Implications of intratumour heterogeneity for treatment stratification

Authors

  • Andrew Crockford,

    1. Translational Cancer Therapeutics Laboratory, Cancer Research UK London Research Institute, London, UK
    Search for more papers by this author
    • These authors contributed equally to this study.
  • Mariam Jamal-Hanjani,

    1. Translational Cancer Therapeutics Laboratory, Cancer Research UK London Research Institute, London, UK
    2. UCL Cancer Institute, London, UK
    Search for more papers by this author
    • These authors contributed equally to this study.
  • James Hicks,

    1. Cold Spring Harbor Laboratory, NY, USA
    Search for more papers by this author
  • Charles Swanton

    Corresponding author
    1. Translational Cancer Therapeutics Laboratory, Cancer Research UK London Research Institute, London, UK
    2. UCL Cancer Institute, London, UK
    • Correspondence to: Charles Swanton, Translational Cancer Therapeutics Laboratory, Cancer Research UK London Research Institute, London WC2A 3LY, UK. e-mail: Charles.Swanton@cancer.org.uk

    Search for more papers by this author

  • No conflicts of interest were declared.

Abstract

Despite advances in the diagnosis and treatment of cancer, the majority of advanced metastatic solid tumours remain incurable. Differential gene expression, somatic mutational status, tumour-specific genetic signatures and micro-environmental selection pressures within individual tumours have implications for the success of predictive assays to guide therapeutic intervention. In this review we discuss the evidence for genetic and phenotypic heterogeneity and its potential implications for clinical decision making. We highlight areas of research that could be improved in order to better stratify patient treatment. We also discuss the predictive potential of patient-derived models of tumour response, including xenograft and cell line-based systems within the context of intratumour heterogeneity. Copyright © 2013 Pathological Society of Great Britain and Ireland. Published by John Wiley & Sons, Ltd.

Introduction

Cancer diagnosis and treatment has advanced significantly over the last decade, with an armoury of therapeutic strategies for oncologists to choose from [1, 2]. Despite this, metastatic tumours or tumours not amenable to surgical resection remain incurable, and response to treatment is often of limited duration. A compelling contributor to such unsatisfactory outcomes is intratumour heterogeneity (ITH). Light microscopy allowed pathologists such as Rudolf Virchow to report differences in tumour morphology in the 1800 s and later, in the nineteenth century, haematoxylin and eosin (H&E) tissue staining, pioneered by Wissowzky, allowed scientists such as Von Hansemann and Broders to identify heterogeneous tumour tissue characteristics, such as asymmetrical mitoses and anaplasia [3, 4]. Studies from the early 1970s demonstrated functional phenotypic heterogeneity, such as differential drug metabolism, drug efficacy and metastatic potential, highlighting the important clinical implications of ITH [5-7]. Modern genomics have provided an insight into the degree of tumour heterogeneity at nucleotide resolution and have revealed differential phenotypes and prognostic signatures both within single biopsies and between biopsies of the same tumour [8]. This review discusses the evidence for ITH, focusing on its impact on clinical decision making and the use of patient-derived xenograft models to predict therapeutic outcome.

