The clinical effectiveness and cost-effectiveness of genotyping for CYP2D6 for the management of women with breast cancer treated with tamoxifen: a systematic review

Enzymes, genes and pharmacogenetics

Differences in the response of individuals to the same drug at the same dose may occur as a result of interindividual differences in enzymes (e.g. CYP2D6) responsible for metabolising the drug. These differences may be inherited and occur as a result of differences in the genes (e.g. CYP2D6) that encode the enzyme.

In humans, each gene is composed of two alleles, one inherited from each parent, and a person may have two copies of the same allele (homozygous) or one copy of two different alleles (heterozygous). Alleles that differ from the normal or common form are known as polymorphisms [variant (vt)], while a normal allele is referred to as wild type (wt). It is from these differences that an individual’s genotype is derived, for example the homozygous wt (i.e. wt/wt) genotype.

A phenotype is the observable physical trait of an organism, which, in pharmacogenetics, relates to an individual’s reaction to a drug, usually as a result of the way in which the drug is metabolised. The phenotype is largely determined by the overall genetic make-up of a person, although it may also be influenced by environmental factors (e.g. diet and smoking).

The cytochrome P450 (CYP450) enzyme system, to which CYP2D6 belongs, has been identified as a major metabolic pathway for many drugs and a source of interindividual variability in patient response. It is believed to play a prominent role in the way in which tamoxifen (TAM) is metabolised and thus may explain differences in responses in individual patients to the same dose as it is known that TAM is metabolised to its active metabolites (which are thought to affect patient response, rather than TAM itself) by a number of CYP450 enzymes (including CYP2D6).

Based on studies that have examined the urinary metabolic ratios of drugs such as debrisoquine and/or dextromethorphan to their metabolites (4-hydroxydebrisoquine and dextrorphan, respectively), an association between CYP2D6 genotypes (genetic make-up) and phenotypes (response to treatment) is believed to exist. It is thus also believed that patients experiencing a normal response at a normal dose of TAM would be CYP2D6 extensive metabolisers (EMs). These individuals are thought to be homozygous for the wt allele. Patients experiencing reduced clinical effects owing to deficient alleles are referred to as poor metabolisers (PMs) and are thought to be homozygous (and possibly heterozygous) for the vt allele.

However, there are a number of different vt alleles, some which result in decreased enzyme activity and others that result in a complete lack of enzyme activity (i.e. the differing extent to which the drug is metabolised). PMs must possess at least one of these complete lack of function alleles (e.g. *4; see Table 1).

Patients are sometimes also considered to be intermediate metabolisers (IMs) if clinical effects lie somewhere between EMs and PMs. Generally, IMs are thought to possess at least one decreased activity allele (e.g. *10). Patients are also sometimes considered to be ultrarapid metabolisers (UMs) when there are multiple copies of an allele (e.g. *2 × 2, *2 × 3, etc.). However, multiple copies of an allele do not necessarily result in increased activity. Furthermore, for CYP2D6 there is no uniformly agreed way in which to relate genotype to phenotype. While it is acknowledged by all that a patient with a wt/wt genotype would be an EM and a patient with the *4/*4 genotype would be a PM, some would also classify patients heterozygous for these alleles differently, for example a patient with the wt/*4 genotype could be considered an EM, IM or PM.

Genotyping for CYP2D6

There is growing anticipation that genotyping for CYP2D6 may be used to assist in treatment decision-making. A number of these tests have been developed and are described in the literature, and have been used for a wide range of drugs and diseases, not just TAM and breast cancer. However, not all tests will be the same.

Table 2 presents examples of three possible CYP2D6 tests that could be used. As can be seen, in test A, patients are simply tested for *4. Those who are found not to possess *4 are considered to be wt. Even with this simple test, it is possible to classify a patient with the wt/*4 genotype in three different ways: EM, IM or PM. As the number of alleles tested for increases (test B), the chances of detecting IMs and/or PMs are increased and the classification is complicated somewhat by the inclusion of the decreased activity allele (*10) in test C.

These examples can also be used to show that the construction of phenotypes can change the way in which results are interpreted. Thus, if we had 30 patients and found from test A that 15 had the wt/*4 genotype and that 10 of these patients had a side effect from taking TAM that was not detected in the other 20 patients then, depending on which classification we used, we would describe these patients as being IM, PM or EM. Consequently, we would assume from this sample of patients that there was an association between the phenotype and the side effects.

In addition, these examples also show that as a larger number of alleles are tested for, the chances of detecting IMs and PMs are increased. For example, a patient with the *3/*5 genotype identified by test B and labelled a PM would not have been detected as a PM by test A, which did not test for these two alleles, and so he or she would have been classified as wt/wt, i.e. EM. Test C may also be unable to identify this patient as a PM, not because of the number of alleles tested but because of the types of alleles tested, here this patient being identified as *3/wt. Thus, the types of alleles tested for are just as crucial as the number tested.

To date, the majority of these tests are designed bespoke, ‘in house’, for specific research projects, often using commercially available technologies such as TaqMan^® (Roche Molecular Systems). The only commercially available complete test that is available and used in clinical practice, albeit rarely, is the AmpliChip^® (Roche Molecular Systems), which tests for 33 different alleles.

