are not recurrent at the breakpoint level but result in recurrent CN changes (Figure 3). In tumor M050AAB, we identified chromothripsis of chr17 present as an additional derivative chromosome, effectively leading to chr17q gain. The other four candidate CGRs consist of multiple unbalanced translocations that result in one or more unfavorable chromosomal alterations at the same time (Figure S7). For example, tumor M263AAB carries a CGR connecting multiple unbalanced translocations that result in chr10q loss, chr11q loss, and chr17q gain. This shows that single instances of CGRs can have a large impact on the tumor genome and that markers of poor prognosis can be interdependent and physically linked.

Amplification of oncogenes via ecDNA can be a potent cancer driver event as illustrated by MYCN amplification and the PAX7::FOXO1 fusion genes. Furthermore, we identified potentially pathogenic ecDNA/amplicon-type CGRs in seven fusion-negative RMS tumors, resulting in the amplification of MDM2, MYCL, IGF1R1, and YAP1 (Figure S8). On chr12, we identified a hotspot overlapping with the MDM2 oncogene that consists of ecDNA-like breakpoints from three RMS tumors and likely reflects a pathogenic MDM2 amplification (Figure S8). This includes breakpoints from MYOD1-mutated tumor M365AAD, where we observed a pattern of interconnected focal gains that did not meet the amplicon threshold but nevertheless suggests the presence of multiple subclonal amplifications (Figure S9). Of special interest is the ecDNA identified in tumor M911AAA resulting in the rare fusion gene PAX3::WWRT1 (Figure S10), which was detected in both the primary and relapse samples (Figure S4). Previously, we found that the transcription profile of this particular tumor clustered with canonical fusion-positive RMS tumors, which form a high-risk subgroup.²⁷ Moreover, for three of these seven fusion-negative RMS with ecDNA/amplicons resulting in oncogene amplification, we found no alterations in cancer driver genes outside of these complex events (Figure 3).

In conclusion, for 42 of 56 tumors carrying a CGR, we identified potentially pathogenic candidate CGRs affecting cancer-type specific driver genes or unfavorable chromosomal alterations. Furthermore, twice as many patients whose tumors have a CGR affecting a driver gene experienced an adverse clinical event such as progression, relapse, or death, as compared to those without such a CGR (41% vs. 19%, p < 0.05, odds ratio 2.8; Figure 3). This cohort-wide observation is consistent with earlier examples from specific cancer types and CGR types that tumors carrying CGRs increase the risk of poor patient outcomes.^4,15,28,29 Our results show that CGRs tend to be pathogenic events that play an important role in tumorigenesis and are likely indicators of tumor aggressiveness.

Discussion

SVs are important drivers of pediatric cancer, but the prevalence and biological relevance of complex rearrangements has remained elusive. We systematically characterized patterns of extrachromosomal DNA, chromoplexy, and chromothripsis in a cohort of 120 primary tumors across 5 types of pediatric solid tumors. CGRs are prevalent, with 47% (n = 56) of tumors having at least one complex event and instances of the different CGR types occurring in multiple cancer types. Hotspot regions that are recurrently altered by CGR breakpoints often overlap with cancer genes and tend to contain breakpoints from the same cancer type, suggesting that selection pressures play a role in shaping the structural variation landscape. Overall, we identified candidate pathogenic CGRs in 35% of all tumors that affect genes or chromosomal alterations known to be relevant in these cancer types, half of which were known drivers with underlying complex rearrangements (e.g., EWSR1::FL1 fusion genes, MYCN amplifications). In addition, we identified CGRs affecting genes or regions previously associated with poor outcomes, such as MDM2 amplification in fusion-negative RMS due to ecDNA/amplicons and chr1p loss in HBLs due to focal chromothripsis. In conclusion, CGRs are prevalent and highly pathogenic in pediatric solid tumors.

