Promises and perils of using genetic tests to predict risk of diseaseBMJ 2020; 368 doi: https://doi.org/10.1136/bmj.m14 (Published 07 February 2020) Cite this as: BMJ 2020;368:m14
- Ian A Scott, director of internal medicine and clinical epidemiology and professor of medicine12,
- John Attia, professor of medicine and clinical epidemiology and consultant34,
- Ray Moynihan, assistant professor and honorary associate professor56
- 1Princess Alexandra Hospital, Woolloongabba, QLD, Australia
- 2University of Queensland, Brisbane, QLD, Australia
- 3University of Newcastle, Callaghan, NSW, Australia
- 4John Hunter Hospital, Newcastle, NSW, Australia
- 5Institute for Evidence-Based Healthcare, Bond University, Robina, QLD, Australia
- 6Sydney Medical School—Public Health, University of Sydney, Sydney, NSW, Australia
- Correspondence to: I A Scott
Low cost genetic testing is increasingly being used by patients and the public to predict risk of developing disease in asymptomatic people in the hope that more precise risk stratification might facilitate targeted interventions for reducing risk
The proliferation of genetic variants might cause clinicians and citizens to misread their clinical relevance, potentially leading to overestimation of risk, overdiagnosis, and overtreatment
In appraising the value of genetic testing for clinical decision making, consideration must be given to validity, predictive accuracy, clinical utility, potential harms, cost effectiveness, and feasibility of use in routine care
Moving from traditional genetic testing for rare monogenic disorders within families to wider polygenic testing for common diseases in heterogeneous populations requires robust evidence of benefits and harms of this paradigm shift
Increasing numbers of patients and clinicians are undertaking low cost genetic testing in asymptomatic people to identify genetic variants that might predict risk of developing diseases. By early 2018, an estimated one in 25 citizens of the United States had undergone genetic testing, more than double the rate in the previous year.1 Although testing for risk of monogenic diseases such as cystic fibrosis in people with family histories is often appropriate, extending testing for polygenic diseases such as cardiovascular atherosclerosis to people with no family history is problematic and might cause harm.
Tests for approximately 75 000 genetic variants are now commercially available from companies such as 23andMe, Navigenics, and deCODE Genetics, which can be ordered on the internet by consumers anywhere in the world and are increasingly advertised in lay media in the US, Canada, Australia, and various European countries.2 These tests aim to predict individual risk of developing coronary artery disease, diabetes, Parkinson’s disease, dementia, and some cancers.3 Most rely on DNA assays that screen for known single nucleotide polymorphisms associated with specific diseases. Next generation sequencing, however, in which the entire DNA sequence is read rather than just loci of known variation, has generated many new variants, the disease associations of which await investigation. These are called “variants of uncertain significance.”
Predictive genetic testing aims to more precisely stratify people by risk and thus enable targeted interventions for reducing risk such as lifestyle counselling, closer surveillance, more intense screening, and pharmacological prophylaxis. But the proliferation of genetic variants and related risk scores might cause clinicians and consumers to misinterpret their results. People at low risk may be overdiagnosed as having a “pre-disease state,”4 which might never manifest as overt disease, causing overtreatment with potential harm.5 Conversely, falsely reassuring tests in people at high risk might result in loss of close surveillance and preventive intervention. In this analysis, we appraise the value of predictive genetic testing with the aim of raising awareness among both clinicians and consumers of the limitations of these tests and potential harm from their more widespread, indiscriminate use.
Predictive accuracy of genetic variants
Whether a variant is truly associated with a specific disease and has a clinically meaningful genetic effect in causing it can be contentious. The absolute risk of disease in an individual is governed partly by the number and frequency of genetic variants influencing disease risk, the magnitude of genetic effect of each variant, and the prevalence of the disease in the general population.
High quality studies detailing variant-disease associations are limited, as all relevant variants for many diseases await identification. Analyses of thousands of single nucleotide polymorphisms have yielded false positive disease associations by applying too lenient a threshold for calling a variant rare and presumably harmful.6 Replication studies of the same variant-disease associations can yield conflicting results,7 and studies of disease enriched populations, such as certain families or ethnic groups, might yield non-generalisable associations, as most studies to date are limited to populations of European ancestry.