Incidence/prevalence and health impact

Breast cancer is the most common cancer affecting women in the UK. In England and Wales, in 2007, around 45,000 new cases of breast cancer were diagnosed1 and there were nearly 11,000 deaths due to breast cancer. 2 Breast cancer incidence rates increase with age; around 80% of breast cancers occur in women aged > 50 years, and women have a one in nine lifetime risk of developing breast cancer. 3 Breast cancer prevalence is around 172,000 women in the UK according to the most recently published data. 4 This relatively high prevalence rate has been attributed to high incidence rates combined with 5-year survival rates of > 75%. 5

Aetiology

Breast cancer is the uncontrolled, abnormal growth of malignant breast tissue affecting predominantly women. The strongest risk factor for breast cancer (after gender) is age – the older the woman, the higher her risk – but other genetic and hormonal risk factors have also been identified in the aetiology of breast cancer. 5,6

Carriers of the breast cancer 1 (BRCA1) or 2 (BRCA2) gene mutations7,8 and women with a family history of breast cancer9 both have an increased risk of developing breast cancer. Higher concentrations of some endogenous hormones appear to increase breast cancer risk. 10 Risk factors associated with endogenous oestrogen – including early age at menarche, late natural menopause, later age at first full-term pregnancy and never breastfeeding – are all associated with an increased risk of breast cancer,11 while childbearing and a higher number of full-term pregnancies increase the protection. 11 Risk factors associated with the use of exogenous hormones, such as oral contraception, oestrogen replacement therapy and combined anti-oestrogen therapy, increase the risk of breast cancer, as do other factors such as breast density (a risk factor independent of endogenous hormones), a body mass index of > 25 kg/m² in postmenopausal women, moderate to heavy alcohol intake and a sedentary lifestyle. 11 Patients with a history of breast cancer12 and radiation exposure13 are also at increased risk.

Pathology, clinical staging and diagnosis

Breast cancer is classified into clinical stages according to tumour size, spread of cancer to lymph nodes and distant metastases. A number of different classification systems exist, including the tumour/nodes/metastasis (TNM) staging system developed and maintained by the American Joint Committee on Cancer (AJCC)14 and the Union Internationale Contre le Cancer (UICC). 15 In this system, ‘T’ refers to the size of the tumour and its spread, ‘N’ to the number of lymph nodes involved and ‘M’ to the presence of metastases (Table 3). The TNM system can be categorised further into disease stages (Table 4).

The stage of disease is an indication of prognosis. Data reported by Cancer Research UK in 2004 and cited by Ward et al. 6 suggested that the 5-year survival rate was around 90% for those with stage I disease, dropping to 75% for stage II, 42% for stage III and 14% for stage IV.

Alternatively, many clinicians in the UK use prognostic tools, such as the Nottingham Prognostic Index (NPI)16 or the web-based tool ‘Adjuvant! Online’. 17 The NPI takes into account three of the major prognostic factors, namely tumour size, lymph nodal status and grade according to the following formula:

The formula gives scores, which fall into the following categories:

• excellent-prognosis group ≤ 2.4

• good-prognosis group > 2.4 and ≤ 3.4

• moderate-prognosis group > 3.4 and ≤ 5.4

• poor-prognosis group > 5.4.

The 10-year predictive survival rates are as follows:18

• excellent-prognosis group = 96%

• good-prognosis group = 93%

• moderate-prognosis group = 53%

• poor-prognosis group = 39%.

‘Adjuvant! Online’ also incorporates tumour oestrogen receptor (ER) status and patient comorbidity, and provides an estimate of the potential benefit of treatment, derived from clinical trial data. This programme also has the feature of a modifiable prognostic calculator to factor in other known poor prognostic features, such as lymphovascular invasion and human epidermal growth factor receptor 2 (HER2) expression.

Identification of studies

The search aimed to identify all studies relating to the genotyping of CYP2D6 in the management of breast cancer, specifically related to TAM treatment. The following databases were searched on 19 June 2009: MEDLINE, EMBASE, The Cochrane Library (Cochrane Database of Systematic Reviews and Cochrane Controlled Trials Register), Web of Science (for the Science Citation Index and Conference Proceedings Citation Index) and the Centre for Reviews and Dissemination databases (Database of Abstracts of Reviews of Effects, NHS Economic Evaluation Database and Health Technology Assessment). Searches were not restricted by publication type. Because CYP2D6 genotyping is a relatively new area, and because the earliest study49 identified in the previous review of pharmacogenetics of TAM treatment was from 2003,47 searches were limited to the year 2000 and onwards. To assess the link between endoxifen plasma concentrations and clinical outcomes, a further search of MEDLINE was conducted on 21 July 2009, in which the inclusion criteria were extended to include studies considering the link between endoxifen concentrations and clinical outcomes, regardless of whether or not subjects had been genotyped for CYP2D6. The search strategies are listed in Appendix 1.

There was additional searching of the Human Genome Epidemiology Network Published Literature database, Proceedings of the American Society of Clinical Oncology, the San Antonio Breast Cancer Symposium and the European Society for Medical Oncology. Current research was identified from database citations through searching the National Research Register), the Current Controlled Trials register, the Medical Research Council Clinical Trials Register and the US National Institutes of Health website (ClinicalTrials.gov). Relevant reviews were hand searched in order to identify any further studies. Further studies that became known to the authors via relevant conferences or e-mail alerts from an automatically updated search of the Scopus database were also included as they became available, up to and including 17 March 2010.

Two reviewers (NF and RD) independently screened all titles and abstracts. Full-paper manuscripts of any titles/abstracts that were considered relevant by either reviewer were obtained. The relevance of each study was assessed (NF and RD) according to the inclusion and exclusion criteria listed in Box 2. Studies that did not meet the criteria were excluded and their bibliographic details were listed alongside reasons for their exclusion. Any discrepancies were resolved by consensus and, where necessary, a third reviewer was consulted.

Data extraction strategy

Data were extracted by one reviewer (NF) using a standardised data extraction form in Microsoft Word 2007 (Microsoft Corporation, Redmond, WA, USA) and checked independently by a second (JH). Disagreements were resolved by discussion.

Quality assessment strategy

As no universally accepted quality assessment criteria exist for assessing studies of pharmacogenetic testing, a tool based on elements of checklists developed to assess the methodological quality of prognostic factor studies50 and pharmacogenetic studies51 was used to assess specific issues considered important in terms of the reliability of such studies. Quality was independently assessed by two reviewers (NF and YD) and disagreements were resolved by discussion.

Methods of data synthesis

The results of the data extraction and quality assessment are summarised in structured tables and as a narrative description. Prespecified outcomes were tabulated and discussed within a descriptive synthesis. Additional relevant outcomes [breast cancer mortality and recurrence-free survival (RFS)] were also included.