Genetic intratumour heterogeneity

The tumour clonal evolution model proposed by Peter Nowell (1976) provides a framework for ITH based on Darwinian selection over time, generated by genetic instability and genetic drift [9]. Premitotic and mitotic dysfunction can result in chromosomal instability (CIN), whilst mutagenic exposure and aberrant DNA repair can result in cumulative somatic mutations [10-13]. Nowell proposed that genomic instability could produce subclonal populations that can exist within a single tumour, and that certain subclones may be out-competed by others, due to lack of fitness under a selection pressure. These favoured subclones may harbour growth advantages, resulting in selection and expansion. Recent studies of haematological malignancies have demonstrated that this process can be non-linear, resulting in branched evolutionary growth, generating genetic diversity [14-16]. Ding et al proposed two major patterns of evolution during treatment associated with drug resistance in acute myeloid leukaemia (AML), in which there is either the expansion of pretreatment-resistant subclones harbouring advantageous genetic and/or phenotypic traits, or the gain of resistant mutations over time in an otherwise non-resistant subclone [15]. Schuh et al found multiple subclones in pretreated chronic lymphocytic leukaemia (CLL) cases, demonstrating that clonal mutational profiles can change over time and can be heterogeneous between patients [16]. Branched evolutionary growth has also been documented in solid tumours [8, 17, 18] and is apparently consistent with the clonal evolution model. Through multi-region sequencing of clear cell renal carcinomas (ccRCCs), we have found that 63–69% of identified mutations in two tumours were not present in all regions sampled and that genetic heterogeneity resulted in differential phenotypes in terms of mammalian target of rapamycin (mTOR) signal transduction activity [8]. Spatial separation of subclones may be encouraged by the presence of physical tissue barriers between tumour regions, limiting subclone intermixing, and mutations with near-equal fitness in individual subclones that may hinder clonal succession and selective sweeps. Such divergence may exacerbate the development of heterogeneous somatic copy number alterations (SCNAs) and somatic mutations within a tumour, leading to multiple regionally separated phenotypes [19]. The important principle is that multiple subclones can branch from a single lineage (as defined by SCNA breakpoints or point mutations) and flourish independently in the primary tumour. The accumulation of new events in these subclones, on top of those pre-existing in the lineage, makes the strong prediction that continued tumour growth is not the result of repeated initiation by stem-like cells, but results from a combination of proliferation and selection on the population as a whole.

Clinical implications of heterogeneous gene amplification

Gene expression and SCNAs can be readily detected by a variety of methods available in the clinical setting. Early founder SCNAs or actionable somatic mutations will likely be present in the majority of tumour subclones, present throughout the tumour, and therefore act as potential therapeutic targets and clinical biomarkers. However, there is increasing evidence demonstrating that genetic ITH results in differential spatial and temporal biomarker expression within a single lesion, suggesting that a single biopsy at a defined time point may not be representative of the tumour's genetic potential [20, 21]. Geyer et al demonstrated heterogeneous SCNAs in breast cancer that correlated with distinctive cellular morphology, suggesting a phenotypic consequence of genetic ITH [22]. Breast cancer biopsies are commonly assessed for oestrogen receptor (ER), progesterone receptor (PR) and human epidermal growth factor receptor 2 (HER2) status. These correlate with prognosis and determine choice of therapy; for example, the use of tamoxifen or trastuzumab in ER-positive or HER2-positive tumours, respectively [23]. Immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) analyses have identified heterogeneity of HER2 expression and amplification, suggesting that sampling bias may contribute to false-negative results [24-28]. FISH and IHC studies by Niikura et al demonstrated that 24% of primary HER2-positive tumours had HER2-negative metastases, indicating the potential need for biopsies from metastatic sites at recurrence to re-assess HER2 status [29]. Lindstrom et al demonstrated temporal heterogeneity of ER, PR and HER2 status between primary and metastatic breast tumours, reporting discordant frequencies of 32.4%, 40.7% and 14.5%, respectively [30]. Other studies in breast cancer have shown heterogeneity in markers such as p53 and mib-1 [31], and in a panel of signalling proteins using reverse-phase protein arrays [32]. HER2 over-expression has also been shown to display heterogeneity in mucinous ovarian and gastric tumours, providing evidence for HER2 ITH in other tumour types [33-35].

