Mutations in the HCC1187 Genome

We assembled a list of all known sequence-level mutations in the HCC1187 genome (Table S6 in File S2). Non-synonymous coding-sequence mutations were compiled from the coding sequence screen of Wood et al [3] and other, targeted, re-sequencing studies [19]. In total, there are 85 known sequence-level mutations, comprising 75 base substitutions and 10 indels.

Structural mutations, i.e. translocations, inversions, duplications and deletions, and the resulting fused or interrupted genes, were compiled by combining our previous analysis of the karyotype [12], [13] with SNP6 array-CGH copy number data [20] and massively parallel paired-end sequencing [14] (Tables S4, S5 in File S2).

Our previous analysis had identified the breakpoints of chromosome rearrangements larger than about 3 Mb [12], and mapped all balanced breakpoints to gene level, but many unbalanced rearrangements had been mapped only to 1 Mb resolution. SNP6 array data allowed us to map these unbalanced breakpoints more precisely, to around 10 kb resolution, and detect deletions of less than 3 Mb. Paired end sequencing data identified the junctions of around 40 percent of the known rearrangements to sequence level.

Smaller-scale rearrangements, below the resolution of our previous analysis, were also apparent in the SNP6 array data–13 small deletions ranging from 0.26 kb to 2.3 Mb with a median size of 257 kb were predicted. There were also 24 small duplications ranging from 11.7 kb to 2.8 Mb, median size 320 kb. All of these duplications and deletions were absent in the matched normal lymphoblastoid cell line, HCC1187BL. Many of these features were likely to be small interstitial deletions or “head to tail” tandem duplications. Indeed, five of the 13 deletions and 17 of the 24 duplications were confirmed by structural variants detected by the paired-end sequencing [14]. We identified broken genes and possible gene fusions for all these additional structural changes (Tables S1–S6 in File S2). (Paired end sequencing also uncovered further apparent structural variations that were below the resolution of SNP6 segmentation [14]. These were not included in the present analysis, though we checked that they predicted no additional fusion genes).

These structural rearrangements gave rise to at least twelve expressed fusion transcripts, confirmed by RT-PCR and Sanger sequencing: RGS22-SYCP1, CTAGE5-SIP1, PLXND1-TMCC1, SEC22B-NOTCH2, KLK5-CDH23, BC041478-EXOSC10, AGPAT5-MCPH1, SUSD1-ROD1/PTBP3, SGK1-SLC2A12, RHOJ-SYNE2, PUM1-TRERF1 and CTCF-SCUBE2, some of which have been reported previously ([12]–[15] and Table S4 in File S2). Of these twelve, the first four were predicted to form an in-frame fusion product.

HCC1187 Endoreduplicated during its History

The HCC1187 karyotype is hypotriploid and highly rearranged, like most breast cancers. The karyotype is highly likely to have evolved via successive chromosome loss, unbalanced translocation and endoreduplication, since this is the predominant pattern in breast tumors (Fig. 1) [17]. We therefore looked for signs of endoreduplication.

The main evidence that endoreduplication had occurred was that a high proportion of the genome had been duplicated precisely once. To make this clearer, we worked out which chromosome segments derived from which parent by analysing how many copies of each genomic segment had the same alleles, using SNP array data (Fig. S1 in File S1). We were able to assign almost all chromosome segments in the karyotype to one or the other allelotype (Fig. 2). This showed that many chromosome segments were present in two copies of the same parental origin, and most of the remainder appear to have evolved from a pair of copies (Fig. 2B). For example, all segments of chromosomes 6 and 7 are present in two copies, while there are two complete copies of chromosome 16 derived from the same parent, one of which has been split by a balanced translocation.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. The structure of the HCC1187 genome.

