Study Procedures

The workflow for the CASCADE program is shown in Fig 1. Once recruited, the CASCADE coordinator communicates with the patient, their family, and the associated health care providers regarding the patient’s status and ongoing willingness to participate. Following the patient’s death, the next of kin specifies when the body can be transferred for autopsy. Autopsies were conducted by a specialist forensic pathologist assisted by the research team. Tumour tissue was harvested from multiple representative metastatic sites. Each lesion was divided into portions that were immediately snap frozen in liquid nitrogen and also fixed in formalin for subsequent paraffin embedding. Frozen samples were used for sequencing where possible. Prospectively collected fresh frozen samples were available for some cases. For all fresh tissues, frozen sections were reviewed by a pathologist to confirm the presence of tumour, quantify necrosis, and estimate tumour cellularity. Premorbid formalin-fixed, paraffin-embedded (FFPE) archival samples from primary tumours and metastatic lesions were also obtained.

Download:

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

Fig 1. Workflow diagram giving overview of methodology.

https://doi.org/10.1371/journal.pmed.1002204.g001

Whole Exome Sequencing

DNA was extracted from sections cut from frozen tissues or FFPE blocks. For three of the cases, ER1–ER3, whole exome libraries were prepared using the Roche-NimbleGen SeqCap EZ exome version 3. For case TN1, libraries were prepared using the Illumina Nextera Rapid Exome. All libraries were sequenced on the Illumina Hiseq platform. TN1 had DNA submitted for single nucleotide polymorphism (SNP) array analysis with the Illumina Human Omni 2.5S beadchip, and other cases had DNA submitted for SNP array analysis on the Affymetrix Genome Wide SNP 6.0 array. Blood was available for all cases for reference germline DNA.

After adapter trimming with cutadapt [10], raw sequence data were aligned to the human reference genome GRCh37 with BWA [11]. Two samples for TN1 exhibited low sequencing yield and a large number of likely artefactual mutations due to FFPE processing, and were excluded from further analysis. Fresh frozen samples were sequenced to a mean depth of 105, with FFPE samples sequenced to a mean depth of 75. Following alignment, BAM files were processed according to GATK best practices [12]. Variants were called using MuTect [13] with default parameters. Indels were called with VarDict [14] and pindel [15].

Called variants were annotated with ANNOVAR [16]. Variants were filtered as follows: exclusion of variants in 1000 Genomes and the Exome Sequencing Project [17,18]; variants with an allele frequency ≥2% or with 2 or more supporting reads in the normal; and allele frequency <5% in fresh tissues or <10% in FFPE tissues. Indels were manually reviewed and excluded if adjoining >4 homopolymer runs or repetitive regions.

Copy number was called from whole exome sequencing (WES) data using the facets package, which also provided purity and ploidy estimates [19]. For the Illumina bead chips, raw data were tQN normalised [20], and segmented allele-specific copy number was called using OncoSNP [21]. For the Affymetrix SNP chips, samples were processed using the Aroma package, applying TumorBoost and PSBCS [22,23]. Output of facets was compared to the SNP chip data, and, in some cases, the parameters were adjusted to exclude less likely combinations of copy number and purity. As SNP arrays were not available for the FFPE samples, whole exome copy number was used for all analyses. Copy number segmentation determined from SNP arrays was used as orthogonal validation of WES copy number calls from facets.

To determine the functional significance of novel nonsynonymous variants, the following steps were taken: search COSMIC [24] and TCGA [25] for previous reports (across all tumour types); prioritise variants with a high CADD score [26]; map variants to functional domains of the gene, if possible; check if variant falls in a mutational hotspot using MutationAligner [27] and Cancer HotSpots [28]; and review literature (using GeneRIF queried via MyGene.info [29]). Annotation of CNA was limited to genes found to be recurrently altered in the literature [30,31].