Glioblastomas (GBMs) have been shown to have mosaic patterns of gene amplification, frequently observed in the EGFR and PDGFRA genes [36, 37]. Therapeutic response to single-agent inhibitor therapy can be poor and may be associated with the heterogeneous amplification of such genes [38-40]. Szerlip et al found that GBM tumours showed heterogeneous SCNAs for PDGFRA and EGFR. FISH analysis of tumour tissues identified cells with either EGFR or PDGFRA amplification, and in most cases a small population harbouring both events (dual-amplified clones). All tumours assessed contained regions where EGFR was exclusively amplified, whereas this was not the case for PDGFRA. Regions containing PDGFRA amplified cells were always accompanied by EGFR or dual EGFR + PDGFRA amplified cells (Figure 1) [37]. Experiments carried out on cell lines derived from these tumours also demonstrated the presence of subpopulations with dual amplified EGFR + PDGFRA (100% EGFR and 25% PDGFRA). Treatment with dual inhibitor combinations of gefitinib and imatinib resulted in repressed downstream signalling, whereas treatment with a single agent only induced weak pathway down-regulation at the population level [37]. In keeping with these findings, Sottoriva et al demonstrated that the most common SCNAs in 31/38 GBMs were not ubiquitous throughout all tumour regions [41]. Common genes implicated in early, intermediate and late phase tumour growth (eg EGFR, PDGFRA and GLUT9, respectively) were identified. Spatial and temporal ITH was demonstrated, with early driver events undergoing selective genetic sweeps before branched events, resulting in diversity [41]. Taken together, these data suggest that targeting only a single amplification may prove therapeutically ineffective, due to the presence of subclonal driver events present in some cells and not others.

Figure 1.

Potential GBM EGFR and PDGFRA subpopulations. EGFR amplification may exist exclusively, suggesting the possibility that single biopsy may miss regions with abnormal EGFR and PDGFRA SCNAs. Over-expression of EGFR and PDGRA in a single lesion may result in ineffective gefitinib and imatinib monotherapy.

Targeted agents can select for the up-regulation of an adjacent oncogenic pathway, resulting in therapeutic resistance. Katayama et al revealed that crizotinib treatment of EML4–ALK sensitive non-small cell lung cancer (NSCLC) tumours could select for KIT amplification, thereby resulting in decreased sensitivity. However, ‘pathway rewiring’ can be a heterogeneous event, as different crizotinib-resistant regions in the same tumour utilize EGFR over-expression in contrast to KIT [42]. This indicates the importance of evolving heterogeneous tumour rewiring programmes during treatment, which may require deeper appreciation to anticipate the tumour's next evolutionary move in order to achieve more effective and prolonged tumour responses. Since differential rewiring may occur within distinct tumour regions, more efficient tumour sampling methods at different time points during the disease course may be necessary to overcome pathway adaptation. Sharma et al demonstrated non-genetic based drug resistance in lung cancer cell lines after EGFR TKI and cisplatin treatment [43]. Stochastic epigenetic remodelling led to the production of resistant cells, potentially due to pathway rewiring. Withdrawal of drug treatment resulted in resensitization over a defined period of time. This mechanism can explain retreatment responses observed in the clinic, and how drug resistance can occur in an irregular fashion within a population, independently of a defined genetic mechanism. It is therefore apparent that the detection and prediction of non-genetic drug resistance is a challenging feat, given the absence of genetic markers.

Clinical implications of heterogeneous somatic mutations

Drug resistance induced by somatic mutations or somatic copy number alterations (SCNAs), leading to oncogene over-expression, can be detected in order to make an informed treatment decision. Low-frequency somatic mutations, insertions and deletions, resulting in functional modification of drug targets, require greater resolution and sequencing depth. KRAS activation acts as a biomarker for anti-EGFR therapy resistance in colorectal cancer and is therefore sequenced in order to identify activation status prior to treatment stratification. However, the distribution of mutated active alleles can be heterogeneous within a single lesion and between metastases. Baldus et al investigated primary and metastatic colorectal adenocarcinomas and showed heterogeneity for KRAS-activating mutations versus wild-type in different regions of the same primary tumour (8%) as well as between paired primary and metastatic tumours (31%), implying that a single biopsy may not indicate whether anti-EGFR therapy is likely to be therapeutically effective [44]. In contrast, a larger study by Vakiani et al showed > 90% concordance for five genes (KRAS, NRAS, BRAF, PIK3CA and TP53) between matched primary and metastatic colorectal tumours [45]. A possible reason for the differences between these two studies may be different technical sensitivities and pre-analytical variables (eg tissue processing) [46].