A) Spectral karyotype as in [44]. Chromosomes are named A-Z and a-k based on their relative sizes as in [12]. Cytogenetic description of the karyotype is in Table S1 in File S2. B) Circos plot [45] of the HCC1187 genome: Chromosome ideograms around the outside, oriented clockwise pter to qter. Moving inward, the pale grey and dark grey boxes are chromosome segments observed by array painting [12] with their chromosome of origin indicated. Their parent of origin (light grey and dark grey) was deduced from the number of allelotypes given by PICNIC segmentation (Fig. S1 in File S1). Note that assignment of parents 1 and 2 does not transfer between chromosomes. Dark blue line, total copy number, equivalent to array CGH, from PICNIC. Red line, copy number of the minor allele; where this is zero, the genome is homozygous. Chromosome segments that share a translocation breakpoint were assumed to have the same parental origin. Inner links represent interchromosome translocations identified previously [12]–[14].

https://doi.org/10.1371/journal.pone.0064991.g002

Inferring the Genome State before Endoreduplication

Having clear evidence that endoreduplication had occurred in HCC1187, we were able to infer the state of the genome immediately before it doubled (Fig. 3). To do this, we assumed that the simplest possible sequence of events had happened, in particular that endoreduplication accounted for almost all duplications. The resulting picture of the likely history of the karyotype almost exactly fits the suggested monosomic pattern of evolution (Fig. 1), with each unbalanced translocation leading to loss of one chromosome, plus some whole-chromosome losses.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 3. Chromosome segments in HCC1187 and their most probable state before endoreduplication.

Chromosome ideograms are drawn around the outside as in Fig. 2. Outer rings are array painting segments as in Fig. 2. Inner rings are chromosome segments that must have been present before endoreduplication. Coloured circles are different types of mutations, on the outer chromosome segment on which they were observed: truncating (red), non-synonymous (blue), small deletion (yellow), small duplication (black), expressed gene fusion (light blue). Mutations that were on two copies of a chromosome segment probably occurred before endoreduplication and are also shown on the inner, pre-endoreduplication genome. Dashed grey boxes on chromosome 1 and 11 indicate regions where parental origin was undetermined, because PICNIC segmentation suggested additional rearrangements had taken place.

https://doi.org/10.1371/journal.pone.0064991.g003

We next determined whether the somatic mutations most likely occurred before or after endoreduplication. As most loci in this genome had duplicated only once we could infer whether a mutation happened before or after endoreduplication: if the mutation occurred before, it would be present in two copies after the duplication, whereas if the mutation occurred after endoreduplication, it would only be present on one of two copies. This classification would be wrong if gene conversion had occurred, i.e. copying of an allele from one chromosome to another, but all of the 83 sequence-level mutations analysed below were found on only one parent of origin, implying that gene conversion was rare or absent in this cell line, as is typical for epithelial cancers [21], [22].

We placed each structural mutation before or after endoreduplication, according to whether the translocation junction or deletion was duplicated, or involved only one copy of a pair of participating chromosomes (Figs. 2 and 3). For the three regions of the genome that were triplicated we assumed that one duplication had occurred at endoreduplication and another had occurred later. We determined the copy number of gene fusions, and confirmed their chromosomal location by fluorescence in situ hybridization (FISH) (not shown). For small deletions and duplications, we determined the copy number of the relevant region relative to flanking regions from array CGH segmentation to assess whether the segment bearing the deletion or duplication had itself been duplicated. Earlier deletions and duplication showed a copy number shift of two or more whereas later events had a copy number shift of only one.

Seven of the twelve fusion transcripts were classified as before endoreduplication; two, CTCF-SCUBE2 and BC041478-EXOSC10 were classified later. AGPAT5-MCPH1 and SUSD1-ROD1/PTBP3 and KLK5-CDH23 were undetermined, as their allelic copy number could not be resolved by array CGH or FISH. We were able to place seven deletions earlier, and these were all the homozygous deletions. Five deletions, all heterozygous, were placed later, with one undetermined. We could unambiguously place 14 small duplications relative to endoreduplication: seven earlier and seven later.

To assign point mutations to one or two copies of particular chromosomes, we isolated the individual chromosomes in a cell sorter and re-sequenced the mutated exons (Fig. 4). We confirmed this analysis for selected genes by measuring the relative number of mutant and wild-type copies using pyrosequencing (Fig. S2 in File S1). We were able to place 75 of the 85 previously described sequence-level mutations before or after endoreduplication, with only 10 undetermined. Of these ten, two were on a chromosome that was too small to be resolved in flow sorting, and 8 were not possible to score, either because they were found on single-copy genome segments, or they were found in a region where parent of origin could not be determined. Two reported mutations, in ZNF674 and HUWE1, were not found in our sample, therefore presumably occurred in other stocks of the line. They could therefore be classified as later (Fig. 3, Table 1 and Table S6 in File S2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 4. Point mutations on chromosome 6, and whether they occurred before or after endoreduplication.