Furthermore, individual variants usually exert small genetic effects such that, despite occurring relatively frequently (in >5% of people), their cumulative effects on risk of any disease are weak; for example, variants account for only 2-5% of the variation in the incidence of many common conditions such as coronary heart disease.8 By contrast, variants that do not themselves cause disease can sometimes be highly predictive of disease by being linked and inherited with other pathogenic genes. Most variants do not increase the disease risk by more than 50%,9 which translates into only slight increases in absolute risk (<5% in most cases).
Genetic effects also depend on the degree of expressivity and penetrance of the variant. Familial disease linked to fully expressive and highly penetrant but rare variants accounts for only 1-10% of all disease incidence.9 Non-genetic factors such as environmental exposures, lifestyle, smoking, and diet often bestow greater risk of future disease.
Polygenic risk scores might predict disease risk more accurately
Polygenic risk scores might improve predictive accuracy by combining incremental information provided by hundreds or thousands of different variants into one result. These scores vary according to the number of variants, the specific type of statistical model used to combine risks associated with individual variants, and how widely applicable they are. But they have limitations. First, most polygenic risk scores take no account of non-genetic risk factors, although newer scores that integrate genetic, clinical, and laboratory variables might be more accurate.1011 Second, different commercially available scores can yield markedly inconsistent risk estimates for the same disease.12 Third, most polygenic risk scores only consider people with or without well defined disease at a fixed point in time. They are unable to adjust risk estimates according to the time of disease onset (such as early onset obesity or late onset Alzheimer’s disease) or spectrum of disease (in type 2 diabetes, from glucose intolerance to frank diabetes). Fourth, predicted increases in relative risk of disease might still translate to very small increases in absolute risk; in other cases, they yield no additional information to single gene tests (box 1).1314151617
Some limitations of polygenic risk scores
One polygenic risk score predicts a 90% increase in risk of dementia over eight years, but this translates to only 0.6% increase in absolute risk13; others provide no more information than single allele tests such as APOE ε4.14
Polygenic risk scores can identify only small proportions of the population who have more than three times the risk of developing coronary artery disease (8%), atrial fibrillation (6.1%), type 2 diabetes (3.5%), inflammatory bowel disease (3.2%), or breast cancer (1.5%).15
Increasing the number of variants in polygenic risk scores (to as many as 7 million) yields little change in predictive accuracy compared with more simple scores.16
To be useful, polygenic risk scores should discriminate well between people who develop disease and those who do not (as measured by areas under the receiver operator characteristic curve of 0.8 or more); should have score cut-off points at which reasonable sensitivity and specificity can be obtained; and should be well calibrated.17 Very few meet all these criteria.16
Clinical utility of predictive genetic testing
Genetic tests are most useful to clinical decision making when they fulfil four criteria (box 2) 18192021: they are targeted to people more likely to harbour a particular variant, facilitating more cost efficient testing in those with higher pre-test probability and less chance of false positive results; they are the only way of reliably identifying high risk individuals with no family history or other risk factors, or their estimates of risk are substantially better than those calculated using current clinical risk prediction rules; if preventive actions effective in reducing risk of disease developing exist; if decision making, informed by genetic test results, leads to more preventive actions being adopted by people, with consequent better outcomes. Emerging guidelines focus on these criteria for several conditions.222324
Assessing genetic tests for clinical utility
An example of a near ideal test identifies the BRCA1 and BRCA2 mutations, which, although rare (<1% of female population) and seen in only 10% of all breast cancers,18 convey a 50-80% and 15-40% absolute lifetime risk of breast and ovarian cancer, respectively.19 Algorithms exist for estimating pre-test probability of these mutations, which facilitates targeted testing.20 Carriers of the mutations are recommended, at the very least, to undergo intensive screening for breast cancer or to consider risk reducing mastectomy, oophorectomy, or both, which yield better outcomes than usual care in young persons.
