Mutation Analysis of Cell-Free DNA and Single Circulating Tumor Cells in Metastatic Breast Cancer Patients with High Circulating Tumor Cell Counts

Breast cancer is the most common cancer in women and ranks as second in the world in terms of cancer deaths. Recent studies using next-generation sequencing (NGS) have shown that cancer evolves in the patient (1–3), and therapies can induce the evolution of clones of cells that are refractory to the treatment (4, 5). Much research has been devoted to developing bioassays that can help in the selection of therapies as breast cancer evolves, including characterization of circulating tumor cells (CTC) and the tumor-derived fraction of circulating cell-free DNA (cfDNA) termed circulating tumor DNA.

Although a number of commercial platforms are under investigation for isolation and characterization of CTCs without the use of a surface marker, CELLSEARCH (Janssen Diagnostics), which enumerates CTCs of epithelial origin [CD45⁻, EpCAM⁺, and cytokeratins (CK) 8, 18, and/or 19 positive cells], is currently the only FDA-approved platform in clinical use. EpCAM-positive CTCs as measured by CELLSEARCH are rarely detected in patients with primary breast cancer and many patients with metastatic disease also have few EpCAM-positive CTCs (0–5 per 7.5-mL blood). When detected, the number of EpCAM-positive CTCs reflects both the effects of systemic therapy (6–8) and prognosis (9). In contrast, circulating tumor DNA is detected in early and late stages in a high proportion of cases (10–12) and can persist for many years after supposedly successful therapy (13), despite no evidence of overt distant metastases. In metastatic breast cancer (MBC), the dynamics of circulating tumor DNA compare favorably to serum CA15-3 and CTC counts determined by CELLSEARCH in reflecting changes in tumor burden (14).

Metastases are heterogeneous, both between metastatic sites and also between the cells that compose each metastasis. Early metastases are composed of a greater proportion of cells with EMT and “stem-like” characteristics, whereas larger metastases are dominated by more differentiated, heterogeneous patterns of cells, more closely reflecting the heterogeneity seen in primary tumors (15). Taking all these considerations into account, together with the well-established phenomenon of clonal evolution (16, 17), it is evident that a single conventional tissue biopsy will not necessarily provide accurate information regarding appropriate therapy. An additional layer of complexity is suggested by the role of the immune system in patients with cancer, since immunodestruction of persisting micrometastases could mean that the circulating tumor DNA is predominantly a reflection of destroyed cancer cells; whereas persisting CTCs could be envisaged as a more accurate guide to appropriate therapy.

Previous studies have compared mutations in cfDNA to either biopsies of metastases or CTCs (2, 18). We reported emergence of ESR1 mutations in cfDNA, and in a single patient with a very high CTC count (>3,000 EpCAM-positive CTCs in 7.5-mL blood), the ESR1 p.D538G mutation was detected in cfDNA and matched CTCs (2). Since cfDNA is in principal easier to obtain than CTCs for molecular analysis, it is essential to determine whether mutations in cfDNA are the same as those seen in individual CTCs, as heterogeneity is frequently observed in metastatic biopsies (19, 20). Here, we compared total cfDNA levels and CELLSEARCH CTC counts with serum CA15-3 and alkaline phosphatase (ALP) levels in patients with MBC. We performed deep amplicon sequencing of 50 genes and compared mutation profiles in cfDNA and individual EpCAM-positive CTCs isolated from the same blood sample in 5 patients with high CTC counts (≥100 CTCs in 7.5-mL blood). Results show that cfDNA mutation profiles reflect mutational heterogeneity seen across individual CTCs and that total cfDNA levels and ≥5 CTCs are significantly associated with overall survival. Circulating tumor DNA is a useful substitute for metastasis biopsy (18). These data suggest that circulating tumor DNA reflects persisting EpCAM-positive CTCs in patients with high CTC counts and could enable monitoring of the metastatic burden but requires confirmation in larger groups of patients.