Heterogeneity of drug target somatic mutations may also lead to variable drug efficacies among tumour subclones, a mechanism commonly observed in chronic myeloid leukaemia (CML). Subpopulations of imatinib-resistant cells with kinase domain mutations have been identified in patients prior to treatment with imatinib, which, after several months of treatment, can expand from a low- to a high-frequency subclone as a result of selection pressure [47]. Solid tumours also harbour target-resistant subclones. A common mechanism of resistance to EGFR inhibitors in NSCLC is the T790M mutation. Su et al demonstrated that patients with underlying resistant subpopulations harbouring the T790M mutation prior to treatment, have shorter progression-free survival times with EGFR-targeted therapy [48]. The findings of Taniguchi et al and Chen et al demonstrated EGFR mutational heterogeneity in different regions and between paired metastases of the same tumour [49, 50]. Similarly, mutations in mTOR can also be heterogeneous within a single lesion. We performed next-generation sequencing of ccRCC primary tumour sites and metastases, revealing activating mTOR mutations in seven of eight primary sites, with no evidence of the mTOR mutation in three metastatic biopsies [8]. This is noteworthy, since preclinical studies have suggested that mTOR pathway activation (eg phosphorylated S6 kinase) might provide a biomarker for rapamycin analogue drugs [51]. Heterogeneous pathway activation within primary and metastatic sites will certainly confound such approaches.

Extensive genetic ITH raises the possibility that individual subclones can have mutations in the same target but at different amino acid residue positions, resulting in variable drug affinity and sensitivity between subclones of the same tumour. The prevalence of the EML4–ALK fusion event in unselected NSCLC is in the range 1.6–8.6% and can be treated by crizotininb [52]. Choi et al investigated resistance in response to TKIs in NSCLC with the EML4–ALK fusion protein [53]. Seventy-three cDNA clones derived from a pleural effusion aspirate were sequenced, revealing novel somatic mutations in the EML4–ALK kinase domain (46.6% G→A and 15.1% C→A). Although the PCR analysis spanned both aberrations, they were never found in the same clone, suggesting that they evolved from distinct separate subclones [53]. Katayma et al demonstrated that crizotininb treatment acts as a catalyst for acquired multi-drug resistance by selecting clones with variable EML4–ALK mutations [42]. Four different mutations were identified after treatment (three missense, L1196M, G1202R and S1206Y, and one insertion, 1151Tins), each producing varying degrees of sensitivity to crizotinib and next-generation EML4–ALK TKIs. Induced crizotinib-resistant cell lines CR1 and CR2, cultured in parallel, were found to have different resistance mutations (L1196M with amplification and 1151Tins, respectively) [42]. In cell lines with no EML4–ALK mutations, EGFR receptor and ligand over-expression led to phosphorylation of downstream targets AKT/ERK in the presence of crizotinib, inducing resistance. Pathway and growth suppression were only observed after dual administration of crizotinib and gefitinib [42]. As multiple mechanisms of resistance can arise independently from the same cell line, it is not surprising that tumours also have such capacity. This may result in tumour subclones containing simultaneous, yet distinct, mutations within the same target. Ultra-deep sequencing analyses from tumours prior to treatment and at resistance may give a more definitive picture of the extent of target heterogeneity in EML4–ALK tumours and the potential resistance mechanisms that may occur over the course of therapy.