Meta-analyses were planned in which binary outcomes were to be compared in terms of odds ratios, using either a fixed-effects or random-effects approach, depending on the degree of heterogeneity (to be assessed by visually inspecting the forest plots and by calculating the I²-statistic,52 which measures the proportion of variation across studies that is due to genuine differences rather than due to random error). In view of the controversy surrounding possible confounding from population stratification, and in keeping with the approach suggested in The HuGENet HuGE review handbook,53 in which studies differed in terms of the ethnicity of included patients, separate effect estimates were planned for each ethnic group. When studies differed in terms of their study design, sensitivity analyses were planned including only studies of the same study design. However, heterogeneity of the alleles genotyped, phenotypes derived, patients included and outcomes measured (see Results, below) precluded any planned meta-analyses.

Given the absence of either clinical utility studies or meta-analyses of the clinical validity data, attempts were made in an exploratory analysis to measure the clinical sensitivity and specificity of testing for particular alleles, as recommended by Flockhart et al. 54 in an American College of Medical Genetics statement. Data to calculate sensitivity and specificity were derived from the number of events reported in studies in the text/tables.

Number of studies identified and included

The literature search yielded 1186 citations after duplicates had been removed. Of the titles and abstracts screened at screening stage one, 57 were assessed in detail at screening stage two. At this stage, 27 citations55–81 (reporting on 23 studies) were excluded (see Appendix 2), leaving 30 citations to be included (22 studies41,73,82–101 reporting on clinical outcomes by CYP2D6 status and eight studies49,73,89,102–106 reporting on endoxifen plasma concentrations by CYP2D6 status). No studies were found that met the criteria for clinical utility.

Following completion of the search in June 2009, a further nine citations were identified that met the inclusion criteria for the review (Figure 2): five studies82,107–110 reported clinical outcomes by CYP2D6 status, another111 presented additional data for one of these studies,109 two112,113 reported endoxifen plasma concentrations by CYP2D6 status, and one114 included data on clinical outcomes (which also included patients from a study88 previously identified by the literature search) and endoxifen plasma concentrations by CYP2D6 status, in two separate studies. 114

FIGURE 2.

Identification of eligible studies.

The separate search for the link between endoxifen concentrations and clinical outcomes yielded 4998 citations after duplicates had been removed. Of these, none met the criteria for inclusion into the review.

Two ongoing studies69,115 which are of some relevance to clinical utility but which do not meet the inclusion criteria have also since been identified. Both of these studies have been presented as conference posters. Details of these ongoing studies are provided in Appendix 2.

It is further apparent that many of the different studies included in the review are in fact reporting on the same cohort of patients but with a few subtle differences, such as using only a specific subgroup of patients, considering different genotypes, taking into account concomitant medication that inhibits CYP2D6 or analysing different alleles and genotype classifications. As these cohorts share the same patients and study characteristics, it is preferable to consider the quantity and quality of research available by cohort rather than individual study or paper. Thus, in total there are 25 cohorts (and where reference is made to the cohorts as a whole, rather than specific studies, the latest fully published study in the table is used to derive the name of the cohort, e.g. the cohort including the studies by Jin et al. ,105 Borges et al. ,102 Henry et al. 87,107 and Rae et al. 95 is referred to as the ‘Henry et al. cohort87’).

Quality assessment of included studies

Given that there are 25 distinct cohorts, the quality of each cohort as opposed to individual study is summarised here (see also Appendix 3). While the majority of the studies of these cohorts were published as full papers in peer-reviewed journals, it is important to note that six82,86,90,98,109,113 cohorts have reported findings only at conferences.

All but seven41,49,86,90,93,98,113 of the cohorts were explicit about both the source population from which the study population was derived, and the definition of the study population itself. Six82,86,90,98,109,113 cohorts had reported their studies only as abstracts, not full papers, and so space was limited to present this information. Compared with the typical sample sizes required to provide sufficient power to detect a range of typical genetic effect sizes for various minor allele frequencies,51 the majority of the samples in this review are small.

The majority (n = 12) of cohorts41,73,83,87,93,96,97,99–101,113,114 presented the rationale for the alleles tested for, with all but two49,112 providing rationale for CYP2D6 per se. All described the test used for genotyping and/or the specific procedure, with TaqMan or AmpliChip being the most commonly used in 1283,86,87,91–94,96,97,104,108,114 and six41,82,87,90,109,113 cohorts, respectively. Three cohorts91,97,104 reported quality control methods and seven cohorts87,91,96,97,104,114 reported on the Hardy–Weinberg equilibrium (two of these cohorts were reported in the same paper114).

In around half (n = 12) of the cohorts it was clear there were missing genotype data,41,82,83,87,96–101,108,112 reasons being provided in seven of these. 82,83,87,97,99,108,112 Only three of the cohorts, all abstracts,86,90,113 failed to present the number of patients contributing to each analysis.

Characteristics of included cohorts

The cohort characteristics are summarised in Table 6, where it is clearly evident that the size of the cohorts varied, with the smallest containing 12 subjects49 compared with the largest of 288082 [the International Tamoxifen Pharmacogenomics Consortium (ITPC) cohort that included patients from three published studies: Matthew Goetz, Mayo Clinic, Minnesota, USA, 2010, personal communication]. Generally, however, cohorts included between 60 and 300 patients, two other exceptions being two larger cohorts of 67799 and 1361108 patients, this last cohort itself including patients from two previously published cohorts. 83,96

Unsurprisingly, all seven cohorts that measured endoxifen plasma concentrations were followed up prospectively. 49,73,87,104,112–114 All other studies were analysed retrospectively, using archived samples. In five cohorts,92,96,100,101,112 alongside data on patients receiving TAM, additional comparative data were provided for patients not receiving TAM.

Seven cohorts49,83,87,90,92,94,98 were solely from the USA, 1041,86,91,96,99,100,104,109,112,113 solely from Europe (including two from the UK91,109), six73,88,93,97,101,114 from Asia, one108 a combination of US and German patients and one82 from 12 unspecified ITPC project sites in the USA, Europe and Asia. The average duration of the studies varied considerably, from 4 weeks49 to 11.4 years,83 although, where average duration data were provided, all retrospective analyses were of at least 5 years’ duration.