Despite the relatively low mutation burden of pediatric cancer, CGRs occurred in many tumors, including in those with few CN alterations and SVs. Also, many tumors (n = 31) carried only a single CGR, and many CGRs appeared to be clonally present. These observations suggest that the formation of CGRs in pediatric cancers is often a one-time event occurring early in tumor evolution and in absence of (ongoing) genome-wide instability. In contrast, adult cancers have higher mutation burdens and are more heavily altered by both simple and complex rearrangements, with chromothripsis occurring in ∼30% of tumors and chromoplexy in ∼18%.^5,30 Disruption of TP53 has been linked to a higher prevalence of chromothripsis in both pediatric and adult cancers, but it is not a prerequisite for complex rearrangements to occur.^17,30 This further supports the notion that complex SVs do not reflect generic genomic instability but rather reflect a single event with specific mutational mechanisms. The comparison of CGRs between different types of pediatric and adult cancers has been hindered by inconsistencies in criteria for (specific) CGRs patterns and requires striking a balance between consistent definitions and tailored detection methods. Since CGRs were first discovered in adult cancers, criteria for their detection and classification were made with high background mutation rates in mind.¹¹ To accommodate observations of complex events in germline genomes, some studies utilized relaxed criteria, which makes it difficult to compare outcomes and can result in obfuscation of CGR types.⁷ Although dedicated detection tools such as ShatterSeek³⁰ can provide uniformity, we found that this statistical approach was less suited to our dataset of pediatric cancer genomes with very few breakpoints. Alternatively, previous studies in NBLs and pediatric brain tumors used de novo discovery of SV signatures based on the decomposition of SV and CN patterns.^31,32 Although these approaches are powerful, they require larger uniform datasets, and the resulting signatures can be difficult to interpret and compare between studies. Therefore, we identified CGRs agnostic to type by detecting clusters of SV breakpoints and performed categorization into the different CGR types afterward. This allowed for a systematic investigation of CGRs across pediatric cancer genomes.

Among pediatric solid tumors, we observed distinct rearrangement patterns likely arising from an interplay between similar mutational mechanisms and different selection pressures. Although differences in three-dimensional genome structure can also contribute to the observed cancer-type specific patterns,²⁰ the fact that often only a single CGR occurs in a pediatric solid tumor genome and that it regularly affects known driver genes suggests that both negative and positive selection contribute to the observed complex structural variation landscape. Also in pediatric brain tumors, complex SVs have been suggested as cancer drivers and found to occur as single events.³¹ However, while we observed similarities in CGR types among solid cancer types, Yang and Yang identified (complex) SV signatures that are unique to specific brain cancer types and may reflect differences in drivers of genomic instability.³¹ This potentially contradictory finding regarding mutational mechanisms in pediatric brain and solid tumors can arise from how (complex) SV patterns are identified in both studies. While we focused on three previously established CGR types, Yang and Yang decomposed SV patterns into a large collection of signatures, and from this, inferred differences across cancer types.³¹ More generally, both studies support that complex rearrangements are a way to acquire multiple alterations at once, potentially affecting multiple cancer genes and chromosomal gains/losses. We hypothesized that this gives tumors a selective advantage, in line with the findings in pediatric brain tumors that positive selection explains the prevalence of the observed SV patterns despite the low activity of mutational processes.³¹

For EWS and NBL, it has been established that complex rearrangements play a role as cancer drivers.^1,3,4 In EWS, we identified CGRs underlying half of the driver fusions but detected no chromothripsis or ecDNA, which is in line with previous studies reporting a 42% and a 4% prevalence of chromoplexy and chromothripsis, respectively.^28,33 In NBL, we can distinguish two subgroups with different CGR patterns and a third group largely devoid of CGRs: (1) MYCN-amplified (MNA) tumors; (2) tumors carrying chromothripsis, chromoplexy, and related events; and (3) hyperdiploid NBLs, which are generally classified as lower risk.²⁵ For all tumors with MYCN amplification due to ecDNA, it was the only complex event present in the tumor genome, stressing its importance as a tumor driver. Moreover, we found differences in amplicon structure such as inclusion of enhancers or co-amplification of additional cancer genes, some of which previously have been associated with poor outcomes.^4,21 The second subgroup of tumors represent high-risk non-MNA tumors that carry CGRs resulting in derivative chromosomes. Our observations fit the three mutational scenarios recently proposed by Rodriguez-Fos et al., namely (1) reactive oxygen species and replication stress, (2) homologous recombination repair deficiency, and (3) chromosome missegregation.³² However, to our knowledge, we are the first to report CGRs in NBLs resulting in multiple unfavorable chromosomal alterations such as chr1p loss, chr11p loss, and chr17q gain.^3,25 Considering the physical linkage of chromosomal alterations could provide additional insights and ultimately improve clinical decision-making.

