Modeling cancer rearrangement landscapes

Abstract

Cancer genome sequences contain footprints of somatic mutational processes, whose analysis in large tumor sequencing datasets has revealed novel mutational signatures, correlative features of variant topography, and complex events. Many of these analytic results have yet to reconciled with decades of mechanistic genome integrity research performed in controlled model systems. However, a new generation of genome-integrity experiments combining computational modeling, data analytics, and high-throughput sequencing are emerging to link mechanisms to patterns. Conversely, analytic studies evaluating quantitative footprints of specific genome integrity hypotheses will be critical in fitting naturally occurring mutational patterns to the predictions of a particular mechanistic model. Such quantitative and mechanistic studies will form the foundation of an emerging systems biology of genome integrity.

Introduction

Cancer genomic mutational landscapes are shaped by the forces of selection, mutagenesis, and repair. In recent years, massively parallel sequencing has enabled the first unbiased large-scale surveys of somatic genome variation in a variety of cancer types [1]. Though primarily motivated by the search for driver alterations and novel cancer genes, these studies have also revealed (and have been occasionally confounded by) the footprints of known and novel mutational processes [2,3]. These data-driven, statistical analyses provide a fascinating counterpoint to decades of genome integrity research, where rigorous experimentation has dissected the biochemistry and cell biology of DNA damage and repair in model systems such as yeast and mammalian lymphocytes [4].

Though the interaction between cancer genomics and genome integrity fields has been mostly ad hoc, it has already generated several important successes. This includes examples where mechanistic intuition has been statistically validated through cancer genome analyses, and where cancer genome hypotheses have been modeled in experimental systems. Though the most mature area of research has been the study of single nucleotide variants (SNVs), the biggest challenges and opportunities for interactions between computationalists and experimentalists lie in the analysis of DNA rearrangements. We argue that these interactions form the seeds of an emerging field of the systems biology of cancer genome integrity that refines mechanistic models of somatic DNA repair pathways through quantitative analysis of high-throughput sequencing-based signals.

SNV landscapes

The study of single nucleotide variants (SNVs) patterns provides a paradigm for cancer genomics-driven mutational process research [5]. Analytic work in this area has focused on three key directions: mutational signatures, variant topography, and the identification of complex events. As many somatic SNVs are not linked to a known biological mechanism, they provide many fascinating starting points for future experimental investigation.

Somatic SNVs arise under the combined action of exogenous mutagens (e.g. UV light, tobacco carcinogens, industrial chemicals), DNA repair defects (e.g. base and nucleotide excision repair, mismatch repair), endogenous mutational mechanisms (e.g. activation-induced cytidine deaminase), and selection [2]. Cancer genomes bear distinct signatures of these mechanisms, which can be discovered de novo through factorization of SNV count matrices cataloguing 192 stranded or 96 strand-collapsed tri-nucleotide sequence contexts (e.g. TpCpA > T, where C is the mutated base) [6]. Analysis of these signatures has been used to identify footprints of aging [7], acquired DNA repair defects in POLE [8] and ERCC2 [9], and mutagenic exposures (e.g. aristolochic acid) [10].

Correlative analyses of variant topography have linked regional fluctuations in SNV density to replication timing [3,11], post-translational chromatin modification [12,13], transcription factor binding [14–16], and tissue-specific chromatin accessibility patterns [17]. At the forefront of SNV mutational processes research are investigations into the regional fluctuations in specific base-context and strand-specific SNV signatures [8,18]. Such approaches have been used to attribute specific mutational processes (e.g. mismatch-repair deficiency) to replication vs transcription-based mechanisms.

Finally, WGS analyses have been used to identify complex mutational events that generate multiple SNVs in cis, such as kataegis. Kataegis refers to a phenomenon of strand-coordinated clusters of cytosine mutations in the TpCp* context, occurring near somatic rearrangements and affecting specific cancer types (e.g. breast cancer) [19,20]. Kataegis SNV patterns may derive from APOBEC enzyme activity [21–23].

Complex rearrangement landscapes

Cancer DNA rearrangements have been linked to specific mechanisms of DNA breakage and repair, including nonhomologous end joining (NHEJ), microhomology mediated end-joining (MMEJ), DNA replication stress, telomere fusions, and breakage fusion bridge cycles (BFBs) [24]. Though many of these mechanisms have been convincingly established in model systems (e.g. yeast, mammalian B-cells), their relevance to cancer DNA structural variation has not been comprehensively evaluated in large pan-cancer sequencing datasets.