Where cohorts provided data on TAM dose, this was 20 mg/day, the exceptions being two Swedish cohorts,99,100 in which average doses were higher, and the pharmacokinetic study of plasma concentrations by Bonanni et al. 112 in which they were lower. The majority of patients in these cohorts also received their dose for 5 years, the Swedish cohorts99,100 again being notable exceptions where 1–2 years’ dose duration was not uncommon and three other cohorts where it varied from a matter of weeks49,73,104 in pharmacokinetic studies to an average of 3 years. 97

Five cohorts83,87,96,108,114 were explicit in prohibiting adjuvant chemotherapy, while four41,92,100,109 were explicit in stating that this was permitted, concomitant chemotherapy in the other studies being uncertain. Data on CYP2D6 inhibitor use were also often either lacking or incomplete, with four93,101,113,114 cohorts explicitly prohibiting their use and five83,86,87,91,109 accounting for these in their analysis (noticeably, where there was more than one study for any given cohort this account was made in the more recent studies83,86,95,107).

It is also known that the study of endoxifen plasma concentrations by Lim et al. 73 also included patients with metastatic disease. No other study appears to include patients with metastatic breast cancer except Nowell et al. ,92 in which 5% of patients have metastatic disease and possibly also Wegman et al. 100 in which the inclusion criteria state that patients were required to have either histological verified lymph node metastases or a tumour diameter > 3 cm. It is also impossible to be sure in five of the pharmacokinetic studies49,102,104,112,114 and three of the efficacy studies whether or not the studies included patients with metastatic breast cancer. 94,98,100

Very few cohorts provided information on ethnicity, this being mentioned in just under half (n = 12) of the cohorts. 73,83,87,91–94,97,98,104,109,114 When this information was provided, Caucasians or Asians were represented in the study by at least 90% of all participants in all studies except Nowell et al. 92 in which there were 81% Caucasians and 19% African Americans, and Wang et al. 98 who simply stated that their population was ‘ethnically diverse’.

Not all cohorts provided data on the hormonal status of their patients. In six cohorts, all83,96,99,100,116 or nearly all (96%)108 of the women were known to be postmenopausal. In 10 others, there was a mix, although in only two instances41,93 was there a minority of postmenopausal women (40% and 22%, respectively). Similarly, not all cohorts provided data on hormone receptor status. Where these data were provided, all, or nearly all (> 90%), of the women were ER+ (or ER+ and/or progesterone receptor positive) in nine41,83,87,96,99,104,109,112,116 and five73,86,93,97,108 cohorts, respectively. In the remaining five cohorts91,92,100,101,114 in which this information was known, the proportion of ER+ women varied between 67% and 82%.

Less than half of the cohorts provided data on nodal status (n = 13)83,86,91,92,96,97,99–101,108,109,114,116 or tumour size (n = 12). 83,91,93,96,97,99–101,108,109,114,116 It was noticeable that the proportion of LN+ patients varied considerably across the cohorts, from no such patients97 to 89%,100 with a minority of patients being node positive in the majority of the studies. The proportion of patients with a tumour size ≥ 2 cm also appeared to vary significantly across the studies, from 27%101 to 72%,99 with a further two cohorts reporting only tumour sizes ≥ 3 cm in just under one-quarter of patients. 83,91

Fifteen cohorts measured efficacy. 41,82,83,86,91–93,96,97,99–101,108,109,114 Standard breast cancer study outcome measures, such as OS and DFS, were utilised; however, the definitions of these same outcomes often differed from study to study. In addition, depending on the cohort, the analysis was adjusted for in 14 cohorts. 82,83,86,91–93,96,99–101,104,108,109,114 Six cohorts41,83,87,90,94,98 reported on AEs and seven cohorts49,73,87,104,112–114 reported on endoxifen plasma concentrations. No cohort reported on health-related quality of life.

Derivation and classification of phenotypes

An important finding from our review was that currently there is no consensus about how CYP2D6 phenotypes should be derived from their genotypes and how they should thus be compared. In the current review, a large number of classifications and comparisons were utilised by different cohorts, and in some instances within cohorts. The different classifications used are summarised in Table 7, where it is evident that 10 cohorts82,83,87,90,91,96,104,108,112,113 used standard phenotypes (PM, IM, EM and UM), even though these were not always classified in the same manner from study to study, while others considered enzymatic function or simply compared genotypes. This heterogeneity makes comparisons across studies problematic, which is compounded further when one considers each cohort genotyped for different alleles, which may also be summarised as follows:

• Number of cohorts that:

  1. – genotyped for *4 only = 5

  2. – genotyped for *4 plus at least one other allele = 16

  3. – genotyped for *10 only = 3

  4. – genotyped for *10 plus at least one other allele = 12

  5. – genotyped for both *4 and *10 = 9

  6. – used the AmpliChip = 6.

For the purposes of this review, when considering study outcomes, ‘standardised comparisons’ are made in the tables and text in which alleles are simply considered to be wt (i.e. normal function), null (i.e. loss of function) or vt (any allele that is not wt, which includes null), and phenotypes are considered to be EM (wt/wt), IM (wt/vt) or PM (null/null) [this classification is the same as classification 1 for Test C in Table 2 (where null = *3, *4, *5 and vt = *10)]. It should be noted that for the purposes of these comparisons, UMs are likely to be classified as EMs. This is because not all genotyping methods are able to detect UMs and where studies have used methods that can, UMs appear to be classified with EMs, for example in Schroth et al. 108 Thus, the following comparisons can be considered:

• PM versus EM

• IM versus EM

• PM + IM versus EM

• PM versus EM + IM

• Asian patients genotyped for the *10 allele

• other comparisons that do not fit these categorisations.

Differences in cohort characteristics by genotype or phenotype

The cohort characteristics in Table 6 are for all patients in the cohort, regardless of CYP2D6 status. Only eight cohorts83,91,93,97,99–101,108 provided any of these data by genotype or phenotype, which is perhaps unsurprising given the retrospective nature of these studies (these data possibly not being originally collected).