Subclonal Inference

superFREQ is a cancer exome clonality inference tool that takes advantage of multiple samples from the same individual by tracking both single nucleotide variants (SNV) and CNA across samples [32]. This allows the detection of highly subclonal somatic mutations that are present at higher cell fraction in other samples. An advantage of superFREQ is that it estimates and propagates statistical and systematic error sources throughout the analysis, thus decreasing the number of false positives and allowing downstream analysis of uncertainties. It also accepts aligned sequence data as input directly. Briefly, the pipeline performs the following steps: (1) GC bias correction, (2) differential coverage analysis with limma-voom [33], (3) examines variant positions shared between individual samples, flagging variants for base quality, mapping quality, strand bias, stuttering, or other artefacts detected in the pool of normal control samples, (4) summarises differential coverage and SNPs for each gene, (5) recursively clusters neighbouring genes with sufficiently similar differential coverage or SNP frequencies until a segmentation of the genome is achieved for each sample, (6) summarises consensus coverage and SNP frequencies for each segment and renormalizes, taking accuracy of coverage and SNP frequency into account, (7) calls CNAs and clonality in each segment based on coverage and SNP frequency, (8) calculates clonalities of somatic SNVs using local CNA, (9) tracks clonalities of SNVs and CNAs over multiple samples to determine if the same or different alleles are gained/lost between samples, (10) clusters mutations (SNVs or CNAs) with similar clonalities in all samples into clones, (11) sorts clones into a tree structure, with smaller clones being assigned as subclones when the sum of the clonalities of disjoint subclones cannot be larger than the clonality of the containing clone. Not all mutations are allocated to subclones by superFREQ, usually due to highly deranged copy number, which does not permit the cancer cell fraction to be reliably determined. Others have noted this problem with subclonal reconstruction [6].

Case TN1 displayed a high degree of copy number variation, including large swathes of loss of heterozygosity (LOH) affecting over 50% of the genome. Accurate clonal inference with superFREQ was not successful. As an alternative, mutation clonality was determined using PyClone [34], with allele-specific copy number provided by the R package facets [19]. Mutation genotypes were built using “parental_copy_number” mode. For mutations in regions of LOH, genotypes consistent with LOH were assigned a prior weight of 10, and other genotypes down weighted to 0.1. All other genotype prior weights were left as 1. From the PyClone output, the median clonality for each mutation (or cancer cell fraction) was clustered via affinity propagation using the Schism package [35]. Mutation clusters were reviewed to separate agglomerated private subclones if needed, and then the reviewed clusters were supplied to Schism’s genetic algorithm to construct a multi-sample phylogeny. To validate this approach, high depth validation sequencing of mutations found in more than one sample was subjected to the same process to check the consistency of the tree with the WES data.

Validation

From each case, mutations and indels were selected for high depth validation based on their contribution to the subclonal phylogeny and biological interest. An Ion AmpliSeq custom panel was designed for the variants of interest in each case, and multiplex PCR was performed as per the manufacturer’s protocol. Libraries were sequenced on the Ion Torrent PGM sequencer (Life Technologies) to a median depth of 770. Wild-type and mutant reads from validation loci were extracted from BAM files using the GenomicAligments R package [36]. Taking loci where there were at least 20 reads, a mutation was considered validated if there was a significant deviation from an expected error rate of 1/200 using the p-binomial test [1].

Mutational Signatures

Mutational signatures were ascertained in a two-step process. Somatic mutations were filtered to have allele frequency >10% in FFPE samples to avoid detecting formalin fixation artefact and >5% in fresh samples. To increase detection power and avoid spurious signature detection, all unique mutations from each patient were pooled together and used as input to deconstructSigs [37]. “default” normalisation was used as per the authors’ instructions, with a minimum signature contribution of 0.06. The 30 signatures found in the latest COSMIC classification were used [38,39]. The contribution of the restricted set of signatures found in the pooled mutation set was then examined in each sample. To test the robustness of signature detection, pooled mutations were randomly downsampled to between 10% and 90% of the original mutations, and deconstructSigs was re-run on 1,000 such downsampled sets.

ctDNA and Immunohistochemistry

ctDNA was assayed as previously described [40]. For immunohistochemistry, heat-induced antigen retrieval was performed in 1× citrate buffer (Thermo Scientific). Samples were blocked in 2% bovine serum albumin in tris-buffered saline/Polysorbate 20 solution and endogenous peroxidase inactivated in 1.5% H2O2. Samples were incubated with primary antibodies, including AGF2 (Abcam), CHD4 (Abcam), and Cullin1 (Cell Signalling Technology). Biotinylated species-specific secondary antibodies were used at 1:300 (Dako) followed by the avidin-biotin-complex (ABC) method prior to visualisation with 3,3′-Diaminobenzidine (DAB) chromogen (Dako). Bright-field microscopy was performed on an Olympus BX-51 microscope. Murine inguinal mammary fat pads were used as normal control tissue.