ecDNA has been associated with oncogene amplifications and poor outcomes in multiple cancer types.^4,13,15 The impact of ecDNA on tumor biology is profound, with its structure leading to severe dysregulation of genes and non-linear inheritance contributing to intratumor heterogeneity.¹⁴ The persistence of ecDNAs without centromeres is likely due to positive selection, as indicated by computational models,^14,34 cancer-type-specific differences in what genes are amplified,^14,15 and changes in the abundance of ecDNAs as response to drug treatment.¹⁶ Although oncogene amplifications have been reported in RMS,^2,35 the prevalence of ecDNA/amplicon-type complex events in RMS was unanticipated. For three patients in whom we identified PAX fusions arising from ecDNA in the primary tumor, we verified the presence of the same ecDNA breakpoints at relapse (M911AAA and M947AAA) or in organoid culture (M157AAB) (Figure S4).²² This indicates that the ecDNAs are likely selected for, since they are maintained throughout tumor evolution. Furthermore, the M947AAA tumor acquired MYCN amplification during relapse, which is linked to its PAX7::FOXO1 fusion in a single ecDNA/amplicon-type CGR. This resulted in a similar construct as present in tumor samples from M157AAB. Complex ecDNA structures like this can arise from clustering together of separate ecDNAs in hubs, followed by recombination into a single structure.¹⁴ Although MYCN amplification is common in fusion-positive RMS,³⁶ we demonstrate for the first time co-amplification of MYCN and PAX7::FOXO1 on a single amplicon structure. Since we identified ecDNAs containing PAX7, FOXO1, and MYCN in tumors from two different patients, it does not seem to be an incidental finding but rather reflects the difficulties in detecting these events without WGS. Furthermore, there have been conflicting reports regarding the association of MYCN amplification with outcomes in fusion-positive RMS³⁶; however, the underlying rearrangements were not considered in these studies. In addition to driver fusions, we identified CGRs underlying oncogene amplification in fusion-negative RMS (e.g., MDM2, YAP1, IGF1R), which could provide leads for targeted therapies.³⁵ Moreover, previous studies found that MDM2 was among the oncogenes most often amplified as circular ecDNA across cancer types.^15,37 Not only did circular amplicons achieve higher CNs but also the oncogene expression was higher corrected for CN due to increased chromatin accessibility and possible enhancer rewiring.¹⁵ Since short-read sequencing is limited in its ability to fully resolve amplicon structures, this highlights a need for additional technologies such as long-read sequencing or circularization for in vitro reporting of cleavage effects by sequencing.³⁸ Even then, ecDNAs can be heterogeneous within a tumor cell population, and they are able to transition between extrachromosomal states and integrate into chromosomes.^14,16 This flexible nature also contributes to their pathogenicity. Detection of these highly amplified CGRs is a first step toward assessing their clinical relevance.

By considering complex rearrangements, we identified potentially pathogenic variants in WT and HBL. For example, in three WTs lacking a driver SNV, chromoplexy breakpoints disrupted WT1. WTs are usually regarded as genomically quiet since they carry few recurrent genetic alterations,²³ and studies looking specifically into CGR patterns are lacking. However, we found that WT genomes can contain “hidden” impactful variation, such as novel derivative chromosomes, that can be uncovered and studied using WGS. Similarly, HBLs carry few mutations and have a paucity of known genetic drivers outside of CTNNB1-activating mutations.³⁹ However, four of the seven tumors recurrently altered chromosome 1p via a CGR, three of which were classified as focal chromothripsis. These findings contrast with previous studies that found no chromothripsis in primary HBLs and reported chromoplexy in only ∼10% and ecDNA/double minutes in ∼5% of tumors.³⁹ To summarize, we identified both known examples of complex cancer driver events, as well as additional likely pathogenic CGRs in cancer types where it was not anticipated. Since we also observed that patients experienced a clinical event twice as often when their tumor carried a CGR affecting a cancer driver gene, the clinical implications of CGRs warrant further study in a larger cohort.

Limitations of the study

We have focused on the three most established classes of CGRs: chromothripsis, chromoplexy, and ecDNA. Nevertheless, existing bioinformatic tools were not suitable for detecting these methods either due to the overall low mutation rate in pediatric cancer or the small sample size. Within the cohort, we detected a large number of CGRs not meeting the criteria of chromothripsis, chromoplexy, and ecDNA that were placed in the “complex other” class. While we established that CGRs frequently occur in pediatric solid tumors, further research is required to catalog all CGR mechanisms. In addition, we report a higher rate of clinical events in patients with a CGR compared to those without; this observation requires uniform clinical follow-up data in a large independent dataset to confirm its clinical relevance.