Strategies to overcome the therapeutic challenges posed by subclonal heterogeneity have been proposed, seeking to exploit different phenotypes of resistant clones [54, 55]. Gatenby et al have suggested dosing regimes that would encourage competition between resistant and sensitive clones and promote disease stability [54]. Using a primary human melanoma xenograft model, in which vemurafenib-resistance was selected for by continuous administration of the drug, Thakur et al demonstrated that resistant tumours were drug dependent for their continued proliferation [55]. Drug cessation led to the regression of tumours, implying that drug-resistant subclones within the tumours had a fitness disadvantage in the absence of the drug. Such drug dependency may be exploited through an intermittent dosing schedule, as Gatenby et al suggested, thereby preventing drug resistance (Figure 2).

Figure 2.

Clonal selection and drug holiday strategy. A tumour made up of multiple subclones is treated with first-line therapy. Treatment selects for resistant subclones and these out-compete sensitive clones due to a fitness advantage, a process known as a genetic 'bottleneck'. New clones can arise over time as a result of tumour evolution. If a drug holiday is introduced, a shift in relative fitness occurs. The mechanisms adopted by the resistant clones can be detrimental in the absence of drug (eg BRAF over-expression). This results in re-expansion of sensitive clones that suppresses the resistant clones. The second round of treatment promotes resistant clone outgrowth and the cycle continues. Drug holiday cycles may promote stable disease for longer and delay the emergence of a completely resistant tumour.

Consequences of tumour microenvironment heterogeneity

The tumour microenvironment can promote or inhibit proliferation and survival, with potential consequences on cellular phenotype that can lead to differential sensitivities to chemotherapeutics [56, 57]. Blood vessel distribution throughout a tumour can be disorganized and heterogeneous, resulting in differential uptake of chemotherapeutic agents by cancer cells. Galmarini et al investigated the effects of heterogeneous blood perfusion in 25 patients with head and neck tumours, treated with a cisplatin–bleomycin regimen. Histopathology demonstrated that 52% of highly perfused regions achieved complete tumour regression, compared with just 4% of those with poor vasculature [58]. As well as limiting drug delivery, poor blood perfusion results in low oxygen concentrations (hypoxia) that may result in the selection for p53–/– clones [59], up-regulation of the drug efflux pump MDR1 and altered pH [60, 61]. Whilst these examples are a gross over-simplification of the degree of genetic and phenotypic adaptations to diverse microenvironments, in such a setting, pro-apoptotic agents could potentially be less effective due to an absent p53 pathway, with low drug concentrations further reduced by efflux. Drugs such as vincristine and doxorubicin, which are weak bases, can become ionized in this acidic environment, potentially reducing cellular drug uptake.

The degree of interaction between the tumour and the extracellular matrix (ECM) is also important. Muranen et al demonstrated ECM-dependent survival in ovarian cancer cell spheroids treated with mTOR inhibitors, using 3D culture techniques [62]. ECM-unattached cells underwent apoptosis when exposed to mTOR inhibitors, since they are dependent on PI3K pathway activation to survive. In contrast, ECM-attached cells treated with an mTOR inhibitor induced EGF and integrin signalling, facilitating an adaptive survival phenotype in response to PI3K pathway shutdown. Treatment with anti-EGF and integrin compounds resulted in better treatment response in animal models [62]. Overall, increasing evidence supports the likely influence of heterogeneity of stromal contact in therapeutic response, potentially selecting for cells that can tolerate drug treatment and contribute to disease progression.

Intratumour heterogeneity and predicting clinical response using cell lines and xenografts

