The emerging field of polygenic risk scores and perspective for use in clinical care

Abstract

Genetic testing is used widely for diagnostic, carrier and predictive testing in monogenic diseases. Until recently, there were no genetic testing options available for multifactorial complex diseases like heart disease, diabetes and cancer. Genome-wide association studies (GWAS) have been invaluable in identifying single-nucleotide polymorphisms (SNPs) associated with increased or decreased risk for hundreds of complex disorders. For a given disease, SNPs can be combined to generate a cumulative estimation of risk known as a polygenic risk score (PRS). After years of research, PRSs are increasingly used in clinical settings. In this article, we will review the literature on how both genome-wide and restricted PRSs are developed and the relative merit of each. The validation and evaluation of PRSs will also be discussed, including the recognition that PRS validity is intrinsically linked to the methodological and analytical approach of the foundation GWAS together with the ethnic characteristics of that cohort. Specifically, population differences may affect imputation accuracy, risk magnitude and direction. Even as PRSs are being introduced into clinical practice, there is a push to combine them with clinical and demographic risk factors to develop a holistic disease risk. The existing evidence regarding the clinical utility of PRSs is considered across four different domains: informing population screening programs, guiding therapeutic interventions, refining risk for families at high risk, and facilitating diagnosis and predicting prognostic outcomes. The evidence for clinical utility in relation to five well-studied disorders is summarized. The potential ethical, legal and social implications are also highlighted.

Introduction

Precision medicine is the tailoring of patient management based on phenotype, clinical history and genetic architecture. There is an increasing appreciation for utilizing this approach to identify individuals at increased risk of developing complex disease and subsequently mitigate risk through medical and behavioral interventions. A large portion of genetic risk for complex diseases stems from the interactive effect of multiple low- or modest-risk single-nucleotide polymorphisms (SNPs). Combining the weighted values of each of these variants results in the generation of a polygenic risk score (PRS), also known as a genetic or genomic risk score, or polygenic score (PGS) (1). More recently, studies have begun to report clinical and personal utilities of PRS (2–6), resulting in commercial availability (7). Despite recent advances, there are still limitations to PRSs and debate in the literature over their clinical utility (8,9). In this article, we will review the development and evaluation of PRSs to stratify disease risk and the clinical utility of PRSs across five diseases. The potential ethical, legal and social implications are also discussed.

Figure 1

The process for creating, calculating, testing and validating a PRS.

Genome-wide association studies—the origin of PRS

Genome-wide association studies (GWAS) have broadened our understanding of the genetic architecture of disease by identifying thousands of SNPs associated with diseases such as various cancers, coronary artery disease (CAD) and psychiatric disorders (10). By convention, SNPs are used to refer to genetic variants that are relatively common in the population, e.g. have a minor allele frequency (MAF) of >0.05 or >0.01 depending on the study. In general, disease-associated SNPs identified by GWAS confer a modest disease relative risk of 1–2, with the majority <1.2. This contrasts with high penetrance variants (e.g. in BRCA1/2 associated with breast cancer) identified through familial studies, which are rare (in aggregate population frequencies are ~MAF 0.006; with each individual mutation very rare), and associated with a high relative risk of >5 (11,12). In general, the most strongly associated SNPs identified through GWAS are not causal, rather they are markers for a functional variant in linkage disequilibrium (LD) and are thus called linked or marker SNPs. There is some evidence to suggest that the causal SNPs are more likely to be located in regulatory as opposed to exonic regions (13).

The validity of PRS is inherently dependent on the rigor and comprehensiveness of the GWAS on which it is based. Both data sets must be of highest standards and the assumptions and limitations must be adequately addressed (14). Over the past decade, there has been significant growth in the number of published GWAS and the cohort sizes (10). Unfortunately, these numbers do not reflect an increase in ethnic diversity, with approximately 80% of published GWAS conducted in populations of European ancestry (15). Naturally, the lack of diversity impairs the transferability of risk estimates and clinical utility to other ethnic populations (15–17). Concerted research efforts are underway to address this issue (18).