It was noticeable in the Goetz et al. cohort83 that, compared with the cohort as a whole genotyped by paraffin tissue, more patients in the *4/*4 group had a tumour size ≥ 3 cm (38% compared with 22%) and were LN+ (69% compared with 38%). 84 Patients with this genotype were also more likely to be older (median age 73 years compared with 68 years) and have had a mastectomy (92% compared with 83%). It should be noted, however, that the number of patients with the *4/*4 genotype was small (n = 13) compared with those with other genotypes (n = 177).

Similarly, in Newman et al. ,91 compared with patients with other phenotypes, patients with the PM phenotype had a larger tumour size (42% compared with 21% had tumour size > 3 cm) and were more likely to have one or more positive lymph nodes at diagnosis (55% compared with 39%). These patients were also more likely to have had a mastectomy (67% compared with 49%) and be ER+ (92% vs 76%), but the median age in both groups of patients was similar (43 years compared with 45 years). Again, the number of patients with the PM phenotype was very small (n = 12) compared with those with other genotypes (n = 103).

Wegman et al. 100 also presented demographic data but these were for the group of patients as a whole, i.e. including both the patients who received TAM and the control group who did not. For each of the groups compared, *4/*4, wt/*4 and wt/wt, the proportion of patients who were node negative but with a tumour > 3 cm was similar in each group (11%, 9% and 12%, respectively). However, some differences were evident in terms of patients who were node positive with a tumour > 2 cm (33%, 58% and 44%, respectively) and ER+ (44%, 22% and 31%, respectively). Once again, the number of patients with the *4/*4 genotype was extremely small (n = 9) compared with those with other genotypes (n = 217).

In their later (separate) study, in which all patients received TAM, Wegman et al. 99 reported no real differences in tumour size or nodal status (tumour ≥ 2 cm being 71% for *4/*4, 75% for wt/*4 and 71% for wt/wt; patients with node > 0 cm being 71%, 69% and 69%, respectively). As with the previously mentioned studies, the number of patients with the *4/*4 genotype was small (n = 34) compared with those with other genotypes (n = 643).

The only study to compare patients with decreased activity as a whole (n = 716, of whom 79 were PM) with EMs (n = 609) was by Schroth et al.,108 who reported no real differences in tumour size > 2 cm (48% compared with 47%), lymph nodal status (36% compared with 31%) or age at diagnosis (median age was 66 years in both groups) between groups. No real differences were evident for any other cohort characteristics presented by the authors.

Xu et al. 101 reported differences in cohort characteristics in Asian women across genotype groups, namely *10/*10, wt/*10 and wt/wt. While lymph nodal status appeared similar across the groups (6%, 8% and 7%, respectively, were node positive), more patients with the *10/*10 genotype appeared to have larger tumours (32% > 2 cm compared with 23% and 21%, respectively) and fewer were ER+ (86% compared with 92% and 96%, respectively). In this cohort, there were almost as many women with the *10/*10 genotype (n = 72) as with other genotypes (n = 80).

Toyama et al. 97 reported few differences between the three groups of Asian female patients, with the notable exception that more women with the wt/wt genotype had a tumour > 2 cm (59% compared with 44% in each of the other genotype groups). Data were not presented on nodal status. Age differed only slightly (median 56 years in the*10/*10 group compared with 60 years in the other two groups). The vast majority (> 96%) of patients were ER+ in all groups. The proportion of patients with the *10/*10 genotype was relatively small (n = 28) compared with other genotypes (n = 126).

In Okishiro et al. ,93 fewer differences were apparent when patients with the *10/*10 genotype were compared with patients with all other genotypes, the number of patients with tumour size > 2 cm and LN+ in each group being comparable (29% vs 28% and 43% vs 45%, respectively) and median age being similar (47 years compared with 46 years). However, it was noticeable that there were more patients with the *10/*10 genotype with ER+ breast cancer (92% compared with 60%). The number of patients with the *10/*10 genotype was again relatively small (n = 40) compared with other genotypes (n = 133).

While other studies neither presented cohort characteristics by genotype/phenotype nor reported any differences between groups, seven other cohorts did adjust for prognostic factors. 82,86,92,96,104,109,114

Efficacy

The efficacy of TAM treatment was considered by genotype/phenotype, using the following comparisons as described above (see Derivation and classification of phenotypes):

• PM vs EM

• IM vs EM

• PM + IM vs EM

• PM vs EM + IM

• Asian patients genotyped for the *10 allele

• other comparisons that do not fit these categorisations.

Unfortunately, not all clinical end points measured by the cohorts were defined. Where end points were defined, it was apparent that different cohorts used different definitions for the same end points. Given these inconsistencies and/or a lack of information to correctly classify a clinical end point [e.g. information on censoring would enable recurrence outcomes to be classified as DFS, RFS or time to recurrence (TTR)], the efficacy outcomes are presented below in terms of OS, breast cancer mortality (i.e. mortality attributed only to breast cancer and not from any cause as with OS) and outcomes such as DFS, RFS and TTR, which can be considered relating to relapse/recurrence.

Overall survival is considered the least ambiguous and most clinically relevant clinical end point. 117 Seven cohorts82,83,91,92,96,97,108 examined OS by CYP2D6 status. One108 of these studies also includes 350 patients from two of the other included cohorts,83,96 and the large ITPC cohort82 also contains data from three published data sets (Matthew Goetz, personal communication). Two cohorts considered breast cancer mortality97,101 and 13 cohorts41,82,83,86,91,93,96,97,99,100,108,109,114 reported outcomes relevant to relapse/recurrence.

Breast cancer mortality

Two97,101 cohorts of Asian patients were examined for breast cancer mortality by *10 genotypes (Table 17). The majority of women had ER+ breast cancer in both cohorts, although this varied from 82% to 96%. Three-quarters of women in Xu et al. 101 were postmenopausal; these data were not reported in the other cohort. 97 While the proportion of patients with tumour size ≥ 2 cm varied in the cohorts (from 27% to 48%), few if any patients were LN+ (0–7%). Adjuvant chemotherapy was not permitted in Xu et al. ;101 it was unclear if patients received this in the other cohort. 97 Interestingly, in one cohort a greater proportion (32%) of patients with the *10/*10 genotype than other genotypes had tumours ≥ 2 cm, whereas in the other cohort97 a greater proportion (59%) of patients with the wt/wt genotype had a tumour size > 2 cm.