Genetic, phenotypic and tumour microenvironment heterogeneity is likely to have implications for the ability to recapitulate accurate cancer biology in cell lines, cell line xenografts (CLXs) and patient-derived tumour xenograft (PDTX) models. Cell lines derived and cultured from a single biopsy may not represent the overall phenotype of the original tumour. Selection of a low-frequency subclone within a biopsy, for example, or spatial sampling bias, could result in a discordant phenotype [5, 7, 63, 64]. Barranco et al created melanoma cell lines derived from a single melanoma lesion [5]. Cell lines were treated with the pro-drug arabinosylcytosine (cytotoxic to S-phase cells) in a manner such that cytotoxicity should correlate approximately with the proportion of cells in S-phase. In this instance, one clone correlated positively where other clones were less sensitive to treatment, even though the S-phase fractions were similar. The amount of arabinosylcytosine activating enzyme was reduced by approximately 50% in the least sensitive clone, possibly reducing the amount of activated drug and thereby increasing S-phase survival [5]. In another study, Wang et al derived two cell lines from a single bile duct carcinoma and showed differential morphology, growth rate, tumourigenicity, hypoxia tolerance and therapeutic sensitivities between clones [64]. Similarly, Heppner et al created three cell lines from a single murine adenocarcinoma, and tested drug sensitivity in vitro and in vivo. All three cell lines displayed differential cytotoxic therapy sensitivity, as well as differential drug-induced metastatic potential [7]. These studies demonstrate how ITH can give rise to phenotypically different cell lines derived from a single tumour biopsy, and suggest that their use as a predictor for therapeutic efficacy may be limited.

PDTX models have recently been proposed as an improvement over cell lines and CLXs for preclinical drug sensitivity analyses, and as a drug efficacy prediction tool [65]. Typically, small human tumour fragments (∼3 mm3) are obtained from surgically resected tumours or metastatic biopsy specimens and implanted, either subcutaneously or orthotopically, into immunodeficient mice and propagated over a series of generations [65]. Daniel et al compared gene expression among PDTXs, their derivative cell lines and CLX [66]. Expression profiles between PDTX and CLX models differed substantially (152 genes), with greater differences between cell lines and PDTXs (315 genes). These data suggest that cell lines and CLX models may differ substantially from the original PDTX. Furthermore, when tumour-specific genes were compared between the primary tumour and the PDTXs, only one of three models closely resembled the original tumour genetic signature [66]. In contrast, Guenot et al noted stable gene expression and similar histopathology of seven PDTXs compared to tumours [67]. Julien et al used a colorectal cancer (CRC) PDTX panel using tumours from different patients, and in some cases two regions from the same tumour, to demonstrate the use of such models as a drug discovery tool [68]. PDTXs derived from 49 patients were treated with four chemotherapeutic agents (5-FU, oxaliplatin, ironotecan and cetuximab) and treatment response assessed by tumour growth inhibition [68]. In terms of drug response, PDTXs from different patients showed intertumoural heterogeneity. In cases where multiple samples from the same tumour were engrafted, different drug responses were apparent, demonstrating intratumoural heterogeneity.

A number of studies have suggested that the PDTX method may be representative of disease course and adopted for clinical use [69-71]. Fichtner et al established colorectal cancer PDTXs using surgically resected specimens from 15 patients [69]. Five of these patients went on to receive chemotherapy (5-FU, oxaliplatin and irinotecan) for synchronous metastases. The authors suggested a correlation between the PDTX and patient treatment response, based on palpable tumour size in the mice and radiological imaging in the patient. Villarroel et al used a PDTX to predict treatment response in a patient who had undergone surgery for pancreatic cancer [71]. Having consented prior to surgery, a PDTX was established whilst the patient received chemotherapy with gemcitabine. The PDTX was treated via the peritoneum with either a control, mitomycin C, cisplatin or gemcitabine. After 4 months of gemcitabine treatment, based on a favourable PDTX response to mitomycin C, the patient was treated with five cycles of this agent, resulting in maintained disease response for 22 months until further progressive disease. At this point, the patient was given two further cycles of mitomycin C with response, but complicated by renal failure. The patient was then treated with three cycles of cisplatin, based on the PDTX findings, again resulting in a response. In another study, inclusive of the PDTX derived by Hidalgo et al, PDXT models for 14 patients with refractory solid or early-stage, poor-prognosis tumours of different types were established and treated with 63 drugs in order to predict the choice of chemotherapy drug [70]. Eleven patients were treated with 17 treatment regimens based on their corresponding PDTXs, and 15 of these regimens resulted in durable partial remissions.