A) Deducing the parental origin of chromosome 6 segments: the simplest explanation for the allele combinations (blue and red lines on the aCGH plot) in terms of parental origin. Both copies of chromosome 6 I (chromosome 6 fragments are designated 6 I, 6A, 6D as in ref. [12]) originate from parent 1 and the chromosome 6 segments of 6A and 6D originate from parent 2. Several small copy number steps are omitted for clarity. B) Sequence traces show whether mutations are on each isolated chromosome. HSD17B8: Chromosome 6I (2 copies) homozygous G>T mutation (black arrow); chromosome 6A and 6D, no mutation. NCB5OR: Chromosome 6, heterozygous mutant (black arrow). C) The likely evolution of the segments of chromosome 6: unbalanced translocation of one copy of chromosome 6 was followed by duplication of both chromosomes during endoreduplication. HSD17B8 was mutated on each copy of chromosome 6I, but not on 6A or 6D, while NCB5OR/CYB5R4 was mutated on only one copy of chromosome 6I. The pre-endoreduplication state was likely to be one normal copy of chromosome 6 with the other having a mutation in HSD17B8 and having suffered unbalanced translocation. The NCB5OR/CYB5R4 mutation occurred after endoreduplication.

https://doi.org/10.1371/journal.pone.0064991.g004

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Summary of Mutations in HCC1187 and their timing.

https://doi.org/10.1371/journal.pone.0064991.t001

Earlier and Later Mutations

Overall, the proportion of mutations classified as occurring before endoreduplication (earlier) was fairly similar for structural and point mutations (Table 1): 27/48 (56%) of structural changes (translocations, deletions and duplications) and 34/75 (45%) sequence-level changes were classed as earlier (Fig. 3 and Tables S4–S7 in File S2).

Among the structural mutations that could be classified, 13/22 (59%) of chromosome translocations were in the earlier group, while 7/14 (50%) of small duplications were earlier and 7/12 (58%) of small deletions were earlier. For fusion genes, 7/9 (78%) were classified earlier and, interestingly, all three in-frame fusion transcripts that could be classified were classified as earlier.

Of the classifiable sequence-level mutations, 58 were missense mutations, of which 23/58 (40%) fell early. To try to uncover ‘driver’ mutations within this group, we applied the Sorting Intolerant from Tolerant (SIFT) algorithm [23] to all of the point mutations, 47 of which could be scored by this method. Of the missense mutations predicted to be non-functional (tolerated) and so more likely to be random, 9/28 (32%) were earlier, while 7/19 (37%) mutations predicted to be functional (deleterious) were earlier (Table 1, Table S6 in File S2). Wood et al. [3] also identified genes likely to be drivers as ‘candidate cancer genes’ (CAN) based on their mutation rate and several other bioinformatic estimates of functionality [3]. The nine CAN genes showed a bias towards the earlier category, six classified earlier (INHBE, KIAA0427/CTIF, MYH9, PCDHB15, RNU3IP2/RRP9, TP53) and three in the later category (ABCB8, KIAA0934/DIP2C, NCB5OR/CYB5R4).

Strikingly different from the overall distribution of mutations in HCC1187 was the proportion of sequence-level truncation mutations in earlier rather than later categories: All eight classifiable INDEL mutations happened earlier, and combining this figure with nonsense mutations showed 11/13 (85%) protein truncating mutations happened earlier. This difference in proportion (11/13 truncating vs. 23/58 missense) is statistically significant (p<0.01 for chi-squared test with continuity correction).

Non-random Timing of Mutation Subsets

The distribution of mutations between earlier and later could uncover selective pressure for a mutation to occur at a particular stage in tumor development, or a change in the level of genetic instability. We therefore estimated the number of random and non-randomly timed mutations given the proportions of different mutation classes above.