Subclonal Architecture Defines Metastatic Spread

Inferring subclones and constructing subclonal phylogenies is a complex problem, requiring high sequencing depth and/or multiple samples to increase sensitivity and reduce the space of possible tree reconstructions [41–43]. By utilising multiple samples from each patient, we were able to construct consistent subclonal phylogenies using mutation and copy number information (Figs 2–5). Private subclones (that is, subclones that are found only in one lesion) are not displayed in these figures for clarity, as they do not contribute to the core tree structure; therefore, all clones displayed were found in at least two spatially distinct sites.

Download:

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

Fig 2.

Treatment history (A) and subclonal structure (B and C) for ER1, inferred by superFREQ. Pre- and postmortem samples are shown. AC: doxorubicin, cyclophosphamide; CMF: cyclophosphamide, methotrexate, 5-fluorouracil; Cyclo-MTX: cyclophosphamide, methotrexate; Gem-Carbo: gemcitabine, carboplatin; Everolimus-Ex: everolimus, exemestane. P1: multiple samples from archival breast primary; A1, A3, A4: right liver lobe metastases; A2: left liver lobe metastasis; A5: right adrenal lesion.

https://doi.org/10.1371/journal.pmed.1002204.g002

Download:

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

Fig 3.

Treatment history (A) and subclonal structure (B and C) for ER2, inferred by superFREQ. Pre- and postmortem samples are shown. AC: doxorubicin, cyclophosphamide; Gem-Paclitaxel: gemcitabine, paclitaxel; Everolimus-Ex: everolimus, exemestane. P1: multiple samples from archival breast primary, including pretreatment breast core biopsy; M1: liver core biopsy; A1: dural metastasis; A2, A3, A4: left liver lobe metastases; A5, A8: para-aortic nodal metastases; A6, A7: right liver lobe metastases.

https://doi.org/10.1371/journal.pmed.1002204.g003

Download:

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

Fig 4.

Treatment history (A) and subclonal structure (B and C) for ER3, inferred by superFREQ. Pre- and postmortem samples are shown. Cape-Cyclo: capecitabine, cyclophosphamide; P1: multiple samples from archival breast primary; M1: ovarian metastasis; M2: supraclavicular lymph node biopsy; A1: brain metastasis; A2, A9: left and right hilar lymph nodes; A3: left lung lower lobe metastasis; A4: left paravertebral soft-tissue metastasis; A5, A6: anterior and posterior left liver lobe lesions; A7: perinephric metastasis; A8: right upper lobe lesion; A10: rib metastasis; A11, A12: right liver lobe metastases.

https://doi.org/10.1371/journal.pmed.1002204.g004

Download:

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

Fig 5.

Treatment history (A) and subclonal structure (B and C) for TN1, inferred using PyClone and Schism. Pre- and postmortem samples are shown. FEC: 5-fluorouracil, epirubicin, cyclophosphamide; TCH: docetaxel, carboplatin, trastuzumab; AURKA: aurora kinase; P1: multiple samples from archival breast primary (post-neoadjuvant chemotherapy); M1: liver core biopsy; A1: left upper lobe metastasis; A2: left liver lobe metastasis; A3, A4, A5: left lung metastases; A6: right liver lobe metastasis.

https://doi.org/10.1371/journal.pmed.1002204.g005

All ER-positive cases had founding clones containing known driver alterations. These founding clones developed secondary alterations before giving rise to diverse subclones across multiple metastatic sites. By way of example, in ER1 shown in Fig 2B, we inferred the presence of an intermediate blue clone not found in isolation. This follows from (1) the absence of this clone in the primary tumours, indicating the blue clone is a true subclone; (2) the absence of the orange subclone in two samples where the red subclone has a high clonality, indicating the red subclone did not arise in the orange subclone; and (3) two different subclones (red and orange) with the blue clone as an ancestor. This intermediate blue clone was marked by an FGFR4 tyrosine kinase domain mutation and a splice site mutation in MGA, which counter-regulates MYC activity. The blue clone also shows deletions in genomic regions containing known tumour suppressor genes RB1, SPEN, CASP9, and FANCA. These mutations and CNAs likely conferred metastatic potential, as all subsequent metastatic subclones arose from this intermediate clone, as described above. The FGFR4 mutation could not be detected in three spatially separated samples of the primary tumour, even with high depth validation sequencing, although this does not rule out very low prevalence subclones at the time of diagnosis. Emergence of an unheralded “lethal subclone” following treatment has been documented in many tumour types, including breast cancer, but here it associated with delayed disease relapse.