By contrast, identifying the LRRK2 or GBA variant for Parkinson’s disease confers a modest (15-80%) increase in risk of the disease developing, has attracted no reliable pre-test probability scores, and does not assist care as no proven prevention strategies exist for this condition.21RETURN TO TEXT
Many variants purported to indicate increased risk of specific diseases might add little prognostic information beyond age, sex, family history, and known risk factors. One review of 23 studies of variants for type 2 diabetes showed no increase in predictive accuracy compared with traditional risk factors, despite differences in study design, sample size, participants’ race or ethnicity, or number of variants.25 By contrast, new polygenic risk scores for coronary heart disease have reclassified up to 12% of patients from intermediate to high risk,2627 with higher rates of adherence to statin treatment noted among individuals with high genetic risk.28 Recently developed polygenic risk scores for breast29 and prostate cancer30 identify high risk individuals who might benefit from more intensive screening.
Many clinicians, however, consider genetic tests irrelevant31 or use them inconsistently,32 and we are currently lacking clinical trials that confirm changes in consumer behaviour or improved outcomes from using genetic data. Most available studies are limited by small samples, short follow-up, and self-reports focusing on intentions rather than actual behaviour.33 A trial in which participants were randomly allocated to receive information of their genetic risk of diabetes, in addition to lifestyle modification, did not show any differences in short term weight loss or motivation to change behaviour34 or in six year incidence of diabetes.35 A review of 18 trials in which people were informed of their genetic risk of various diseases showed no effects on smoking, diet, physical activity, alcohol use, or participation in screening and behavioural support programmes.36 More recent trials using polygenic risk scores to identify those at higher risk of developing diabetes or cardiovascular disease have not fared much better.3738 The premise that having more genetic information motivates people to alter risky behaviours remains unproven.39
Harms of predictive genetic testing
Harm to healthy people might ensue from test results that are incorrect (false positives or false negatives), that are misinterpreted by clinicians who fail to account for family and personal medical history or other risk factors, or that disclose a condition that would never have manifest clinically. Genetic variants with little or no pathogenicity can still evoke worry and unnecessary care.40 Conversely, knowing one has almost double the average risk of developing Alzheimer’s disease, but with no recourse to proven preventive interventions, might simply cause psychological distress,41 and even thoughts of suicide, as claimed by more than 10% of respondents to one survey.42 Misinterpretation of BRCA genetic test results has resulted in inappropriate surgery.43
False reassurance can arise from negative tests that do not include, or fail to identify, pathogenic variants or interacting genes (23andMe, for example, tests for only three of over a thousand potentially cancer related variants of the BRCA1 and BRCA2 genes). Disclosure of genetic risk to employers, insurers, and government agencies might result in discriminatory practices against individuals deemed to be at higher risk.44
Cost effectiveness of predictive genetic testing
Targeting preventive measures to selected patients at high genetic risk should reduce costs compared with intervening in larger numbers of people at average risk. However, economic evaluations of the same genetic test can yield contradictory findings. One review found that BRCA testing in all patients with ovarian cancer and their first and second degree relatives was cost effective in reducing incidence of, and mortality from, breast and ovarian cancer in variant positive relatives,45 but another review did not.46 A review of 128 economic evaluations of genetic testing found that only 12% of predictive tests that were analysed were definitely cost saving. For the remaining tests, the results varied considerably from study to study due to uncertain estimates of clinical utility, ill defined time horizons and discount rates, different technologies against which genetic tests were compared, and limitations in identifying and quantifying use and costs of downstream health services.47
Predictive genetic testing in routine care
The rapid introduction and commercialisation of genetic tests have occurred with little regulatory oversight, implementation planning, or professional education and decision support.48 Clinical scenarios in which a genetic test might be useful, cost effective, and devoid of harm need to be clearly defined.49 Genetic test results must be communicated to individuals in readily interpretable and actionable formats. Consultations with expert clinical geneticists can improve test interpretation and patient counselling, but their availability is being outpaced by the availability of genetic tests, a problem exacerbated by the advent of whole exome and whole genome sequencing.