The study protocol was approved by the Riverside Research Ethics Committee (Imperial College Healthcare NHS Trust; REC reference number: 07/Q0401/20) and conducted in accordance with Good Clinical Practice Guidelines and the Declaration of Helsinki. All patients gave written informed consent prior to participation. We recruited 112 patients with radiologically confirmed MBC (Table 1). Twenty milliliters of blood was taken into EDTA tubes (BD Biosciences) and processed to plasma for cfDNA (21), and 7.5 mL blood was taken into CellSavepreservative tubes for CTC capture and enumeration with the CELLSEARCH CTC Test (Janssen Diagnostics) as described previously (22). CA15-3 and ALP results were obtained from patient notes.

cfDNA was isolated from 3 mL of plasma using the QIAamp Circulating Nucleic Acid Kit (Qiagen) according to manufacturer's instructions. Isolation of DNA from lymphocytes and quantitation of total cfDNA was as described previously (21). cfDNA levels (ng/mL) were converted to copies/mL plasma assuming 3.3 pg DNA per haploid genome. Formalin-fixed, paraffin-embedded (FFPE) tissue blocks with matching hematoxylin and eosin (H&E) were retrieved from the histopathology archive and reviewed by a consultant histopathologist. Two separate regions of tumor were then removed from the block using a 1-mm tissue microarray needle. DNA was extracted from the FFPE tissue core using the Qiagen GeneRead Kit according to manufacturer's instructions. CTC isolation by DEPArray, QC to identify those samples with a genomic integrity index (GII) of 3 or 4 suitable for sequencing (20) and Ampli1 whole-genome amplification were carried out as a service by Silicon Biosystems. Details of the exact numbers of single cells recovered and analyzed for each patient are given in Supplementary Table S1.

AmpliSeq reactions were set up using 10 ng WGA DNA, an average of 8 ng cfDNA and 10 ng FFPE tumor tissue DNA. All samples with sufficient DNA (including all individual cell samples) were sequenced using 2 AmpliSeq panels: an in-house 30-amplicon panel, designed using Ion AmpliSeq designer software (https://www.ampliseq.com), covering mutation hotspots in ESR1, PIK3CA, TP53, and ERBB2 (Supplementary Table S2) and the Custom Cancer Hotspot Panel v2 for Ampli1 WGA DNA (49 genes, >2,200 COSMIC mutations; Silicon Biosystems, see http://www.siliconbiosystems.com/fee-for-service and Supplementary Table S3). Three genes (PIK3CA, TP53, and ERBB2) were on both panels making a total of 50 genes surveyed for hotspot mutations. Overlapping mutations were reviewed in Integrated Genomics Viewer (IGV) if they were only detected on one panel. In total, the two panels surveyed about 2,200 COSMIC ID mutations.

Sequencing on a 316 chip using the Ion PGM and data analysis was as described previously (2). In brief, sequencing data were accessed through the Torrent Suite v.4.2.0. Reads were aligned against the human genome (hg19) using Alignment v4.0-r77189 and variants called using the coverageAnalysis v4.0-r77897 and variantCaller v4.0-r76860, respectively. Because of the sensitivity of the in-house panel, for cfDNA and FFPE tumor tissue, variantCaller was configured to call on high stringency, somatic variants with downsampling set to 6,000 and gen_min_coverage set to 6. COSMIC IDs were obtained using COSMIC v71 (http://cancer.sanger.ac.uk/cancergenome/projects/cosmic/; last cited on 21st February 2016). Variants in individual CTCs and white blood cells (WBC) were analyzed using a preset high stringency germline threshold, as recommended by the manufacturer, which is optimized for high-frequency variants and minimal false-positive calls. The allelic dropout rate (ADR) and false-positive rate (FPR) was calculated as described by Leung and colleagues (23). All mutations with a quality score below 25 were omitted, and all variants detected in the first or last 10 bases of an amplicon were omitted as likely mispriming events (2, 24). ANNOVAR (25) was used to annotate all private mutations with refGene ID, functional consequence (e.g., nonsynonymous), and functional predictions [using SIFT (26), Polyphen-2 (27), and MutationTester (28)]. All variants detected were also manually confirmed across all samples using the IGV 2.3 (29, 30). The mutant allele frequency (termed AF) was calculated as the proportion of total reads at a site, which contained the variant allele [e.g., if you have 200 reads in total and 8 of them have the variant, then the AF is (8/200) × 100 = 4%].