ER2 and ER3 display a different relationship between subclonal structure and metastatic disease. For ER2, a biopsy of the breast mass was the index lesion, and the breast primary was removed after chemotherapy, followed by a liver biopsy after several lines of therapy. All premortem samples displayed a similar subclonal structure, with a subclone (orange) marked by a frameshift deletion in the cancer testis antigen MAGEA3 and a splice site mutation in ROCK1 that was subsequently detected at multiple sites in the liver at autopsy. The absence of this subclone at other metastatic sites, however, shows there was early divergence and parallel evolution of at least two other subclones (blue and red). ctDNA for the AKT1 (E17K) mutation and three ESR1 mutations (D538G, S463P, E380Q) were assayed in plasma taken at the time of liver biopsy and subsequently after death. Fig 6F shows all these mutations were detectable at both time points (albeit at very low levels for the first time point), even though the liver biopsy WES data did not reveal any ESR1 mutations.

Download:

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

Fig 6. Mutational signatures in TN1 and ER2, and ctDNA assays in ER2.

(A) Reconstruction of mutational context plot with deconstructSigs for TN1 using COSMIC signatures 17, 24, and 29. 5′ and 3′ nucleotides are indicated by colour code on x-axis. (B) Contribution of each signature per sample. (C) Subclonal phylogeny for TN1 showing private and public subclones. Private subclones in dark blue. Mutations arising from signature 17 represented in pink along branches. (D) Reconstruction of mutational context for ER2 using APOBEC signatures 2 and 13 detected with deconstructSigs, with origin of key mutations overlaid. Colour codings as for (A). (E) Subclonal phylogeny for ER2, with private subclones in dark blue. (F) Bar charts show ctDNA results at time of liver biopsy and later at death. y-axis unit is copies per millilitre.

https://doi.org/10.1371/journal.pmed.1002204.g006

ER3 was a case of delayed presentation, with 2 years elapsing between initial detection of early stage disease and subsequent metastatic disease, without intervening treatment. No samples could be obtained from the original diagnosis. In a similar fashion to ER1, the primary subclones have a linear monophyletic relationship with the metastatic subclones. This case is distinguished, however, by an ovarian lesion that contains the founding clone only. This ovarian lesion was discovered incidentally during a therapeutic oophorectomy prior to any anticancer therapy. Hence, despite lacking the more complex subclonal structure seen in other lesions, this disease possessed the ability to metastasise very early. Whether this ovarian metastasis would have resulted in clinically significant disease is unknown. The rest of the lesions must have arisen from dissemination of a different subclone, which had acquired the SPEN K1838X nonsense mutation and deletion of the negative modulator of oestrogen signalling DUSP22. These alterations occurred while the patient was receiving endocrine therapy.

TN1 displayed both short linear evolution and early divergence. Fig 5 shows the subclonal structure of the primary is found in all metastatic lesions via the green clone, but the blue and orange clone diverged from a common ancestor derived from an earlier progenitor of the green clone. Fig 5C shows the private subclones, some of which also emerged from the common ancestor. It is noteworthy that this case, with the most aggressive disease history with no response to three different standard chemotherapy regimens, had the fewest common subclones. S1 Fig demonstrates that CDH4, which contains a truncal mutation, is expressed at the protein level.

A key question is whether metastatic disease from one lesion can cross-seed anatomically distant sites. This is difficult to distinguish from a subclinical low prevalence clone seeding two different metastatic sites, a disseminative pattern rather than cross-seeding. All cases showed spatially distinct lesions with very similar subclonal structures (e.g., liver lesions in ER1; para-aortic nodal metastases in ER2; brain, lung, and liver metastases in ER3; lung metastases in TN1). This is suggestive of metastatic cross-seeding. In addition, ER1, ER2, and TN1 showed recurrent subclonal mixtures, raising the possibility of seeding by polyclonal clusters sustained by clonal cooperation, in which two subclones help maintain each other’s survival and confer novel phenotypic traits [3,44]. For such patterns to arise from dissemination alone rather than cross-seeding, polyclonal seeding, or cooperation, multiple waves of dissemination by different subclones from a primary tumour to the same metastatic site would be required.