Rearrangement signature analyses have catalogued spectra of structural variant types (deletion, amplification, translocation) and sequence contexts across multiple cancers, however at a much smaller scale than for SNVs. Though many cancer genomes are highly rearranged, the average genome harbors fewer than a few hundred rearrangements, as opposed to tens of thousands of SNVs [25]. In addition, identification of rearrangements requires WGS data, of which only ~3000 cases have been made available to the community (in contrast to the 15,000 exome cases that have been deposited in data portals). The largest published pan-cancer rearrangement signatures studies have analyzed fewer than 200 samples [25–27]. While some groups have observed short stretches of sequence homology around rearrangement junctions, it is unclear whether the observed patterns provide definitive evidence for alternate rearrangement mechanisms (e.g. MMEJ). Much less is known about the genetic basis of hypermutators for rearrangements than for SNVs, though enrichment of specific rearrangement signatures have been noted in a few tumor contexts: Rag recombinase deletion signature in ETV6-RUNX1 fused acute lymphoblastic lymphomas [28], translocation in exceptional platinum-chemotherapy responder pancreatic cancer [29]. Current international analysis efforts are underway to analyze much larger datasets [30].

The topography of cancer rearrangement patterns has also not been as exhaustively explored as for SNVs. Pan-cancer microarray analyses have correlated copy number endpoint locations with late replicating regions and nuclear proximity as determined by 3D chromatin confirmation (Hi-C) methods [31–33]. Similar analyses have not been, to our knowledge, applied to pan-cancer whole genome sequencing data. However, a recent study analyzing rearrangement patterns in over 500 breast cancer whole genomes has suggested that somatic rearrangement junctions are enriched in early, rather than late, replicating regions [34]. Mechanistic studies incorporating engineered translocations in mouse model systems have suggested that rearrangement junctions occur independently of nuclear proximity [35]. Additional pan-cancer analyses of whole genome sequences will be required to resolve these apparent contradictions, and suggest novel topographic correlations.

Despite the paucity of large-scale statistical surveys, some of the most dramatic findings in cancer genomics have emerged from WGS rearrangement analyses. This includes the discovery of chromothripsis (chromosome shattering [36]) and chromoplexy (chains of balanced rearrangements [37]). The concept of “catastrophic” large-scale rearrangement events has provided an attractive explanation for these phenomena [38], thought alternative explanations invoking progressive iteration of simple events have been argued using computational models [39]. Such modeling approaches provide a lens with which to understand the genomic footprints of BFBs, a complex rearrangement phenomenon described long before the sequencing era [40,41]. Modeling becomes especially critical when BFBs or chromothripsis are observed in conjunction with other rearrangement events [42]. It remains to be seen whether the existing “lexicon” of rearrangement events are sufficient to explain the dense webs of rearrangement junctions observed in many cancer sequence data, or whether additional basic event types remain to be discovered [43].

Fundamental analytic challenges

Several conceptual and data-driven challenges complicate the analysis of rearrangement patterns in cancer genomes: As mentioned above, rearrangement data are sparse, comprising much fewer events per sample than for SNVs. Additional complexity is related to the definition of a rearrangement “event” and rearrangement context. Finally, the myopic nature of short-read sequencing (100–200 bp read pairs generated from 300 to 700 bp DNA fragments), makes it difficult to assemble or phase rearranged loci and detect variants in repetitive genomic regions [44].

The atomic unit of cancer rearrangement data is a junction (Figure 1): a pair of genomic locations and strands that were previously distant but were made adjacent through a rearrangement event. There are four orientations that a junction can take with respect to a particular coordinate system, which is most commonly defined with respect to the p and q arm orientation of chromosomes in a reference assembly, however could also be defined with respect to other frames (e.g. replication origin, transcription origin). A rearrangement event may involve a single junction or multiple junctions, e.g. a deletion or duplication involves a single junction, a balanced translocation or inversion involve two junctions, chromothripsis or chromoplexy involve many more [38].

Figure 1.

DNA junctions: The four possible types of rearrangement junctions are illustrated, each consisting of a pair of breakpoint locations and orientations. The orientations are defined using a particular coordinate system, which is normally anchored with respect to the p–q axis of a chromosome, but can be anchored to other landmarks (e.g. replication origins, transcription start sites). In the standard frame, the left direction is the direction of decreasing coordinates (e.g. towards the p arm) and the right direction is in the direction of increasing coordinates (e.g. towards the q arm). Junctions ends that connect to the left side of a breakpoint are denoted as “−” and those that connect to the right side are denoted with a “+”. Red lines indicate rearrangements connecting two previously separate genomic loci.