Development and validation of PRSs

Literature on how PRSs are developed is evolving, and there is no single universally accepted methodology (Fig. 1) (3). In general, PRSs are constructed by summing the total number of risk alleles an individual carries, weighted by an estimation of the impact of each allele on disease risk, e.g. the individual allele log odds ratio. When developing a PRS, consideration needs to be given to the statistical disease association threshold for determining selection of variants, with some authors restricting inclusion to independent (no LD) variants reaching genome-wide level of significance (P < 5 × 10⁻⁸) (19–21), while others using scores constructed with independent SNPs at varying significance thresholds (pruning and thresholding) or approaches that model millions of genome-wide variants together (a genome-wide PRS) (22–24) (Table 1).

Table 1

Summary of evidence for the utility of PRS across five diseases

Disease	Clinical use	Clinical risk factors	Monogenic risk factors	Method of variant selection for PRS	Studies assessing utility of PRS in non-European populations	PRS added to existing risk prediction models	Cost-benefit analysis	ELSI studies
Breast cancer	Risk-stratified population screening programs (19,25–29) Refining risk for high-risk families (6,30–33) and modifying penetrance of monogenic risk genes (17,34,35) Differentiating between disease subtypes (estrogen positive and negative) (19,28,36) and predicting diagnostic outcomes (37–39)	Age, family history, breast density, age at menarche, age at menopause, nulliparity, age at first childbirth and history of hormone replacement therapy, alcohol consumption, BMI and smoking history	Several high- and moderate-risk genes including: BRCA1/2, CHECK1, ATM, PALB	Restricted (6,17,19,26–28,30–46) Genome wide (25, 29)	Asian American (41), East Asia (40), Taiwan (42), Singapore–Chinese (43), African (44), Hispanic (45), African American (45, 46)	Improved AUC and risk classification: BOADICEA (27,31,32), Gail (26,31,45), BRCAPRO (31), Tyrer-Cuzic (31,45)	Cost benefit to risk-stratified population screening (47)	Early literature limited to a few qualitative (48,49) and cross-sectional studies (50)
Colorectal cancer	Risk-stratified population screening programs (51–55) Refining risk for high-risk families by modifying penetrance of monogenic risk genes (17) and does not modify risk for MMR genes (56)	Age, gender, family history BMI, regular NSAID use, smoking history, physical exercise and fiber, red meat and vegetable intake	Several high and moderate risk genes including: MLH1, PMS2, MSH2, MSH6, MUTYH	Restricted (51–65)	Japanese (58,59), Korean (60,61), Chinese (62,63), Taiwanese (64)	Improved AUC when PRS incorporated with family history and phenotypic risk factors (57,65) Improved risk classification when added to family history only (53–55) and family history with phenotypic risk factors (51)	Potential for cost benefit with increase in AUC value, reduction in testing costs or increased adherence to screening program (66,67)	Early literature is limited to studies assessing impact of PRS based on 3 SNPs (68,69)
Prostate cancer	Risk-stratified population screening programs (20,29,70–72) and improved PPV of PSA testing (73–75) Refining risk for high-risk families and modifying penetrance of monogenic risk genes (76)	Age and family history	BRCA1/2	Restricted (20,70–85) Genome wide (29)	Chinese (78,79), East Asian (80) African (81) (82) Latino (83)	Higher AUC for PRS than for family history (84,85) Improved AUC when combined with family history (84) and clinical risk a score (77) Improved risk classification when added to family history (72,84) and to a clinical risk score	PSA screening for men at a higher polygenic risk could reduce overdiagnosis and is cost effective (86)	Early literature limited to qualitative studies (87,88) and a study assessing impact of PRS to guide PSA testing (89)
Coronary artery disease	Guiding therapeutic interventions for statin therapy (90,91) Refining risk for hypercholesterolemia (92–94)	Age, gender, family history, cholesterol levels, systolic blood pressure, physical activity, body mass index and smoking history	LDLR, APOB, PCSK9	Restricted (16,21,90,91,93,95–101) Genome wide (16,25,29,92,94,102–104)	East Asian (21), Chinese (101), Japanese (100), Pakistani (99), Hispanic (16), Latinos (21,100), African American (16,21,100)	Mixed results: Improvements to Framingham (92) and clinical risk factors (25,92,96–98,102) Minimal changes to Framingham (95), QRisk3 (103), pooled cohort equation (104) and clinical risk factors (97). Reduced calibration when PRS included to model (104)	Cost benefit (105)	Qualitative study (106) and RCT assessing impact of PRS on health behavior and psychosocial outcomes (107,108)
Schizophrenia and bipolar disorder	Facilitating diagnosis (109–111) and predicting diagnostic outcomes (112–114) Response to lithium treatment (115) Moderates effects of SCZ in carries of high-risk copy number variants (CNV) (116)	Several including stressful life events, childhood trauma, socioeconomic status drug use, family history	Limited: rare CNVs and associated risk with some chromosomal disorders	Restricted (110,113) Genome wide (109,111,112,114,116)	Asian (115,117), African (114)	Interaction between PRS SCZ family and socioeconomic status (118)	No studies	Early literature limited to a single qualitative study (119)