The findings are summarised in Table 17. While Xu et al. 101 suggested that, compared with wt/wt + wt/*10, breast cancer mortality was higher for women with the *10/*10 genotype, the CIs were wide and the finding was not statistically significant. From the Kaplan–Meier curve, the other cohort suggested that *10/*10 patients may actually have improved mortality from breast cancer when compared with patients with either the wt/wt or wt/*10 genotypes, but again the findings were not statistically significant.

One cohort also compared breast cancer mortality in patients taking TAM with a control group of patients receiving chemotherapy, the majority of whom had ER– breast cancer. As with patients taking TAM, there was no significant association between patients with *10 alleles and DFS in this group, suggesting that *10 is not a prognostic factor for breast cancer independent of TAM.

Relapse/recurrence

Five studies88,93,97,101,114 from four cohorts93,97,101,114 reported on relapse/recurrence outcomes. Ostensibly two cohorts97,101 reported on DFS and two93,114 reported on RFS, although, as can be seen in Table 18, the outcome definitions were very similar for both DFS and RFS.

The proportion of postmenopausal women in each cohort differed, varying from 22% to 76% in the three93,101,114 cohorts that reported these data. In all cohorts, the majority of women had ER+ breast cancer, although this varied from 74% to 82% in two101,114 cohorts to 91–96% in the other two. 93,97 Differences also existed for tumour size and nodal status, the proportion of women with tumours ≥ 2 cm varying from 27% to 48% and LN+ patients varying from 0% to 17%, although nodal status was not reported in one cohort. 93 Three93,101,114 of the studies explicitly excluded patients with CYP2D6 inhibitors, and two101,114 did not allow adjuvant chemotherapy, these data not being reported in the other cohort(s). There were, however, some differences within cohorts by genotype in the three93,97,101 cohorts that reported these data. In one cohort, a greater proportion (32%) of patients with the *10/*10 genotype had tumours ≥ 2 cm, whereas, in the other,97 a greater proportion (59%) of patients with the wt/wt genotype had a tumour > 2 cm. No such differences were apparent in the third cohort,93 although 90% of patients with the *10/*10 genotype had ER+ cancer compared with 60% with the other genotypes.

Compared with other genotypes, all four cohorts93,97,101,114 suggest that there may be modestly poorer outcomes in terms of DFS and RFS for patients with the *10/*10 genotype, but CIs are extremely wide and a significant difference was reported in only one114 of these cohorts. This cohort still reports wide CIs, suggesting that the finding should be treated with caution.

One cohort also compared DFS in patients taking TAM with a control group of patients receiving chemotherapy, the majority of whom had ER– breast cancer. As with patients taking TAM, there was no significant association between patients with *10 alleles and DFS in this group, suggesting that *10 is not a prognostic factor for breast cancer independent of TAM.

Adverse events

Nine studies41,84,87,90,93–95,98,107 from seven cohorts41,83,87,90,93,94,98 have considered any AEs in relation to CYP2D6 status (Table 20).

The only AEs that appeared to differ by CYP2D6 status in any of the cohorts were hot flushes. 83,87,90 Here EMs and/or IMs appeared to be more prone to suffer hot flushes than PMs,83,87,90 more prone to experience severe hot flushes83,87 and more likely to discontinue treatment because of hot flushes. 87

Endoxifen concentrations

There were eight studies49,73,102,104,105,112–114 from seven cohorts49,73,87,104,112–114 that examined endoxifen concentrations in relation to CYP2D6. Four studies reported mean plasma concentrations102,105,112,113 and three reported median plasma concentrations. 104,112,114 Two studies reported decreases in endoxifen following the administration of an SSRI (paroxetine). 49,102 In five cohorts49,73,87,104,114 the TAM dose was known to be 20 mg/day, whereas in Bonanni et al. 112 the drug dose was only 10 mg/day as half of the patients receiving TAM in this cohort also received ANA 1 mg/day. In the other cohort113 the TAM dose was not stated.

The findings from each individual cohort appear to suggest differences in concentrations between PMs87,104 or those with the *10/*10 genotype73,114 and EMs (or those with the wt/wt genotype). However, in the Caucasian population, one cohort104 reported the levels for the IMs to be closer to those of EMs than PMs. The other cohort87 reported wt/vt genotypes (which may be equated as IM) to have levels closer to vt/vt genotypes. Reduced decreases in mean endoxifen plasma concentrations were also evident in patients taking potent CYP2D6 inhibitors in two cohorts. 49,102

Overall survival

Taking into account the differences between the cohorts, the data suggest that the sensitivity and specificity of testing simply for *4 in the adjuvant setting may be 15% and 73%, respectively, for OS92 (Table 23). A more comprehensive genotype test (in terms of the number of alleles tested) appears to increase sensitivity and specificity to 18% and 83%, respectively. 108

Relapse/recurrence

The data from four cohorts86,92,99,100 suggest that the sensitivity of testing simply for *4 in the adjuvant setting may be between 21% and 37% for relapse/recurrence and specificity may be between 52% and 86% for relapse/recurrence (Table 24). If, in testing for *4, phenotype status is altered based on concomitant CYP2D6 use, based on a small number of patients (n = 84), for relapse/recurrence, sensitivity may be 50% and specificity 73%. 86 Utilising data from the only cohort to test simply for *10 suggests a sensitivity of between 19% and 50% and specificity of between 95% and 96%. If a more comprehensive genotyping strategy is used, data from the largest cohort of Schroth et al. 108 and Schroth et al. 96 (whose patients are actually included in the Schroth et al. 108 cohort) suggest a sensitivity of between 18% and 30% and a specificity of between 86% and 88%.