When multiple junctions map to a given region, determining the events that gave rise to them may not be straightforward. As Figure 2 illustrates, a junction that appears to be a deletion locally (inner junction associated with copy number drop) is more plausibly explained as a duplication when the broader context is considered (outer junction associated with copy number gain). For this reason, the correct inference of rearrangement events from WGS junction patterns requires building models of loci evolving across intermediate and final states. When many junctions map nearby on the genome, this inference may be computationally intensive and fundamentally ambiguous, with many solutions fitting the data equally well. Resolution of this ambiguity may be facilitated with additional data, either snapshots of additional tumor samples in intermediate temporal states of evolution or long-range sequencing data enabling end-to-end locus assembly in space. Technologies providing long-range cis information include single molecule sequencing (e.g. Pacific Biosciences), optical mapping (e.g. BioNano genomics), linked reads (e.g. 10X genomics), and proximity ligation (e.g. Dovetail genomics) [45–47]. The development and application of novel computational modeling approaches [48,49], based partly on phylogenetic techniques employed in comparative genomics, may provide additional inferential power for reconstructing rearranged loci.

Figure 2.

Rearrangement evolution: Junction patterns in cancer are usually the result of the progression of a locus across multiple states of evolution. The derivation of rearrangement events from junction patterns requires modeling their evolution in time and/or their long-range structure in space. The example here depicts the challenge of inferring events from junctions, by depicting two tandem duplication events occurring temporally in sequence (Peter Campbell and Yilong Li, personal communication). The left column shows the evolution of the locus assembly (i.e. string of segments A–E) in somatic coordinates, progressing from its initial state to early and advanced lesions in which the central segments are duplicated. The right column shows the same assembly in reference genome coordinates as a thread of segments joined by rearrangement junctions (purple and red). Simulated read data from rearranged genomes generate coverage patterns on the reference that reflect the copy number of segments A–E. For example, the advanced lesion copy number pattern is consistent with 1, 3, 2, 3, and 1 copies of segments A–E, respectively. Since the purple junction is associated with an intervening copy loss, it would be categorized as a deletion event by most copy number callers. However, given the end-to-end assembly of segments, it is clearly the result of a (late) tandem duplication event.

In the absence of perfect reconstructions of locus evolution or cis-structure, it can still be possible to study sequence signatures or topography with respect to individual junctions (Figure 3). Unlike with SNVs, the sequence and topographic context that must be mined must be gleaned from two or more genomic locations. Instead of simply counting base triplets, sequence context analysis must examine stretches of homology at or near both junction ends, and take into account additional somatic SNVs occurring near the lesion site (e.g. TpC mutations arising in the context of kataegis). This sequence context analysis must be oriented with respect to the junction strands, and also take into account (novel or templated) sequence that may be inserted between the junction ends. Similar consideration must be applied to the analysis of topographic genomic features, such as regional chromatin states or other features (e.g. R-loops) that are gleaned from “*-seq” experiments. A final feature of rearrangement context is the proximity between junction ends, which can be measured using chromatin conformation maps. These analyses can be used to examine relationships between the context (the state of the genome prior to an event) and consequence (the state of the genome following the event) of a somatic junction.

Figure 3.

Junction context: Several notions of junction context are illustrated here, useful for defining and exploring sequence based signatures and regional genomic topographies that might be associated with rearrangement events. Since junctions connect two loci, the immediate and regional sequence context of each breakpoint must be considered (e.g. immediate microhomology in blue, distant homology in red oriented using junction-centric coordinates (where coordinate 0 is the black–gray interface and black denotes positive coordinates (so the coordinates increase away from the junction in two directions). In addition to the reference sequence context, one might consider inserted sequence at the junction sites and the locations of additional somatic mutations (e.g. kataegic SNVs at the arrowheads) occurring in cis to the rearrangement site as context features. In addition to sequence context, numeric genomic annotations (e.g. conservation signal, GC content, DNA replication timing) or the results of [*]-seq experiments (e.g. ChIP-seq, R-loop annotation, nuclear proximity) may be considered as topographical context features. As has been done for SNVs, these definitions of junction context could be used to define footprints of mutational processes driving rearrangements, either in naturally occurring cancers or model systems. These context features could be measured in reference states (i.e. prior to rearrangement) or rearranged genomes (e.g. tumor samples) to assess both causes and consequences of rearrangements. Such an approach has been informally used in previous studies [28] to identify novel rearrangement signatures.

From pattern to mechanism, and back

The identification of recurrent patterns in genomic data inspires the search for mechanisms, through the construction of experimental systems and conceptual models that are sufficient to generate a set of observations. The prototypical example of this link between pattern and mechanism has been the study of chromothripsis, where elegant engineering of cells in culture have established sufficient conditions for the phenomenon of “chromosome shattering”. Zhang and colleagues employed live cell microscopy and single-cell sequencing to demonstrate a high frequency of chromothripsis induction in the immediate progeny of cells with induced micronuclei [50]. Maciejowski and colleagues induced chromosomal bridges via telomere dysfunction and demonstrated chromothripsis and kataegis through WGS of derived clones [51]. Mardin and colleagues employed chemical double stranded break induction and selection for anchorage independence growth followed by WGS of clones to recapitulate chromothripsis in culture [52]. While these studies have effectively solved a prominent cancer genomics puzzle, they also have also spawned new ones: What is the contribution of these competing mechanisms to chromothripsis events observed in real tumors? Can their signatures be recognized and analytically disentangled using cancer sequencing data?