Droplet digital PCR (ddPCR) was used to validate selected mutations in ESR1 p.Y537C/N/S, p.D538G; PIK3CA p.E545K and H1047R; and KRAS p.G12D. Each assay was performed in duplicate using a Bio-Rad QX200 ddPCR system as described previously (31). Inventoried assays were used for detection of PIK3CA p.E545K (assay numbers dHsaCP2000075 and dHsa2000076) and p.H1047R (assay numbers dHsaCP2000077 and dHsaCP2000078) and KRAS p.G12D (assay numbers dHsaCP2000001 and dHsaCP2000002; Bio-Rad Laboratories) according to manufacturers' instructions. For the ESR1 mutations, a dual-labeled locked nucleic acid (LNA) probe strategy was used (see Supplementary Table S4 for sequences). Thermal cycling conditions for the ESR1 assays were: 10-minute hold at 95°C, 40 cycles of 95°C for 15 seconds and then 60°C for 60 seconds. Raw fluorescence amplitude was analyzed using the Quantasoft version 1.6.6.0320 software and used to obtain the fractional abundance for a given mutation. This was reported as the allele frequency to be consistent with NGS data. Calculation of the allele fraction was performed as described previously (32, 33). The total number of droplets (with and without DNA) was used to calculate DNA copies/μL, then dividing the number of mutant copies by the number of total DNA copies (mutant plus wild-type), and multiplying by 100 to give the percentage (allele fraction) of mutant DNA copies on the basis of Poisson distribution of positive to empty droplets (33).

Correlation analysis of total cfDNA levels, CA15-3, ALP, and CTC counts, was performed using a nonparametric Spearman correlation. P values were 2-tailed and considered significant if P < 0.05. Survival analysis was performed using multiple Cox regression as described previously (14) with each biomarker as a continuous time-dependent variable (Supplementary Table S5). Briefly, for each variable, a model was constructed using a counting process notation (i.e., start, end, event). In our study, start was taken as the date of the baseline blood sample, and the end was taken as an arbitrary cutoff of 01/12/2015.

This was a prospective study set up to identify patients with high CTC counts (≥100 CTCs per 7.5-mL blood) suitable for isolation of individual CTCs by DEPArray. The threshold of ≥100 CTCs per 7.5-ml-blood was based on a previous study by Polzer and colleagues (20), that analyzed individual CTCs from patients with similar high CTC counts but did not compare individual CTCs with matched cfDNA. We recruited 112 patients with MBC to the study (Table 1). Each had undergone multiple lines of treatment. Sixty-one patients were CTC-negative and 51 patients had ≥1 CTCs measured by CELLSEARCH (median, 5; range, 1–701). Serum CA15-3 levels were elevated (>30 U/mL; ref. 34) in 72 patients (64%), ALP was elevated (>100 IU/L) in 40 patients (36%), and both were elevated in 31 (28%). Total cfDNA was obtained for all samples (median, 2757 copies/mL of plasma; range, 30–115,724 copies/mL of plasma). The markers were significantly correlated (Spearman correlation analysis; Supplementary Table S6), except cfDNA and CA15-3 (P = 0.058). High CTC count [≥5 CTCs; HR, 2.8, 95% confidence interval (CI), 1.4–5.7; P = 0.005] and total cfDNA level (HR, 2.2; 95% CI, 1.3–3.6; P = 0.03) were significantly associated with poorer overall survival (Supplementary Table S6).

At the time of the study, we had established that adequate cells could be obtained from the DEPArray System for patients with high CTC counts (≥100 CTCs per 7.5-mL blood) based on a previous study (20). We therefore chose 5 patients with high EpCAM-positive CTCs for single CTC isolation by DEPArray: 4 had estrogen receptor (ERα)-positive metastatic disease and one was triple-negative. Clinical details for the 5 patients are given in Supplementary Data 1 and treatment timelines are in Supplementary Fig. S1. FFPE primary tumor tissue was available for 4 patients. Four had serial cfDNA and 1 had serial CELLSEARCH CTC samples (Fig. 1). Individual CTCs were recovered successfully for each patient. In total, 42 individual EpCAM-positive CTCs, 16 individual WBCs as germline and QC controls, 5 CTC pools and 6 WBC pools were isolated by DEPArray. Of these 40 individual CTCs, 15 individual WBCs and all cell pools passed QC checks (genomic integrity index ≥ 3; Supplementary Table S1) for sequencing (20).

All CTC and WBC samples, matching cfDNA and FFPE tumor tissue DNA were analyzed by deep amplicon sequencing (Fig. 1) using 2 panels covering about 2,200 COSMIC mutations in 50 cancer genes (for details of the custom 30-amplicon panel, see Supplementary Table S2, for full details of the Ampli1 custom cancer hotspot panel, see http://www.siliconbiosystems.com/fee-for-service). A high average read depth was obtained with both panels: 929× for the Custom Cancer Hotspot Panel v2 for Ampli1 WGA DNA (Silicon Biosystems; ref. 20) and 4,537× for the 30-amplicon custom panel (Supplementary Table S7). The ADR across individual CTCs and WBCs and the FPR for single-nucleotide variants (SNV) were 25.01%, SEM 2.68%, and 0.6%, SEM 0.32%, respectively (Supplementary Table S8), within the range reported previously (23). Mutations were only called in individual CTCs and WBCs if these were present in 2 or more single cells (23). Overall only 3 individual CTCs had no coverage at a specific mutated base (Table 2). In each case, the mutation was called in other CTCs and the matched plasma cfDNA.