Although there are benefits to using PDTX models to investigate cancer biology and pharmacology in the laboratory, it should be noted that they are subject to the same sources of discordance described above for cell lines and CLX. Using them to predict patient clinical responses may require a more cautionary approach in light of such frequent genetic and functional intratumour heterogeneity. A further potential caveat is the time required to successfully produce models for drug testing. This can often take months, during which time the patient's tumour genetic and microenvironment parameters may have changed (Figure 3). Clinical therapeutic regimens can be complex and difficult to perform in mice, and the differences in drug administration may also result in differences between the PDTX model outcome and actual clinical response. PDTX models also require immune-compromised mice to prevent rejection of the engrafted tissue. Increasing evidence in support of the role of the immune system in determining treatment response suggest further caveats to this approach [72]. Cytotoxic drug treatments can induce inflammatory mediators, such as ATP, resulting in activated local dendritic cells and macrophages, leading to CD4+ and CD8+ T cell response [72]. Ma et al demonstrated the importance of immune system-dependent activity of anthracyline-based chemotherapy using murine fibrosarcomas established in syngeneic mice [73]. ATP released by dying tumour cells stimulated the local differentiation of dendritic cells, resulting in tumour antigen presentation to CD8+ T cells and aiding anti-tumour immune response. The immune response can also suppress the antitumour effect produced by therapeutics, for example, macrophage infiltration or specialised CD4+ cells that prevent autoimmunity (Tregs) [72, 74]. Importantly, immune-compromised PDTX models are unable to reproduce the spectrum of immune-mediated cell death and drug resistance observed in human patients, further complicating their use as drug efficacy prediction tools.

Figure 3.

PDTX models and clinical response predictions. Region 1 from a patient's tumour is engrafted into an immune-compromised mouse, propagated and expanded. When enough mice are available for agent testing, the mice are administered a test compound and tumour response is measured. Drugs that induce responses are considered to be clinical candidates for the patient. The time course of the procedure presents a potential problem, as the patient's tumour will continue to evolve through time. Progressive tumour evolution may result in a discordant phenotype relative to the PDTX, and possible treatment failure. In addition, if only one tumour region is propagated, other regions may have different genetic and microenvironments, leading to differential drug response.

Voskoglou-Nomikos et al investigated the correlation between phase II clinical trial data (breast, colon, NSCLC and ovarian cancers) and the treatment response of cell lines, CLXs and PDTXs to 31 cytotoxic therapies [75]. This study demonstrated that CLXs were not predictive of clinical outcome and panelled PDTX models (two or more PDTXs used to predict response) were only predictive for non-small cell lung and ovarian cancer. Single PDTX models were also not predictive of phase II clinical activity. In a single patient, Ding et al analysed DNA from peripheral blood, primary tumour, brain metastasis and a xenograft derived from the primary tumour [76]. Next-generation sequencing revealed shared mutations in all three samples. However, the metastasis contained mutations not present in the primary, and the xenograft retained all primary tumour mutations but overall had a mutation profile resembling the metastasis. The narrowing of the mutational allele frequency range between the primary tumour and metastasis and xenograft implies the selection of distinct subclones in the metastatic and transplantation process. The similarities between the xenograft and primary tumour, in terms of mutations and SCNAs, suggests that early-passage PDTXs may be of use in functional and therapeutic studies in order to understand primary tumour behaviour. Arguably, xenograft modelling of metastatic disease, in the context of advanced treatment-refractory cancer, is of greater importance. Whether a PDTX model derived from a single metastatic site represents true tumour phenotypic heterogeneity, given the potential for heterogeneity of SCNAs and somatic mutational status between metastases in an individual patient, is unknown and may limit interpretation of such models.