Copy Number and Subclonal Evolution

In contrast to subclonal mutations, copy number changes between samples were relatively stable (Fig 7) for ER1, ER2, and TN1. Areas of LOH were particularly consistent across samples. ER1 and ER2 displayed relatively few CNA, with overall 80%–85% of coding genes unaffected. ER3 showed widespread copy number derangement affecting 40%–60% of coding genes, particularly on chromosomes 8 and 20. The ploidy of ER3 increased from early to late stage disease; the primary, ovarian metastasis, and lymph node biopsies all displayed diploid genomes, with autopsy samples showing triploidy and tetraploidy in some cases (mean ploidy across samples 3.2). There was inter-lesion copy number heterogeneity for ER3. This is reflected in the relative abundance of subclonal copy number events in ER3 called by superFREQ, with a mean of 22 copy number events per subclone for ER3, compared to 7 and 11 for ER1 and ER2, respectively. Chromosomal instability (CIN) was therefore an ongoing process contributing to subclonal evolution in ER3. TN1 showed the most deranged genome, which also had a mean ploidy of 3.9, consistent with genome doubling, with 40%–50% of genes subject to LOH and some form of copy number derangement affecting 70%–80% of genes. Subclonal copy number could not be called satisfactorily for TN1. Although less deranged, in ER1, there was evidence that the primary event was amplification of 1p, prior to driver mutations in PIK3CA (Fig 2B).

Download:

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

Fig 7. Heatmap of copy number across all samples.

x-axis shows all coding genes in order of chromosomal coordinate. Grey areas on map represent difficulty in calling allele-specific copy number in FFPE samples. Horizontal blue lines indicate areas of LOH. Deep red is high level amplification, pale red is amplified with four or more copies, and blue is deleted with less than two copies. White is two to three copies. Ploidy is displayed in last column on right. y-axis terms correspond to Figs 2–5.

https://doi.org/10.1371/journal.pmed.1002204.g007

These findings are consistent with the growing body of work that extensive CIN takes place early in the evolution of a tumour. In particular, a variety of data from single-cell sequencing of tumours and normal cells has shown tumours contain only a few copy number states, without a large number of intermediates, suggesting that large-scale copy number changes appear to occur in a punctuated fashion rather than accumulating gradually over time [8]. A limitation of studying this question in bulk sequencing is that detecting subclonal copy number changes with high sensitivity is difficult, particularly in specimens with CIN and suboptimal purity.

Treatment Resistance

These cases present a unique opportunity to study clonal evolution during therapy. Breast cancer is the prototypical “treatable” solid tumour for which patients receive multiple lines of therapy that usually result in a partial response or stable disease.

There was clear evidence that treatment alters subclonal structure. Disease at autopsy showed subclones that contained mutations conferring chemotherapy resistance. In ER1 one such mutation, TYMS V106M, is at a highly conserved residue in the binding site of thymidylate synthase, adjacent to two mutations known to cause resistance to 5FU in vitro (Fig 2B, yellow branch) [45].

Mutations in the ligand binding domain of ESR1 have recently been discovered as a common mechanism of resistance to endocrine therapy in metastatic breast cancer [46]. ER2 displayed three different ESR1 mutations all in different subclones. Furthermore, an ESR1 S463P mutation occurred in a subclone that already harboured an ESR1 E380Q, presumably in a different allele. As ESR1 mutations can be treated with alternative endocrine therapy and represent a tractable form of endocrine resistance, identifying them in patients failing endocrine therapy is of some importance [47]. ER2 had a liver biopsy while alive that did not show any ESR1 mutations, highlighting the difficulty in managing advanced heterogeneous disease. At the time of the liver biopsy, circulating cell-free DNA in plasma showed detectable low levels of all three ESR1 mutations (Fig 6F), which increased dramatically by the time of autopsy.

ER3 demonstrated discordant clinical responses on several occasions, with disease in the lung responding while disease in the liver progressed and left and right hilar lymph nodes responding differentially to chemotherapy. Fig 5C shows the divergent subclonal structure of lung, liver, and hilar lymph nodes, although this may be a consequence of heterogeneous treatment responses rather than the cause. Fig 7 also shows that differences in copy number can be discerned between sample pairs in ER3, more so than the other cases, as discussed in the previous section.