Although we examined mutation hotspots in 50 genes across 2 amplicon panels by NGS, mutations were only detected by NGS in 4 genes in ESR1, PIK3CA, TP53, and KRAS. Two of these PIK3CA, TP53 were common to both panels. ddPCR was used as validation for ESR1 p.Y537C/N/S, p.D538G; PIK3CA p.E545K and H1047R, and KRAS p.G12D when sufficient remaining DNA was available.

Overall, 4 patients had mutations detected in PIK3CA, TP53, ESR1, and KRAS genes in cfDNA and individual EpCAM-positive CTCs (Fig. 2; Table 2). Levels of circulating tumor DNA varied. The mean plasma mutant allele frequency ranged from 0.3% to 32% and an increase in levels of circulating tumor DNA generally accompanied an increase in total cfDNA (Fig. 2). Mutational heterogeneity was seen between individual EpCAM-positive CTCs isolated from the same blood sample with the majority of CTCs having 1, 2, or no mutations detected. All mutations detected in individual EpCAM-positive CTCs were detected in the matched cfDNA (Table 2). PIK3CA and TP53 gene mutations were also present in the primary tumor DNA, whereas ESR1 and KRAS mutations were unique to cfDNA and EpCAM-positive individual CTCs (Table 2) suggesting potential clonal evolution. Private mutations of unknown significance were also detected in individual CTCs and WBCs (Supplementary Table S9); however, these were excluded as likely amplification artefacts, as they were not detected in the matched CTC pool, WBC pool, or primary tumor (Supplementary Tables S10–S13).

Patient CTCM155 (ER⁺, HER2⁻) had 2 blood samples taken 11 months apart (Supplementary Fig. S1). Plasma cfDNA was only available for the first sample, as insufficient blood was obtained at the second time point for both cfDNA and CTCs. EpCAM-positive CTCs were detected in the second sample after progressing on exemestane and everolimus combination therapy. The primary tumor and baseline cfDNA had a PIK3CA mutation (p.H1047R, AF 70% and 23%, respectively). A second lower frequency ESR1 mutation (p.E380Q, AF 3.9%) was also detected in cfDNA, suggesting that either a subclonal origin or this mutation was acquired on treatment. All 5 individual CTCs and the CTC pool had both mutations (Table 2, Supplementary Table S10).

Patient CTCM138 (ER⁺, HER2⁻) had a PIK3CA (p.H1047R) mutation (AF 25%) and low-frequency subclonal TP53 (p.R175H) mutation (AF, 1.5%) in primary tumor DNA. She was recruited to the study in 2014 some 10 years after diagnosis of breast cancer and having received multiple lines of endocrine therapy. Both mutations seen in the primary tumor were detected in baseline cfDNA (PIK3CA: AF, 32%; TP53: AF, 3.7%) as well as an acquired ESR1 (p.E380Q) mutation (AF, 24.2%). The second blood sample was taken 7 months later just prior to clinical disease progression. Levels of circulating tumor DNA increased from baseline, the 3 mutations persisted and a second ESR1 mutation (p.Y537C) was acquired at a low level (AF, 0.43%; Fig. 2A and D; Supplementary Table S11). Five individual CTCs were isolated from the second blood sample, 2 of which had the PIK3CA (p.H1047R) and ESR1 (p.E380Q) mutations but not the lower frequency ESR1 p.Y537C or TP53 p.R175H mutations (Table 2), although these could have been missed by sampling 5 individual CTCs.

Patient CTCM105 (triple-negative) was recruited to the study in March 2014 with progressive disease. No mutations were detected across the 50 genes in either the primary tumor or baseline cfDNA sample. A second blood sample was taken 11 months later. Six of 9 individual CTCs isolated from this sample had a KRAS (p.G12D) mutation that was also detected in the CTC pool (Fig. 2B and D; Table 2; Supplementary Table S12). Matched cfDNA had the KRAS p.G12D mutation (AF, 3.8%; Fig. 2D) and a low-frequency mutation in TP53 (p.P278R; AF, 2.3%). The low-frequency TP53 mutation might be missed by sampling 9 CTCs; however, it is also possible that 2 metastatic clones were shedding DNA in to plasma, one reflected in CTCs and the other not. As neither mutation was detected in the FFPE primary tumor tissue, these may either have been acquired on treatment or have arisen from a minor subclone missed by tumor sampling.