Although some studies support the use of PDTXs as predictive tools for clinical efficacy, they consist of a small number of cases and are anecdotal in nature. There is therefore a lack of level 1 evidence supporting PDXT-defined treatment strategies in providing benefit above standard of care. In order to demonstrate the potential for using PDTX models in drug efficacy prediction, a sufficiently powered randomized clinical trial would be required, comparing progression-free survival times in patients treated with the best supportive care compared to those treated based on PDTX models. Given the emerging challenges of ITH for clinical prediction models to hold promise, they may have to be integrated with extensive, ultra-deep genomic profiling and phenotypic analyses at different time points in the disease course. There are clear limitations prohibiting this level of analysis in terms of cost and resource utilization in the clinical setting.

An alternative means of avoiding sampling bias associated with single biopsies, CLX and PTDX models, and therefore potentially ineffective clinical decision-making, may be found in a less invasive and potentially more comprehensive method; the genetic analyses of circulating tumour cells (CTCs) and cell-free DNA (cfDNA) retrieved from blood samples. Although clinical CTC analysis is currently limited to enumerating epithelial cells in the blood using the FDA-approved CellSearchTM technology, multiple methods in development are designed to isolate cancer cells in ways that will make genomic analysis possible, including analysis of individual cells at the transcript level [77-80] or genome-wide through combinations of microarrays [81] and next-generation sequencing. Heizler et al [82] used a combination of microarray and DNA sequencing analysis on single cells retrieved from a CellSearch chip and showed that they were derived from the primary breast tumour lineage. At ASCO 2013 two groups presented single-cell SCNA data obtained by next-generation sequencing of prostate cancer CTCs isolated by flow cytometry [83] or retrieval from stained slides [84], and showed both high degrees of SCNA heterogeneity and response to therapy. Similarly, whole-genome sequencing [85, 86] of cfDNA has confirmed that circulating DNA maintains the mutation spectrum of the primary cancer but also evolves during treatment. On the assumption that CTCs and cfDNA are shed randomly from the primary or multiple metastatic sites, these ‘fluid biopsy’ methods [87] have the potential to provide sequential, real-time monitoring of the genetic state of cancer during therapy and to identify and characterize newly arising drug subclones that may signal resistance and the need to alter treatment.

Conclusions

The presence of extensive genetic and phenotypic ITH warrants greater consideration with respect to diagnostic and prognostic tools. Heterogeneous subclonal events may be important for the emergence of resistant subclones and subsequent therapeutic failure. Multi-region tumour sample acquisition, combined with ultra-deep sequencing analyses of individual biopsies, may help address the landscape of heterogeneity, but only if sufficient tissue is available and detection methods are sensitive enough. High-throughput, deeper sequencing techniques may allow greater sensitivity to support the targeting of multiple subclonal somatic events whose emergence would otherwise result in therapeutic resistance. The identification of multiple targets within a single tumour may allow physicians to prescribe combination therapy, but the limitations of this approach are toxicity and the paucity of available small molecule and antibody-based approaches to target the myriad of potential driver events operating in single tumours. Therefore, drug combination approaches need to be researched extensively in order to define drug safety profiles. The need for multi-region tumour sampling is also relevant to biomarker discovery. ITH may present considerable problems in this regard, since potential spurious chance associations between heterogeneous genes and therapeutic response can be made when large datasets are analysed based on isolated tumour samples [88]. This method does not provide functional evidence for the observed response and implies that sampling bias may hinder validation in a heterogeneous tumour, therefore impeding the clinical application of biomarkers discovered using such techniques.

The interest in PDTX models in research is likely to increase greatly and provide more vigorous preclinical testing than traditional cell line methods. However, we argue that randomized clinical trial evidence is required to fully define the role of PDTX analysis in predicting cancer therapeutic response. The combination of PDTX models with high-throughput genomics may improve predictive accuracy, although ways to streamline the process are required for clinical application, ideally prospectively tested within clinical trials. Ultimately, a greater appreciation of the scale of ITH may accelerate clinical trial design and improve clinical decision making.

Author contributions

AC, MJH, JH and CS wrote the paper.

Ancillary