We found additional evidence of this effect when considering the evolutionary trajectory in the context of dominant oncogenic drivers. In ER2 and ER3, there were multiple subclones with alterations in genes that interact with the truncal drivers in each case, AKT1 and TP53, respectively (Fig 8). For ER2, alterations occurred in genes involved in the regulation of phosphatidylinositide phohsphates and PI3K signalling (PLPP4, PNPLA6, and PIK3R5), negative regulators of AKT1 signalling (PLEKHO1, PHLPP1, PID1, and USP12), and downstream effectors of AKT1 (MFN1). For ER3, alterations occurred in coactivators and facilitators of TP53 activity (BACH2, ANKRD11, TP53I11, MYO10, WDR3, and NDUFAF6), negative regulators of TP53 function (G3BP2), and downstream effectors (CD82). These findings suggest that resistance to chemotherapy agents may arise from augmentation of existing oncogenic or tumour suppressor signalling pathways rather than direct resistance through altered drug targets or drug metabolism. This could be one explanation why increasing lines of treatment become progressively less effective, despite having non-overlapping mechanisms of action.

Download:

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

Fig 8.

Evolution of augmented oncogenic signalling in (A) ER2 and (B) ER3. Subclonal phylogenies with private subclones are displayed with alterations that are expected to modulate oncogenic signalling.

https://doi.org/10.1371/journal.pmed.1002204.g008

Mutational Signatures Drive Heterogeneity and Treatment Resistance

Due to relatively low numbers of mutations, clear mutational signatures could be established only in ER2 and TN1. The pattern of mutational signatures in these two cases were robust and maintained a stable prevalence even when the starting mutation pool was down sampled to 10% of the total.

TN1 had an unusual composition of mutational signatures (Fig 6A–6C). There was a strong signal from signature 17 in the COSMIC classification [38,39]. This signature has been reported in breast cancer as well as oesophageal, liver, lung, and stomach cancers, melanoma, and B-cell lymphoma [38]. The aetiology is unknown. As signature 17 produces a characteristic T > G transversion that has relatively little overlap with other signatures, the contribution of the signature to the mutations in subclones can be tracked. Mutations arising from this signature were found primarily in metastatic lesions from autopsy, comprising between 9%–45% of the private mutations in four spatially distinct lesions in the lung and liver. The signature was not active in the primary lesions or premortem liver biopsy. These private signature 17 mutations have subclonal fractions in the 0.2–0.6 range and cluster with other mutations to form private subclones. Assuming that it is extremely unlikely a mutational process could cause the same mutation in multiple cells within a single lesion, each private subclone with a relatively high subclonal fraction of the signature 17 mutations must have arisen from a very restricted population of single cells that expanded in each lesion. In addition, as a mutational process per se confers no proliferative advantage, the prevalence of this signature may represent cells that were able to survive a treatment bottleneck, sustaining DNA damage in the process. The only treatment the patient received after the liver biopsy (which did not contain signature 17) was a pan aurora kinase inhibitor. There was no evidence of response to this therapy. It is unclear how an aurora kinase inhibitor could cause this mutational signature. Signatures 24 and 29 also show activity in TN1 samples. Signature 29 has been associated with the mutagenic effect of tobacco chewing on the oral mucosa, and, although distinct from the classic smoking associated signature 4, there is a high degree of overlap. Signature 24 is reported to occur with aflatoxin exposure. The significance of these carcinogen-induced signatures in a triple negative breast cancer is unclear.

ER2 was dominated by signatures 2 and 13 in the COSMIC classification [39], which both relate to the well-known phenomenon of APOBEC deaminase activity (Fig 6D). This signature was present consistently in both early and late clones and allows us to study how APOBEC may contribute to the evolution of breast cancer. Key driver mutations AKT1 E17K, SPEN, and the ESR1 E380Q mutation are consistent with those induced by APOBEC. However, the ARID1A and the TP53 truncating mutations have a low probability of arising from the APOBEC signature, and the ESR1 S463P and D538G mutations are not consistent with APOBEC. In this case, therefore, APOBEC-related activity alone was able to furnish most of the putative driver mutations and resulted in an ESR1 mutation that causes resistance to endocrine therapy but could not explain all of the ESR1 mutations.