Conclusions

CGRs commonly occur in pediatric solid tumors, and our findings highlight the importance of analyzing them as candidate pathogenic events. Not only do CGRs give rise to known driver alterations such as fusion genes and oncogene amplifications but they can also affect cancer genes usually altered by SNVs/indels or result in relevant CN gains and losses. Interpreting CGRs as single events that affect genome structure in multiple ways enables more refined functional analysis of genetic alterations. Moreover, CN-balanced CGRs are largely an unexplored type of SV and yield promising new candidates in tumors without known driver alterations. Finally, carrying complex rearrangements has previously been associated with tumor aggressiveness and poor outcomes in specific cancer types. Despite the limited size of our cohort, we show that CGRs are associated with a higher incidence of clinical events agnostic to the type of tumor. CGRs affect a significant proportion of pediatric solid tumors, and our findings warrant further research into their role in cancer etiology and progression.

Resource availability

Lead contact

Further information and requests should be directed to and will be fulfilled by the lead contact, Jayne Hehir-Kwa (j.y.hehirkwa@prinsesmaximacentrum.nl).

Materials availability

This study did not generate new unique reagents.

Data and code availability

All WGS datasets used in this article are available in the European Genome-Phenome Archive (EGA) repository under study EGAS00001007565. Data from organoid sample PMOBM000ABW is available in EGAD00001008466. Code is available through https://github.com/princessmaximacenter/structuralvariation/ (tag version 2.0), released on September 18, 2023 (https://doi.org/10.5281/zenodo.13379929).

Acknowledgments

We acknowledge the willingness of all children and their families to donate their material and data for our research. We also want to express our gratitude for the contributions from all pediatric oncologists, research nurses, lab technicians, and biobank committee members for obtaining patient consent, sample collection, and processing, as well as general logistics. We acknowledge the fruitful discussions with and feedback from fellow lab members and Ivo Griffioen. We gratefully acknowledge the financial support provided by the Foundation Children Cancer Free (KiKa core funding) and the Adessium Foundation. The funders had no role in the design of the study, as well as the collection, analysis, and interpretation of the data or writing.

Author contributions

Conceptualization: I.A.E.M.v.B., J.Y.H.-K., P.K., F.C.P.H., and B.B.J.T. Supervision: P.K., J.Y.H.-K., F.C.P.H., and B.B.J.T. Funding acquisition: P.K. Methodology: I.A.E.M.v.B., J.Y.H.-K., and P.K. Investigation and visualization: I.A.E.M.v.B. Software: I.A.E.M.v.B., S.B., A.J., E.T.P.V., M.S., S.d.V., J.B.-H., and H.H.D.K. Resources: M.v.T. and N.S.-W. Data curation: I.A.E.M.v.B., S.B., A.J., E.T.P.V., M.S., S.d.V., J.B.-H., and H.H.D.K. Writing – original draft: I.A.E.M.v.B., J.Y.H.-K., and P.K. Writing – review & editing: I.A.E.M.v.B., J.Y.H.-K., P.K., M.T.M., N.S.W., J.D., M.M.v.d.H.-E., J.H.M.M., J.J.M., W.C.P., and B.B.J.T.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Critical commercial assays
KAPA DNA HyperPlus kit	https://rochesequencingstore.com/catalog/kapa-hyper-plus-kit/	07962428001
Vysis LSI FOXO1 (13q14) Dual Color	https://www.molecular.abbott	Break Apart Rearrangement Probe 30-231023
Deposited data
Raw WGS tumor and normal samples	EGA	EGAS00001007565
Organoid sample PMOBM000ABW	EGA	EGAD00001008466
Software and algorithms
GATK 4	https://github.com/broadinstitute/gatk	4.1.7.0
Manta	https://github.com/Illumina/manta	1.6
Delly	https://github.com/dellytools/delly	0.8.1
GRIDSS	https://github.com/PapenfussLab/gridss	2.7.2
SV analysis code	https://github.com/princessmaximacenter/structuralvariation/	https://doi.org/10.5281/zenodo.13379929
Other
Human reference genome UCSC hg38	https://console.cloud.google.com/storage/browser/genomics-public-data/resources/broad/hg38/v0/	Hg38
Population variants	NCBI	nstd166, nstd186
Population variants	DGV	version 2020-02-25
Repeat annotation	UCSC	Accessed 2021-04-20