Patient CTCM292 (ER⁺, HER2⁻) presented with bone metastases in 2012 and was treated with tamoxifen. She subsequently relapsed and commenced letrozole, which was then switched to chemotherapy with paclitaxel, then epirubicin. Three serial blood samples were taken over a 12 = month period commencing upon letrozole. No mutations were detected in the baseline cfDNA sample, but 2 mutations were detected 6 months later while she was on letrozole (ESR1 p.D538G AF 28% and PIK3CA p.E545K 19.7%) and then reduced at 12 months on paclitaxel (AF, 0.27% and 0.17%, respectively; Fig. 2C). Both mutations were also present in individual EpCAM-positive CTCs isolated from the third blood sample (Fig. 2C, Table 2; Supplementary Table S13).

Last, the fifth patient (CTCM167: ER⁺, HER2⁻) was monitored over an 18-month period and had a marked increase in total cfDNA levels and CTC counts at disease progression (Supplementary Fig. S2). No mutations were identified in the 50 genes analyzed in FFPE primary tumor DNA. Moreover, there was no evidence for acquisition of mutations across the 50 genes in either cfDNA or individual CTCs from 2 serial blood samples (Table 2).

CTCs are an established prognostic indicator in MBC (7, 9, 35). A recent study has shown that it is possible to interrogate both cfDNA and pools of mutation-bearing CTCs from the blood of patients with MBC (36). Here, we present results of mutations in multiple individual CTCs and matched cfDNA in patients with MBC and high CTC counts. First, in 4 patients, we observed heterogeneity in mutations between individual EpCAM-positive CTCs. In all 4 patients, cfDNA profiles provided an accurate reflection of mutations seen in individual CTCs, and in 2 patients, cfDNA had more mutations than those found in CTCs. ESR1 and KRAS gene mutations were present in individual CTCs and cfDNA but were absent from primary tumor tissue. Therefore, these likely reflect either a minor subclonal mutation or were acquired with disease progression. Of note, the primary tumor, serial cfDNA samples, and multiple individual CTCs from 1 of the 5 patients (CTCM167) had no hotspot mutations detected across the 50 genes. Further analysis, such as whole-exome sequencing or array CGH, is needed to identify mutations and copy number variation in these samples.

Second, in 2 patients, mutations were detected in cfDNA [in TP53 (p.P278R) and in TP53 (p.R175H) and ESR1 (p.Y537C)] that were not seen in individual CTCs or the CTC pool isolated from the same blood sample. As these were low frequency (and likely subclonal) mutations, these may have been missed by sampling small numbers of individual CTCs. It is also possible that these were missed because of allelic distortion/dropout, which had an average frequency of 25.46% across the samples analyzed, similar to a previous single-cell genomics study (23). Of note, in a previous study (2), we also identified an ESR1 gene mutation unique to cfDNA that was not seen in a matched pool of >3,000 EpCAM-positive CTCs suggesting that 2 or more metastatic clones were shedding DNA in to plasma, one reflected in CTCs, one not.

Third, although we detected private mutations in single CTCs and single WBCs (mutations observed exclusively in single cells; ref. 37), these were excluded as likely polymerase error during WGA (38, 39) or sequencing error as all were absent from matched FFPE tissue, cfDNA, and CTC/WBC pools. Previous studies have suggested “census-based” sequencing of more than one CTC to distinguish private mutations from polymerase errors (38, 39); whereas other studies have compared results with ultra-deep sequencing of the primary and/or metastatic tumor (40) both approaches were performed here.

Although we sequenced about 2,200 COSMIC mutations across 50 genes in 5 patients, mutations were only detected in 4 genes (PIK3CA, ESR1, TP53, and KRAS) and only 1 or 2 mutations were detected in the majority of samples. This is consistent with a recent study, which used the same platform and a 50-gene panel (41). TP53 and PIK3CA are the most frequently mutated genes in breast cancer; however, a large number of other genes are less commonly mutated (42). Whole-exome or whole=genome analysis of the primary tumor/metastatic tissue would be required to determine additional mutations not covered by the targeted sequencing approach. However, metastatic tissue is frequently impossible to obtain in this group of patients.