We used a statistical model to estimate the number of mutations that showed non-random timing. The model assumed that any given class of mutations is a mixture of non-random mutations that must happen earlier (that is, before endoreduplication) and randomly timed mutations that can happen earlier or later. The randomly timed mutations are classified as earlier with probability p and later with probability 1-p, independently for each such mutation. We find the most likely number, n, of non-randomly timed mutations (the maximum likelihood estimate, or MLE) and its 95 percent lower confidence bound, given an estimate of p. Further details of the model may be found in File S3.

Estimates of p based on total missense mutations or those predicted to be non-functional (see Table 1) are 0.40 ( = 23/58) or 0.32 ( = 9/28), respectively, and a plausible upper bound would be 0.59 ( = 13/22), the proportion of earlier chromosome translocations.

Most classes of mutation, including non-synonymous point mutations, chromosome translocations, duplications, deletions, predicted functional mutations and CAN genes did not show any excess of mutation earlier or later. However, the observed proportion of truncating mutations falling earlier (11/13) suggests that n >0. When p  = 0.4, the MLE is n = 10 mutations that had to happen before endoreduplication, with a lower confidence bound of 6 (File S3) [24]. For p = 0.32 n = 10, lower bound 7. Thus our simple statistical model suggests that a number of the truncating mutations had to occur before endoreduplication.

When we use the high estimate for p, p = 0.59, the MLE was n = 9, but the lower confidence bound is 0, so data from more tumors would be required.

The Earlier Versus Later Classification

Endoreduplication in HCC1187 proved to be a useful milestone, because numbers of structural changes and point mutations were fairly equally distributed between the earlier and later categories, implying that endoreduplication occurred 32–60% of the way through the mutational history of this genome.

Interpretation of mutation timing depends on the accuracy of our earlier and later classification of mutations. We were confident that the tumor had undergone endoreduplication as it showed two characteristic signatures of this phenomenon: multiple duplicated rearrangements and multiple duplicated homozygous regions (Fig. 2). Given that there had been an endoreduplication, we reconstructed the main steps of HCC1187 karyotype evolution by assuming that the simplest possible sequence of events had happened. Implicit was the assumption that, as far as possible, all duplications had occurred at endoreduplication. The deduced sequence of chromosome changes (Fig. 3) was consistent with monosomic evolution (Fig. 1). Three duplications could not be explained by endoreduplication: these were three chromosome segments of the same parental origin that were present in three copies. The simplest route to these triplications was via endoreduplication followed by an additional single-chromosome duplication.

A few steps in the evolution may have been more complex, but this would not have altered the earlier versus later classification very often. Specifically, if all three triplicated chromosomes had taken a more complex evolutionary route (perhaps duplication followed by endoreduplication, followed by loss), the classification of no more than three point mutations could be affected, moving them from the later category to the ‘undetermined’ class.

Some mutations were omitted from analysis. These were from the complex regions of 10 p and 11 q where the parent of origin could not be determined. The omitted mutations comprised eight non-synonymous missense and two truncating mutations. Even if we consider the most unfavorable case, that the two truncating mutations were classified as later, the MLE for the number of non-random truncation mutations in the earlier group is is n  = 9, with a 95% lower bound of 3 when p  = 0.4, and n  = 5 or 6 with a 95% lower bound of 0 when p  = 0.59.

Timing of Genetic Instability and Other Functions

This earlier versus later classification may help us to understand a variety of issues including the timing and origins of chromosome instability and the drivers versus passengers problem for this particular tumor.

There has been much discussion of when chromosomal instability occurs, for example some have suggested it as a key facilitator of early tumorigenesis, notably causing loss of heterozygosity of APC in colorectal cancers [21]. In contrast, some suggest that the extensive rearrangements of carcinoma karyotypes might be late progression events [7], [8]. Others favor a transient period of chromosome instability, either at ‘crisis’ caused by telomere loss, or at some other catastrophic event such as ‘chromothripsis’, in which one or more chromosomes undergo massive rearrangement, apparently in a single event [9], [10].

We observed roughly equal ratios of structural to sequence-level mutations earlier and later. Although there are other possibilities, a reasonable explanation is that both mutational processes were happening at approximately the same rate for much of the evolution of the tumour. We did not see, for example, a much higher ratio of rearrangements to point mutations in the earlier class, as expected if most of the rearrangements had occurred during a telomere crisis before endoreduplication.