Experimental model and study participant details

To characterize complex structural variation patterns across pediatric solid tumors, we selected patients diagnosed with Ewing sarcoma, neuroblastoma, rhabdomyosarcoma, Wilms tumor and hepatoblastoma for which primary tumor material was included in the Máxima biobank, subject to informed consent.⁴⁰ Informed consent has been obtained for all subjects involved in this study through the Máxima biobank informed consent procedure approved by the Medical Ethics Committee of the Erasmus Medical Center in Rotterdam, The Netherlands, under reference number MEC-2016-739. Approval for use of the subject’s data within the context of this study has been granted by the Máxima biobank and data access committee, biobank request nr. PMCLAB2018.017. Patients were eligible when whole-genome sequencing (WGS) data of sufficient quality was available from matching tumor-normal samples taken within 150 days of diagnosis and the variant calling steps were successfully completed.

Method details

Sequencing library preparation and data pre-processing, including alignment and quality control, was done via the institute’s standardized pipelines and guidelines as described before.^40,41,42 In summary, DNA was isolated from fresh frozen tumor tissue and as a matching normal, DNA was isolated from whole blood. Whole-genome sequencing (WGS) libraries were generated from 150 ng DNA using the KAPA DNA HyperPlus kit and sequenced on a NovaSeq 6000 platform (Illumina). High quality WGS samples were selected requiring a minimum median coverage of 27x for normal samples and 81x for tumor samples. We also included two samples with lower coverage of the tumor (PMABM000DEP with 60x, PMABM000DIX with 80x) that have been successfully analyzed previously.⁴¹ In total, for 120 patients WGS data was available with a median coverage of 37x for the normal and 107x for the tumor samples (Table S6).

Fluorescence in situ hybridization

Fluorescence in situ hybridization analysis was performed using Vysis LSI FOXO1 (13q14) Dual Color, Break Apart Rearrangement Probe 30–231023.

Single nucleotide variants and indels

Somatic single nucleotide variants (SNVs) and indels were identified using Mutect2 from GATK 4.1⁴³ and annotated by variant effect predictor (VEP) (version 104).⁴⁴ First, we filtered high-confidence variants that have tumor allele fraction >0.1 and are located on chromosomes 1–22 and X, excluding ENCODE Blacklist poor mappability/high complexity regions.⁴⁵ Second, to select likely pathogenic variants, we filtered on VEP impact moderate or high and removed variants predicted as benign/tolerated by PolyPhen/SIFT unless they were present in COSMIC.⁴⁶

The tumor mutation burden was defined as the number of nonsynonymous somatic SNVs and indels per megabase. This encompasses the SNVs/indels in protein coding genes on chromosomes 1–22 that also passed the previous filtering steps. For the denominator, we used ∼41 megabase pairs (Mbp) corresponding to the number of coding sequence bases in protein coding genes.

Copy number alterations

Copy number (CN) alteration data was generated by the GATK4 pipeline following their recommended best practices.⁴⁷ Across all analyses, we distinguished four CN call states: gain, loss, loss of heterozygosity (LOH) and neutral (no change). For gain and loss we required at least +/− 0.2 copy ratio log2 fold change (cr l2fc), and for LOH less than 0.4 minor allele fraction (MAF) with absence of gain/loss. The remainder was regarded as neutral. As a proxy for assessing the CN stability across a genomic region, we used the percentage of sequence within 33% of the mean CN for gain/loss and between −0.1 and 0.1 cr l2fc for neutral. Genomic regions with at least 70% of sequence near the target value were regarded as stable.

The fraction of the genome altered by copy number alterations (FGA) was calculated as the number of bases in a gain or loss state, divided by the total number of bases. Hereby excluding the difficult to assess centromeric (acen), variable-length (gvar) and tightly-constricted (stalk) regions. Likewise, the ploidy was derived from the weighted mean of the copy ratio of autosomal chromosome arms.