ESR1 mutations have been linked to increased invasive and metastatic behavior (43). The ESR1 p.D538G mutation leads to ligand-independent activation of ERα and is acquired in patients who have received aromatase inhibitors. The p.E380Q mutation has also been reported, but according to Li and colleagues (44), remains sensitive to antiestrogens. In one patient, we saw evidence of mutational heterogeneity in ESR1, with p.E380Q at a higher level than p.Y537C in cfDNA, suggesting heterogeneous clonal responses to treatment, as reported in a recent study (45). The PIK3CA p.E545K mutation affects the helical region of the P110α catalytic domain (46), whereas the p.H1047R mutation affects the kinase domain and induces a multipotent genetic program in normally lineage-restricted populations at an early stage of tumor development leading to intratumoral heterogeneity (47).

Some studies suggest that ERα-positive/PIK3CA-mutant tumors or ERα-positive cancers with aberrant PI3K signaling are less responsive to antiestrogens than wild-type tumors (48, 49), but other studies have linked these mutations to improved outcome (50, 51), regardless of the type of endocrine therapy used (52). However, PIK3CA mutations have been significantly associated with lower rates of pathogenic response to anti-HER2 therapies (53). Of note the patient who had triple-negative breast cancer had different mutations, in TP53 and KRAS in cfDNA and KRAS only in individual CTCs, both of which have been implicated in resistance to cytotoxic chemotherapy.

It has been suggested that targeted sequencing of cfDNA could be used for monitoring of patients with cancer during treatment or while in remission and that single CTC analysis may then be needed to guide more appropriate therapy upon relapse (54). Here, we describe the relationship between mutations in multiple single EpCAM-positive CTCs and cfDNA isolated from the same blood sample in patients with high CTC counts (≥100 CTCs per 7.5-mL blood). Results from 5 patients with high CTC counts show that cfDNA sequencing finds more mutations than CTCs as well as mutations not seen in the primary tumor. Our CTC data highlight the mutational heterogeneity of single EpCAM-positive CTCs, supporting other recent studies (19, 20). In 2 of the 4 patients with mutations detected in CTCs, mutations were only detected in about half of the CTCs sequenced (Table 2). Together these results suggest that if EpCAM-positive CTCs were to be used, concurrent analysis of multiple CTCs may be necessary for clinical decision-making, particularly when metastatic biopsy is not possible. In aggregate, our results suggest that cfDNA reflects persisting EpCAM-positive CTCs in patients with high CTC counts and could enable monitoring of the metastatic burden for clinical decision-making. Finally, our results also suggest that total cfDNA levels are also an independent indicator of overall survival (P = 0.03), but a confirmatory study is needed in a larger series of patients. Further studies are ongoing to examine the relationship between CTCs and cfDNA using Parsortix filtration (55), an alternative marker independent approach for CTC isolation.

J.A. Shaw is a consultant/advisory board member for Qiagen. No potential conflicts of interest were disclosed by the other authors.

Conception and design: J.A. Shaw, D.S. Guttery, S. Ali, J. Stebbing, R.C. Coombes

Development of methodology: J.A. Shaw, D.S. Guttery, D. Fernandez-Garcia, K. Page, J. Stebbing

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): J.A. Shaw, D.S. Guttery, A. Hills, K. Page, B.M. Rosales, K.S. Goddard, O. Ogle, L. Woodley

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.A. Shaw, D.S. Guttery, A. Hills, D. Fernandez-Garcia, K. Page, R.K. Hastings, J. Luo, S. Ali, J. Stebbing, R.C. Coombes

Writing, review, and/or revision of the manuscript: J.A. Shaw, D.S. Guttery, K. Page, J. Luo, S. Ali, J. Stebbing, R.C. Coombes

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.A. Shaw, A. Hills, B.M. Rosales, K.S. Goddard, R.K. Hastings, J. Luo, L. Woodley, J. Stebbing, R.C. Coombes

Study supervision: J.A. Shaw, D.S. Guttery, J. Stebbing, R.C. Coombes

Other (clinical coordination): L. Woodley

Single-cell isolation was performed by Silicon Biosystems as a per fee service. We thank Dr David Moore for reviewing tissue sections, the Imperial and Leicester ECMCs, the Imperial BRC and the Imperial College Tissue Bank for supporting patient recruitment.

This work was supported by a Cancer Research UK Clinical and Translational Research Committee programme award (C14315/A13462X).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.