The approach taken with the modeling of chromothripsis provides a template with which to mechanistically dissect other complex rearrangement phenomena, whether chromoplexy or BFB cycles. Such experiments may be proposed to follow a common template (Figure 4) following classic reverse-genetics approaches but employing modern genomic-profiling technology and analytic tools derived from machine learning. Such an approach has been employed to study BFBs in Caenorhabditis elegans, where whole genome sequencing of strains defective in telomere maintenance revealed clusters of rearrangements near chromosome ends that were consistent with the proposed BFB signature [53]. It will be important to explore similar experimental avenues in mammalian systems as holocentric C. elegans chromosomes can be stably maintained after fusion events and it is not clear how accurately the rearrangements observed in this system can explain rearrangement signatures observed in cancer genomics. Epithelial cancers that develop after telomere dysfunction in mice show chromosome rearrangements with amplified regions that are adjacent to translocations [54]. This landmark experiment was performed before the advent of massively parallel whole genome sequencing and it would be informative to revisit it and other classic experiments with an eye towards genomic phenotypes in order to conclusively link a major observed feature of cancer genomes with a biological mechanism. The engineered variant patterns generated in these carefully controlled experimental systems could then be analyzed in cancer genome data to close the loop and validate a mechanistic hypothesis.

Figure 4.

A map for computational mechanistic genome integrity experiments: Experiments begin with (1) an initial cell, tissue, or organism type which undergoes (2) perturbations to alter biological context, ablate specific pathways, and/or induce mutagenesis through one of several mechanisms. Then, to enable (3) initial profiling and selection, nucleic acids are extracted from bulk tissue or via multiple-displacement amplification from single cells (+/- time lapse microscopy). Alternatively, clones are expanded and selected through transformation or karyotype screening. Finally, (4) comprehensive genomic profiling is performed, using whole genome sequencing (via standard short reads or long-range sequencing technologies) or other genomic and epigenomic profiling approaches. The same profiling can also be performed on the unperturbed cell type. The resulting variant data and their contextual features are analyzed using machine learning algorithms to identify recurrent patterns and link them to their biological context (e.g. pathway ablation). Among the several paths in this map are several recent studies dissecting the mechanistic basis for chromothripsis [50–52].

Towards a systems biology of genome integrity

A systems biology approach to cancer genome integrity offers insights into the mechanisms that maintain genome integrity and into the biology underlying structural DNA variations. We propose an experimental framework that takes the shape of classic reverse genetic approaches to probe the genomic consequences of chemical or genetic perturbations (Figure 4). This biology-first approach is necessary to circumvent the limitations associated with correlative analyses performed in most bioinformatics-based studies, such as deficiencies in genetic pathways that are not under immediate consideration. Implementation of this framework depends on (1) the selection of a genetically amenable or chemically accessible model to be studied. Model systems include cultured cells or widely used model organisms such as mice or worms.

One important consideration is that the system selected for study has a stable genome. The model system will serve as a control in downstream sequencing-based readouts. For the purpose of structural variant-based studies (2) a genetically or chemically engineered biological context can be engineered to test a specific hypothesis. Interesting chemical perturbations include treatment with chemotherapies of uncertain genomic effect such as ionizing radiation or conventional or next generation anti-mitotic agents such as paclitaxel and Mps1 inhibitors. Genetic manipulations can be introduced to determine the proposed dependency on APOBECs for kataegis-associated mutations [21] or BRCA1 deficiency as the mechanistic cause of the tandem duplicator phenotype [55]. In past work, it has been crucial to perform (3) initial profiling in order to focus sequencing efforts on targets with notable structural variants. Prior profiling has included time-lapse microscopy to target chromosome mis-segregation events [50], karyotype screening to identify cell clones with interesting rearrangements [51], and cellular transformation [52]. Finally, (4) genomic and epigenomic readouts of junctions, chromatin states, and breaks will provide contextual information necessary to build structural variant signatures. Breakpoint junctions detected by whole genome sequencing can be analyzed in the context of nuclear proximity in parallel Hi-C assays. Other sequencing-oriented applications such as END-seq can be used to analyze the genomic factors that influence choice of DNA repair pathway [56].

Conclusion

In this review, we have highlighted key results from cancer genomics mutational process research, while advocating a more formal interaction between cancer genomics and genome integrity fields. We have pinpointed several prominent genome integrity hypotheses that have not yet been addressed in cancer genomics data and cancer genome phenomena that have yet to be modeled in experimental systems. Looking forward, we mention specific experimental and algorithmic approaches that could bolster and enrich this interaction, as experimentalists and theoreticians, computationalists and bench biologists join forces to make sense of a fundamental cancer hallmark.