Structural variants

Somatic structural variants (SVs) were detected with Manta (version 1.6),⁴⁸ DELLY(version 0.8.1)⁴⁹ and GRIDSS (version 2.7.2).⁵⁰ First, we filtered SVs with a minimum of seven supporting reads and removed those with >90% reciprocal overlap with common (>1%) population variants retrieved from the NCBI repository (nstd166,⁵¹ nstd186 [https://www.ncbi.nlm.nih.gov/dbvar/studies/nstd186/]) and from DGV (version 2020-02-25)⁵² accessed on 2021-03-11. Second, we merged SVs called by the three tools based on 50% reciprocal overlap and required detection by at least two tools. Third, we filtered on tumor allele fraction >0.1 for all downstream analyses, except for dedicated analyses into the SV patterns of highly amplified regions (see below) since the allele fractions (AFs) of these SVs can be artificially low due to presence of many reference reads.

To identify whether SVs are likely repeat-mediated, we annotated them with repeats retrieved from UCSC table browser accessed on 2021-04-20.⁵³ First, repeats were filtered by completeness (<50 base pairs (bp) of repeats left) and repeat class (LINE, SINE, LTR, Simple_repeat, Low_complexity, Retroposon) to prevent spurious annotations. To be considered repeat-mediated, we required both the start and end breakpoints within 100 bp of repeat elements of the same class.

CN changes associated with interchromosomal breakpoints

To analyze whether interchromosomal breakpoints (CTX) are associated with chromosome level CN changes, we categorize CTXs as unbalanced, balanced or inconclusive. Hereto, we compared CN states before and after the breakpoint using windows that 1) extend across the full chromosome (arm) and 2) encompass the direct vicinity (5 Mbp flanking regions).

First, we defined windows relative to the chromosome and inferred their CN state and stability (see STAR Methods on CN alterations). Chromosome level windows extend from the 5′ telomere to the bp ("before") and from the bp to the 3′ telomere ("after”). We required both before and after windows to be CN stable (>66%) and sufficiently large to assess (>10 Mbp), otherwise they were regarded as "inconclusive". Second, we compared the CN of the before and after windows using a difference in CN call state or >0.2 cr l2fc as criterion for “unbalanced”. Third, we performed this analysis with the 5 Mbp before and after the CTX breakpoint to verify the chromosome level observations. For the final categorization of the CTX into unbalanced or balanced, we required the local CN changes to either match the chromosome level observation or be "inconclusive", as such allowing for small CN changes around CTX breakpoints that are commonly observed. If the CTX could not be categorized on the chromosome level, chromosome arm-level windows were considered. These are defined similarly but using the nearest centromere instead of both telomeres, and >5 Mbp minimum size. Finally, CTXs within 5 Mbp distance of the telomeres were labeled as 'edge' breakpoints.

In downstream analyses, we used the location of the breakpoint to the nearest telomere as the genomic interval of the CN change of unbalanced CTXs.

Chromosome (arm) level CN alterations

To assess chromosome level CN alterations, we considered the most prevalent CN call state, the CN stability and the presence of unbalanced SVs. CN call states for chromosomes were identified based on the highest fraction of sequence. This allows for identification of numerical CN changes in the presence of focal CN changes, for example if a chromosome has 85% gain and 15% loss due to a large focal deletion, then the fraction in the “gain” call state will be selected as the chromosome-level CN state and further assessed for stability. As criteria for a chromosome level CN alteration, we required a >70% stable sequence of the gain, loss or LOH call state, and absence of an explanatory unbalanced translocation.

For chromosome-arm level alterations, the same approach was used but here we excluded the acen/gvar/stalk regions. In case CN alterations were identified on the full chromosome level, these take precedence over chromosome arm level alterations.

Amplicons

Focal amplifications (amplicons) can arise through different mechanisms and exist in multiple forms, such as circular extrachromosomal DNA (ecDNA) constructs and linear homogeneous staining regions. ecDNAs tend to be smaller with an expected size between 10 kilobase pairs (kbp) and 10 megabase pairs (Mbp) and more highly amplified (>8 copies) than local rearrangements.⁵⁴ To identify amplicons, we selected regions with high CN (>1.9 cr l2fc), corresponding to ∼7.5 copies to account for lower tumor cell fractions. We also applied the same threshold relative to the mean CN of the chromosome to account for the presence of chromosomal gains. For example, >2.4 cr l2fc would be the required CN in a 3n chromosome with mean 0.5 cr l2fc. Next, adjacent CN segments within 6 kbp were merged and we selected amplicons >50 kbp for downstream analysis.