An important value of the classification is that mutations that may cause ongoing chromosome instability must generally be in the ‘earlier’ group as, by definition, they must pre-date almost all chromosome changes, which were quite numerous before endoreduplication. Among the mutated genes that might contribute to chromosome instability were TP53, BAP1 and PAXIP1, and all were indeed classified as earlier. BAP1/UCHL2 (BRCA1-associated protein1/ubiquitin carboxy-terminal hydrolase 2) was discovered as a binding partner of BRCA1 and appears to participate in the DNA damage response by interacting with BRCA1 and BARD1 [25]; [26]. It is a deubiquitinase that controls ubiquitination of Histone H2A, and is also a component of the Polycomb repressive deubiquitinase (PR-DUB) complex. Inactivating mutations of BAP1 have been found particularly in uveal melanomas [27], and as germline mutations that predispose to melanoma [28] and other cancers. Based on this, Stephens et al. [29] identified BAP1 mutations as likely driver mutations in breast cancer. PAXIP1 also seems to be part of the DNA damage signalling system, interacting through its multiple BRCT domains [30]. (Another gene relevant to genetic instability, but at the sequence level, was HUWE1, but we did not find the reported HUWE1 mutation in our stock. HUWE1 is implicated in base excision repair according to the UniProt database).

We also considered how other functions relevant to breast cancer might relate to timing, using the Entrez and Gene Ontology databases (Table S7 in File S2). Although we do not know which mutations were driver mutations, a gene involved in steroid hormone synthesis, HSD17B8, was mutated earlier, while genes encoding covalent modifiers of histones (HUWE1, IPO7, MLL4, PAXIP1, PRKAA2) were, except for PAXIP1, mutated later. Some of these histone modifiers might affect differentiation, while PAXIP1 might affect genome stability. However, other functional groupings were less informative: 17 of the genes mutated in HCC1187 were associated with actin and the cortical cytoskeleton or integrin signalling–functions highlighted in a study of triple-negative breast cancers [31] (HCC1187 is triple-negative)–but they were fairly evenly divided between earlier and later. Similarly the many genes associated with G-protein coupled receptors and Rho signalling, and genes associated with apoptosis, were fairly evenly divided.

Timing and Evidence for Selection

A non-random distribution of mutations between earlier and later categories could be evidence that a substantial number of mutations were drivers rather than passengers. The group of mutations showing greatest deviation from the typical pattern of around 40% earlier was the truncating mutations (and especially indels). A non-random distribution could be explained in two ways: i) the rate of indel mutations was high before endoreduplication and low after, relative to most other types of mutation ii) passenger indels accumulated in the same way as other passenger mutations but more indels accumulated early because they were selected. We consider (ii) most likely because for 9/11 earlier truncating mutations, a chromosome loss before endoreduplication caused loss of the second wild type allele. This is consistent with chromosome instability facilitating early tumor suppressor loss, as has been suggested previously [21]. Indeed, the earlier truncation mutations include known and candidate tumor suppressor genes TP53, BAP1 (BRCA1-associated protein), CTNNA1 (CateninA1) and NFKBIA (nuclear factor of kappa light polypeptide gene enhancer in B-cells inhibitor) [20], [29], [32]–[34]; others were AVPI1, GMCL1L/GMCL1P1, GPR81HCAR, MYH9, SLC4A3, ELP2 and TRIM47.

These data, therefore, support the view that early tumor suppressor loss is consistent with tumor evolving monosomically and that driver mutations that cause gene inactivation will be concentrated pre-endoreduplication. An explanation for this phenomenon is that loss or inactivation of two alleles pre-endoreduplication is more likely than loss/inactivation of four alleles post-endoreduplication [16], [17]. Gain of function mutations are not under the same numerical constraints as tumor suppressors. Where two hits are required to impair tumor suppressor gene function, only a single mutation is required for oncogenic gains of function and we might, therefore, see these mutations either side of endoreduplication. Fusion genes–whose importance in the common epithelial cancers has only recently been acknowledged [35], [36]– were formed throughout the evolution of this tumor but, interestingly, the in-frame fusions (most likely to be translated into functional proteins) were all formed early. Although the numbers are admittedly too small for statistical certainty, this makes the early in-frame gene fusions also good candidates for selected events.