As many as 20% of asymptomatic people who undergo whole exome sequencing show a variant with potential disease risk,50 whereas many other variants identified in less than 1% of the population are classified as variants of uncertain significance.51 Only around 60 pathogenic genes with definite clinical implications have been identified to date, but these occur in no more than 3.5% of asymptomatic people.52 By contrast, whole exome sequencing of an average person can generate hundreds of variants of uncertain significance, with identical variants assigned different degrees of pathogenicity by different laboratories owing to inconsistency in methods used to assess determinative factors and inherent subjectivity in interpretation.53 Up to two thirds of so called disease causing variants might be innocuous,54 but testing for them can incur added harm and costs (box 3).55565758
Examples of potential harm and costs of next generation sequencing
Large databases (such as the Human Gene Mutation Database and ClinVar) that collectively house classifications for over 325 000 variants in more than 4800 genes show that many variants previously reported as pathogenic have population frequencies incompatible with causality for the associated diseases. Despite recent reclassifications, a substantial number remain listed as “pathogenic.”55
In a recent study of 7974 healthy people, whole exome sequencing identified potentially pathogenic variants for kidney and genitourinary disorders in 23.3%. Although reduced to 1.4% after stringent filtering based on minor allele frequencies, this still vastly exceeded the expected rate of monogenic kidney and genitourinary disorders and is likely to lead to genetic misdiagnosis and unnecessary follow-up testing.56
The genetic findings of a randomised trial of whole genome sequencing plus family history versus family history alone among 100 healthy adults in primary care were of uncertain clinical utility, but they prompted clinical actions of unclear value in 11% more patients in the whole genome sequencing group over the next six months.57
In an exploratory study of 12 healthy volunteers who underwent whole genome sequencing, a median of 108 genetic variants were detected per participant, of which five were associated with disease risk and 13 had implications for disease carrier status.58 Of these 18 variants, only one justified clinical action (closer surveillance); 12 were of uncertain clinical significance, attracting considerable disagreement among geneticists. Moreover, whole genome sequencing led to all participants undergoing a median of 1-3 initial follow-up referrals and diagnostic tests per person, incurring a median estimated total cost of between $351 (£270; €318) and $776, for no additional clinical benefit.
Criteria for choosing to use predictive genetic tests
Clinicians, guideline developers, and policy makers need a framework for determining when and for whom genetic tests are indicated and reimbursable. Several frameworks exist,59 but currently there is no consensus on the best approach. To fill the void, we propose the following points for consideration:
Are the genetic variants being tested a valid predictor of disease—is there a proven association with disease and clinically meaningful genetic effects?
Does testing for the variants provide more accurate predictions of disease risk, which better inform clinical decisions and incentivise individuals to engage in effective risk reducing behaviours?
Does genetic testing for the variants pose risk of harm to individuals due to misinterpretation of the degree of disease risk, leading to avoidable intervention, emotional, and psychosocial dysfunction or third party discrimination?
Is testing for the variants feasible in routine practice regarding current reimbursement rules, access to testing modalities, and availability of experts for interpreting test results and issuing appropriate recommendations?
Moving from genetic risk profiling of rare monogenic disorders within families to wider polygenic profiling for more common diseases in large asymptomatic populations carries considerable potential for harm and waste. When choosing predictive genetic tests, clinicians and consumers must avoid commercial hype, ask relevant questions, and advocate for rigorous evaluation.
Contributors and sources: IAS conceived the idea, undertook the primary research, and drafted the first manuscript. JA and RM critically reviewed the manuscript, provided additional references and assisted in redrafting the manuscript. IAS is the guarantor who accepts full responsibility for the finished article, had access to any data from literature searches, and controlled the decision to publish. IAS is a clinical epidemiologist and and RM is an academic and former medical journalist who has both studied and reported widely on inappropriate, non-evidence-based care; JA is a clinical epidemiologist who has analysed results of multiple genome-wide association studies and written about the accuracy and utility of genetic tests.
Competing interests: We have read and understood BMJ policy on declaration of interests and have none to declare.
Patient and public involvement: We did not involve patients or the public in this work which comprised secondary research of published literature. There are no individual patient results to disseminate to these groups.
Provenance and peer review: Not commissioned; externally peer reviewed.