Next, we analyzed whether separate amplicons are connected to each other by SVs (Table S2), which could indicate co-amplification on ecDNA. Hereto, amplicons were annotated by overlapping CGRs (see next section) and we looked into whether low allele fraction SVs connect separate amplicons. For these SVs, the allele fractions are likely artificially low due to presence of many reference reads. Furthermore, we annotated amplicons with overlapping CGRs and analyzed whether SVs have both breakpoints inside the amplicon (no loose ends).

Identification of complex genomic rearrangements

Complex genomic rearrangements (CGRs) are characterized by clusters of breakpoints reflecting the repair of multiple simultaneous dsDNA breaks. To identify clusters of SVs, we used a graph-based approach considering SVs as vertices and drawing edges between SVs that have breakpoints within 5 Mbp. Connections between chromosomes can be formed by interchromosomal breakpoints which are represented as two vertices with an edge. From each tumor’s graph, we identified SV clusters by extracting the connected components, and subsequently categorized them into different types of (complex) genomic rearrangements based on a combination of SV and CN features and manual inspection.

After identifying clusters of SVs, they were categorized into different type of events: ecDNA, pair of SVs, reciprocal translocation, chromoplexy, chromothripsis and 'complex other'. See below the criteria used for each CGR type, and also the order of categorization is of importance.

• 1)ecDNA/amplicon: SV breakpoint within 1 kbp of an amplicon.

We first categorized this CGR type to distinguish potential ecDNA from chromoanagenesis events. Although we could not definitively establish with short-read WGS whether the CGRs represent circular ecDNA constructs, we selected the amplicon criteria to optimize for this¹⁵ and detected closed chains in graphs for many clusters as expected,⁵⁴ but we did not require this as criterium given that SV calling is challenging in highly amplified regions. Instead, we analyzed whether SVs had both breakpoints in an amplicon and observed that there were little to no “loose ends” for most ecDNA/amplicon clusters.

• 2)(non-complex) SV pair: two nearby SVs or one CTX consisting of two breakpoints and a nearby SV.

Although we do not regard two nearby SVs as CGRs, also small clusters of two breakpoints were extracted to allow for detection of chained interchromosomal rearrangements having just two breakpoints on a chromosome.

• 3)(non-complex) Reciprocal translocation: four CTXs connecting two chromosomes, with <2 Mbp distance between the CTX breakpoints indicating a simple rearrangement. Although the presence of small intrachromosomal SVs is allowed, the SVs should form a closed chain and footprints remain <20 Mbp.
• 4)Chromoplexy: a closed chain of SVs connecting two or more chromosomes. No unbalanced CTXs are allowed and all chromosomes have at least one balanced or edge CTX.
• 5)Chromothripsis: high density of SVs, namely at least 10 overlapping SVs or 10 CTXs between a chromosome pair. As a proxy for CN oscillations, we merged all adjacent CN segments with the same call state, and required at least 10 CN state switches within a footprint. Finally, the distribution of SV types should not be significantly different from an equal distribution (chi-squared test p-value >0.05) indicating randomness of fragment joining.
• 6)Complex other: SV clusters not satisfying the criteria outlined above. After manual inspection subset of 12 ‘other’ CGRs were manually assigned to one of the other types (‘complex_sv_class_manual_override’, Table S1).

CN alterations due to CGRs

To associate chromosomal alterations with CGRs, we combined overlap with CN segments and unbalanced translocations. First, we inferred focal CNs relative to the chromosome average and merge CN segments with the same state, allowing for up to 5 Mbp gaps to ignore interruption by small fluctuations. Then, these merged CNs were matched to CGRs if they overlapped at least 5 Mbp, or at least half their size for smaller segments. Second, we added the CN changes due to the unbalanced translocations part of the SV cluster, as defined previously. This allowed us to conservatively infer a set of CN changes likely due to a CGR, and ignore potential co-occurring chromosomal changes such as aneuploidies.

Alterations affecting cancer driver genes and chromosomes

To assess whether a gene is altered, we combined SNV/indels, CNs and SVs and filtered alterations that could affect gene function. The alteration types reported in Table S4 and Figure 3 are based on the following.

• •SNV: SNV or indel predicted by VEP to have moderate or high impact
• •SV: breakpoint inside the gene body, or breakpoint within 1 Mbp of the gene body together with a CN change
• •Complex: the SV breakpoint is part of a CGR
• •CNA: high-impact CN change by itself namely homozygous loss (−1 cr l2fc) or amplification (>1.9 cr l2fc)

Chromosomal alteration types reported in Figure 3.