Requirements for a de novo economic evaluation

In order to undertake an economic evaluation of pharmacogenetic testing within the framework of the wider model of breast cancer care, the following clinical data requirements are considered to be most important:

1. Epidemiological data Data related to allele frequencies for selected genetic variants and how these are distributed in populations; in particular, the data should indicate which groups of patients (as defined by genotype and/or phenotype) would need to be identified by the pharmacogenetic test.

2. Clinical effectiveness Evidence of a link between phenotypes and drug metabolism and data describing clinical outcomes and AEs, including long-term effects of the drugs. It will be necessary to pay special attention to the choice of the time horizon of the economic evaluation which will be related to the length of the treatment (e.g. as stated in guidelines) and ‘carry-over’ effects of the treatment (i.e. delayed effects of the treatment after discontinuation).

3. Test accuracy Data are required on both the sensitivity and specificity of the test and how accurate the test is in linking genotypes to phenotypes and then to clinical events and predictive value of the test.

4. Uptake of the test The degree of test uptake, by patients or clinicians, will have an impact on cost-effectiveness.

5. The impact that pharmacogenetic test results will have on clinicians’ behaviour Data are required including the impact that test results have on prescribing decisions, and how this affects the overall delivery of care.

From a health economics perspective, the key elements that need to be considered when undertaking a de novo economic evaluation of CYP2D6 testing as a management option for women with breast cancer after surgery are discussed below.

Modelling

A simple decision tree could be used to model the sensitivity and specificity of the genotyping test (Figure 3). Beyond this point, a de novo Markov model would be more appropriate. A Markov model structure is considered appropriate because it is assumed that breast cancer is a condition that causes patients to move between a limited number of relevant health states during their lives. This type of model allows a large number of cycles to be simulated without the need to create a new decision tree in each cycle. Figure 4 depicts the schematic model that includes the possible health states and possible transitions between these states.

FIGURE 3.

Decision tree for genotyping for CYP2D6. The Markov model referenced in this figure is depicted in Figure 4.

FIGURE 4.

Markov model for genotyping for CYP2D6.

As can be seen from Figure 4, the Markov model reflects seven health states:

• DFS without AEs: women at risk of an AE owing to the medication received or at risk of any relapse/recurrence

• DFS with AEs: women at risk of any relapse/recurrence with AEs

• contralateral disease: those women with a new primary tumour in the contralateral breast

• locoregional disease: women suffering a locoregional recurrence or ipsilateral second primary tumour

• metastatic disease: women with metastases (detail relating to different sites of metastases could be incorporated if relevant data are available)

• breast cancer death: death from metastatic disease only, as death from either contralateral disease or locoregional disease is unlikely

• non-breast cancer death: death from any cause apart from breast cancer.

Trying to model the cost-effectiveness of this technology seems to be premature given the quantity and quality of the clinical effectiveness and cost-effectiveness evidence available. It is particularly challenging because there are problems with identifying the alleles to test for, derivation of phenotypes and the lack of cost information available.

Patient population

As TAM is considered the standard of care for premenopausal women with ER+ breast cancer, and as anti-oestrogen therapy is not used for patients with ER– breast cancer, the only situation in which genotyping for CYP2D6 might prove to be useful is if alternative treatment options to TAM, namely aromatase inhibitors, are considered to be at least as appropriate. Currently, aromatase inhibitors are recommended for women with ER+ early breast cancer deemed not to be at low risk of disease recurrence, and for women with ER+ advanced breast cancer with no prior history of anti-oestrogen therapy. 20 The decision to offer aromatase inhibitors to higher risk patients is largely based on the results of published RCTs21–27 and systematic reviews28–31 of women with ER+ early breast cancer suggesting a modest benefit for patients taking aromatase inhibitors over those taking TAM. Thus, the population of interest is likely to be postmenopausal women with ER+ breast cancer.

It has been proposed that patients be given TAM or an aromatase inhibitor depending on the test results. This decision may be based on their genotype or phenotype. Thus, it is important to know the distribution of postmenopausal patients with early breast cancer with each of these genotypes or phenotypes. This distribution could be inferred from the studies included in the review.

Concomitant CYP2D6 inhibitor medication

While NICE18 states that neither paroxetine nor fluoxetine should be given to women taking TAM, a few studies included in our review state that women on TAM are taking these SSRIs. 49,86,87,91,97,99,109 To accurately reflect this situation in the model, it is necessary to know how SSRIs change the phenotype of these women, as well as the extent to which women with breast cancer are taking paroxetine or fluoxetine in clinical practice. Ferraldeschi et al. 127 conducted a survey to assess the current practice of breast oncologists in the UK with respect to CYP2D6 testing and SSRI co-treatment. The authors reported that 22% of respondents stated that there was enough evidence to routinely use CYP2D6 testing, 37% required more evidence and 41% were unsure. There was general agreement (80%) that concomitant medications might affect the clinical efficacy of TAM, and 86% indicated that they discuss drug interactions with their patients. Importantly, 93% would not prescribe a potent CYP2D6 inhibitor concomitant with TAM and would prescribe alternative treatments, where available.

Adherence to treatment

Two separate studies from the USA128 and the UK129 have reported that non-adherence to TAM ranges from 13% to 22% after the first year of treatment to 50% in years 4 and 5 of treatment. Data on aromatase inhibitor non-adherence suggests that this rises from 14–22% in year 1 to 21–38% in year 3. 130 However, crucially, there is a lack of data with regard to TAM adherence by genotype. Data on adherence are important because what few data we have suggest that the most common reason for discontinuing treatment is experience of hot flushes,95 which are most common in EMs and IMs. 84,90,107

Timing of test

In the absence of any clinical guidelines or prospective clinical studies, it is difficult to know where one would model the CYP2D6 test along the treatment pathway. Currently, the decision to prescribe TAM or an aromatase inhibitor (or indeed, a switching strategy) is made depending on the risk of disease recurrence. However, other factors are also taken into account, for example in relation to AEs from each drug. Thus, would CYP2D6 testing be required only after risk and these other factors been taken into account or would the results from CYP2D6 be another of these other factors to consider? From a practical point of view, as patients are already tested for ER and HER2 status before they are treated, it may be surmised that CYP2D6 testing would occur at the same time, even if the CYP2D6 test results are considered only at a later stage. Finally, in the absence of any clinical guidelines or prospective clinical studies, it is also unclear which pathways patients would follow after CYP2D6 testing, although current treatment pathways for those treated with TAM and aromatase inhibitors would seem logical.