Interpretations of the earlier and later classes depend on when HCC1187 endoreduplicated. There is some evidence that endoreduplication occurred in vivo in this case. The original ploidy of HCC1187 was not reported, only that shortly after its derivation, HCC1187 had multiple ploidy indices by flow cytometry [37]. However, around 60 percent of mutations occurred after endoreduplication. It would be surprising if so many happened in culture, given that cell lines largely recapitulate the genomic aberrations observed in primary tumors [38]. If endoreduplication happened in vitro, only ‘earlier’ mutations happened in vivo, and all driver mutations will be in the ‘earlier’ set, whereas if endoreduplication happened in vivo (as is often the case in breast tumours [17]), some driving mutations will be present in the ‘later’ group. In either case our estimate of non-randomly timed mutations remains the same.

Comparison with Mutations in Breast Cancers

To attempt to identify mutations in HCC1187 that are recurrent or known drivers, and relate this to timing, we compared sequencing data from breast tumours, both genomic sequencing data [29], [31], [39]–[41] and fusion gene data from transcriptome sequencing [15], [42]. However, this was largely uninformative, because of the heterogeneity and variability of mutations among cancer cases, and the still limited amount of data available. For example, Stephens et al [29] identified 31 genes as targets of driver mutations in breast cancers, from sequencing of 100 exomes and comparison with known drivers in other cancers, but the individual tumours had an average of only 1.7 of these genes mutated, 1.3 if TP53 is excluded. Unsurprisingly, then, in HCC1187 only two of these drivers, TP53 and BAP1, were mutated (both earlier). Similarly, of the genes considered likely drivers by Shah et al [31] in triple negative breast cancers, only SYNE2 and TP53 were mutated in HCC1187–SYNE2 is fused, also earlier. Nevertheless, many other genes mutated in HCC1187 are mutated or rearranged occasionally in the datasets above, so may be recurrent at a modest level (Tables S4, S7 in File S2). Several genes involved in fusions have been reported to be fused or rearranged in other cases–AGPAT5, NOTCH2, PUM1, SEC22B, SGK1 and TRERF1 (all early or unclassified), while several are mutated at sequence level, notably SYNE2. Among the genes in homozygous deletions (all earlier) with four or more reported mutations or rearrangements were SCN1A, FBXL20 and MYO9A; while among the point-mutated genes in HCC1187, apart from the known drivers BAP1 and TP53, ones with four or more reported mutations or rearrangements were CAMTA1, ITIH6/ITIH5L, LHCGR, SPEN, TP53 and ZNF142 (3 of which were earlier, 3 later). Many of the other genes have been found mutated in cancers other than breast, including CTNNA1 and NFKBIA (both earlier, see above). Others have cancer-relevant functions, such as steroid hormone synthesis (HSD17B8, earlier), and covalent modification of histones (HUWE1, IPO7, MLL4, PAXIP1, PRKAA2, all later except PAXIP1) (Table S7 in File S2).

Applicability to Sequencing Data

Our theoretical framework and statistical methods could be applied, in a modified form, to sequencing data from other endoreduplicated cell lines and primary tumours, indeed the idea of placing mutations before or after a duplication event has already been exploited [1], [18]. Endoreduplication is a common process in epithelial cancers, estimated to occur in more than 50% of breast cancers [17], [18]. Endoreduplicated genomes can often be identifed by copy number and allele ratios [18], for example, a large proportion of a recently-endoreduplicated genome will often be present either in four copies and heterozygous, or two homozygous copies (Fig. 4). We relied on flow sorting of chromosomes to quantify our mutations, but the proportion of mutant and reference alleles could be deduced, for example, by counting reads from deep massively-parallel sequencing. Earlier mutations will usually be homozygous in diploid regions, or account for approximately 50% of mutant reads in tetraploid regions. Distinguishing between earlier and later events in large datasets may help identify genes or pathways that must be mutated earlier or later in a given tumour type.

Conclusion

In conclusion, we provide evidence that, in this cell line, chromosome instability and rearrangement was not a late and irrelevant event, and that the great majority of inactivating mutations and expressed gene fusions appear to have happened early, and this suggests that most of them were selected.