• •Complex: CN alterations due to CGRs
• •SV: CN changes associated with interchromosomal breakpoints
• •CNA: Chromosome (arm) level CN alterations

Only genomic regions >5 Mbp are included in the figure and downstream analyses.

To identify alterations relevant for tumor biology, we focused on pediatric cancer genes (Table S7) and more specifically on previously established cancer-type specific associations with genes and chromosomal alterations. The following definitions were used: Cancer driver genes: genes in which alterations confer a selective advantage to the tumor (Table S8). Unfavorable chromosomal alterations: chromosomal alterations that have been associated with poor outcome or high risk (Table S9).

CGRs were annotated with their ‘effect’ based on whether they affect cancer driver genes (driver) or result in unfavorable chromosomal alterations (chrom_alt).

Quantification and statistical analysis

Genomic regions recurrently affected by SVs

Recurrently altered regions were identified using a peak-calling approach based on SV breakpoints. For a conservative measure of recurrence, we counted the number of distinct tumors with a breakpoint in the region, such as to not bias for certain CGR types that inherently have many bps. First, we overlapped all SV breakpoints using 1 Mbp flanking windows and partitioned the genome into regions based on the number of distinct tumors. Next, we mapped each breakpoint to the highest hotspot it overlaps with, having the most distinct tumors. In case of a tie, we selected the region with the highest fraction of overlap with the breakpoint. After this mapping, we re-assessed the number of distinct tumors contributing to each region. Finally, we annotated the genomic regions with cancer genes and assessed the fraction of SVs from each cancer type and CGR type.

CGR hotspots were defined as recurrently altered regions with CGR breakpoints from three or more tumors.

Statistical tests

For group-level comparison of the tumor mutation burden and fraction of genome altered, we used Wilcoxon rank-sum test (two-sided) to compare distinct samples from tumors with and without mutations in TP53 or MDM2, respectively N = 10 and N = 110.

To assess enrichment of CGR breakpoints in repetitive regions, we used the Fisher’s exact test to compare the number of CGR breakpoints mediated by repeats (N = 101) and the number of simple SV breakpoints mediated by repeats (N = 369) with the number of breakpoints that are not mediated by repeats and represent either a CGR (N = 1225) or simple SV (N = 1017). A breakpoint can only be assigned to one of these four categories. In this analysis, 2712 distinct breakpoints across 120 distinct samples were used. The same approach was used to assess enrichment of CGR breakpoints inside and nearby pediatric cancer genes. This was also repeated for CGR breakpoints split per complex SV class. See Table S10 with the breakdown of N number of breakpoints for all the above enrichment analyses. Similarly, to assess enrichment of CGR hotspots for repetitive regions or pediatric cancer genes, Fisher’s exact test was used to compare the SVs in CGR hotspots to SVs in the rest of the genome using the 2712 distinct breakpoints across 120 distinct tumor samples (Table S10).

Finally, Fisher’s exact test was used to compare the prevalence of clinical events in patients that have tumors with or without CGRs affecting driver genes. This involved distinct tumor samples (N = 120), each from one patient. N = 13 of 32 tumors with CGRs affecting driver genes had at least clinical event, compared to N = 17 of 88 tumors without a CGR.

Published: October 14, 2024

Supplemental information

Table S1. CGRs, related to Figures 1 and 3

CGRs annotated with characteristics supporting their categorization as a certain CGR type and to assess their pathogenicity, e.g., affected cancer driver genes and summary of CN changes.

Table S2. Amplicons, related to Figures 1 and 3

Amplicons annotated with genes and CGRs for assessment of physical linkage.

Table S3. Genome-wide SV hotspots, related to Figure 2

Genomic regions recurrently altered by SVs in at least three tumors, annotated with cancer genes. Complex hotspots have CGR breakpoints from at least three tumors.

Table S4. Alterations in cancer driver genes, related to Figure 3

Overview of SNVs, indels, CNs and SVs identified in cancer driver genes (see Table S8).

Table S5. CN changes due to CGRs, related to Figure 3 and Table S1

CN changes associated with CGRs, both overlapping local changes and unbalanced translocations.

Table S6. Complete cohort overview, related to Figure 1 and 3

Patient information related to diagnosis, and if applicable treatment and clinical events, as well as identifiers of the WGS data and quality assessment.