CYP2D6 test availability, cost and accuracy

A number of genotyping tests exist. Many are designed ‘in house’ (using techniques such as TaqMan) to test for specific alleles. Others are offered commercially, such as the AmpliChip. No CYP2D6 genotyping is currently provided by the NHS and so there is no national price list for these tests. Many of these tests are designed specifically for research studies; therefore, the test performance in clinical practice is unknown, although generally genotyping has been shown to have high analytical validity. 45 The greatest uncertainty is in determining clinical sensitivity and specificity and predictive value, i.e. how accurate the test is in linking phenotypes to clinical events. Although we have conducted some exploratory post hoc analyses of sensitivity and specificity for some tests used in the studies included within our systematic review (see Chapter 3, Exploratory analysis: clinical sensitivity and specificity), it should be noted that these are from only a selected sample of studies reporting the necessary data to calculate these values. It is important to note that these are not tests that may be used in clinical practice. They are indicative, however, of the types of data that would be required for an economic model.

Penetrance, the degree of phenotypic expression of genetic variation, is a key parameter for an economic evaluation of pharmacogenetic genetic testing. 131 It is possible to design a test with almost perfect characteristics in terms of test sensitivity and specificity, but the test can still have poor positive predictive value because of the impact of low penetrance, which will affect the cost-effectiveness of the test. We were not able to estimate the degree of gene penetrance and associated positive predictive value for the CYP2D6 test because the clinical data were not able to inform which alleles should be tested. This information is a prerequisite before gene penetrance and test positive predictive value can be established.

Anticipating the importance of the types of tests, alleles tested and their cost to the model, we undertook a survey of several laboratories with regard to their current practices in relation to CYP2D6 testing for patients to be treated with TAM. We chose five laboratories that are currently testing patients for several genes, not only CYP2D6, and, to gain greater insight, chose three from the UK, one from the Netherlands and one from the USA. These laboratories were chosen because they were known to us as laboratories offering the services, often in relation to research studies, and we therefore knew that they were likely to be using the most up-to-date techniques. The survey was conducted between January 2010 and June 2010 and the questions were submitted via e-mail.

The findings from this brief consultation exercise are summarised in Table 25, where it can be seen that there is wide variation regarding the number of alleles tested. The costs also vary considerably by laboratory, from as little as £30 to as much as £500. This wide variation in alleles tested is concordant with the wide variation in the number of alleles tested across the studies included in our review. The wide variation in cost is likely to be due in part to the wide variation in the number of alleles tested, as most assays (e.g. TaqMan) require each allele to be tested individually, so increasing the materials required, time taken, etc. It is perhaps worth noting here that the AmpliChip, which tests for 33 alleles, has been quoted as costing US$500 per test in the USA132 and £300 per test in the UK. 45 The AmpliChip is, to date, the only test that is licensed for use by the FDA and, unlike many tests, is able to test multiple alleles simultaneously.

The number of alleles tested is important to correctly classify patients into their correct phenotype. A recent paper by Schroth et al.,121 re-analysed data from German patients in the large Schroth et al. cohort108 and reported that one-third of patients identified as PMs by the AmpliChip were also identified as PMs by testing for only *4. This proportion rose to 62% when testing for three alleles (*3–*5) and 100% when testing for five (*3–*7). If replicated, the findings from this analysis could suggest that only five alleles need to be tested, if the treatment decision is made on whether or not a patient is a PM. Unfortunately, we do not know this to be true. The clinical evidence from our review suggests that it may be more important to correctly identify EMs (although the evidence relating to endoxifen levels does not seem to support this). Thus, a wider range of alleles would need to be incorporated, in particular those associated with the IM phenotype, such as *10.

Test uptake

The degree of test uptake is an important parameter to consider when evaluating the incremental costs and benefits of pharmacogenetic testing. Low test uptake could affect the cost-effectiveness of the testing, i.e. low uptake reduces the cost-effectiveness. The uptake of pharmacogenetic tests, and whether uptake is driven by the patient or the clinician, are not known and are topics for future research. There are some data that have reported test uptake for breast cancer chemoprevention. 133,134 Uptake rates range between 11.7% in populations with poor genetic counselling and 31% in populations informed by genetic counselling. 133

Time horizon

The time horizon of the model must reflect, at the very least, the duration of the treatment and differences in resource use and/or changes in patient outcomes. Where the time frame of the model exceeds the duration of treatment, it is therefore also important to know how long patients are likely to live for, how long they are likely to be disease free, what other treatments they will receive subsequently, etc. Given that 75% of patients have a life expectancy > 5 years,5 it would seem appropriate to model beyond treatment and until death. As noted above, this means taking into account ‘carry-over’ effects from treatment but (also noted above) may also be problematic given the current lack of established pathways of care. Nevertheless, it would seem feasible to populate a model that followed patients up to death.

Uncertainty

Understanding the impact of uncertainty is a key aspect of any economic model. A model of CYP2D6 testing that is subsequently structured and populated is likely to be an early economic model that in part will aim to inform the need for further research to collect data on key parameters. As a minimum, two key types of uncertainty should be addressed: parameter and structural uncertainty. Parameter uncertainty would be addressed using probabilistic sensitivity analysis, which would allow the analyst to estimate the expected value of perfect information for future research and types of future research required. Structural uncertainty reflecting different possible care pathways should also be explored and included in the model in such a way that allows the analyst to explore the impact of structural uncertainty on whether or not additional evidence is needed. 135

Ovid MEDLINE, 2000 to June week 2 2009

EMBASE, 2000–9 week 24

Web of Science, Science Citations Index and Conference Proceedings Science Index

The Cochrane Library, issue 2

Ovid MEDLINE, 2000 to June Week 2 2009

The searches for the studies for the economics review were identified from the combined searches above.