A restricted PRS has the advantage of reducing type I errors by applying more stringent criteria to variant selection. However, these variants account for a small proportion of disease heritability, thereby limiting the PRS predictive performance. In contrast, a pruning and thresholding, or genome-wide PRS approaches, aim to improve the PRS accuracy by including a wider range of genomic variants. These approaches have the potential to reduce trait specificity, whereby many included SNPs are relevant to multiple diseases/traits (pleiotropy). Furthermore, LD needs to be considered, whereby correlations between nearby SNPs may lead to an overrepresentation of high LD regions in the PRS. In pruning and thresholding, this is addressed by LD pruning, where SNPs are filtered to a single independent SNP for each LD block/associated region. Genome-wide PRSs, computational tools model and account for LD between all included SNPs (e.g. LDpred) (22). Despite the advantages, there are mixed reports on whether genome-wide PRS approaches that model LD improve accuracy over strict LD pruning or a restricted PRS (22–24,120).

When developing either a restricted or genome-wide PRS, it is critical to use separate populations for training and validation purposes. In a GWAS (or more recently GWAS meta-analyses), a specific set of SNPs is genotyped on a training sample, and effect sizes are estimated for each marker’s association with the trait of interest. These weights are then used to assign individualized PGSs in an independent validation population (121). Additionally, independent training datasets can also be used to determine the best performing genome-wide PRS (e.g. determining which P-value threshold to use for pruning and thresholding, or genome-wide approaches), with many studies evaluating several PRSs (25,26). A validation dataset is used to confirm the scientific validity of individual SNPs and the mathematical algorithm used to combine them (14).

There are potential pitfalls when generating a PRS. Risks include using the same dataset for generating and training the PRS (leading to overfitting), the inclusion of redundant SNPs (e.g. not accounting for LD), and not addressing gene–environment interactions that influence risk. In addition, it is critical to address ancestry, recognizing that the accuracy of imputed SNPs and the correlation between PRS SNPs and causal variants, will be population-dependent. Furthermore, direction and magnitude of risk for associated alleles can vary between different populations (15,16,122). Hence, in addition to performing GWAS in diverse ancestries to generate appropriate PRS, there is a pressing need to diversify the studies applying PRS, with recent estimates suggesting 67% of PRS studies exclusively evaluated European populations, 19% focused on individuals of East Asian ancestry, while only 3.8% were cohorts of Africa, Hispanic or Indigenous populations (123). Finally, in the future, it is likely that PRSs will need to account for epistatic interactions between SNPs. Specifically, two or more SNPs may be dependent, whereby their combined effect could be higher or lower than their individual ORs (124).

Evaluating PRS

Until recently, no guidelines existed for reporting and evaluating performance of PRS (1), resulting in significant variability in outcomes across studies. Typically, studies standardize PRS into quantiles, with cases being more likely to fall into the highest quantiles (1 or 5%) than controls (30). Additionally, it is recommended that the predictive ability of PRS, in the form of calibration and discrimination, should be reported in all studies (1). The discriminatory power of PRS has most commonly been assessed by the area under the curve (AUC), a measure of the overall probability that the predicted risk is higher in cases than controls (125). The overall discrimination of PRS is bound by the disease’s heritability, as such this measure will be disease dependent (126). Of note, the AUC has limited ability in predicting diagnostic outcomes (127). Thus, additional measures such as Cox regression and Kaplan–Meier curves are also used. Association between PRS and disease outcomes is then reported with standard epidemiological measures, such as odds ratios and hazard ratios, per standard deviations. Such measures can provide evidence of clinical utility by estimating the percentage of the population with different levels of risk and corresponding risk management strategies.

Polygenic information incorporates high-penetrance (Mendelian) gene risk and the genetics of clinical risk factors, opening the possibility of developing holistic risk calculators that account for the effect of several genetic and non-genetic risk factors. This approach has been evaluated across several diseases and risk models including BOADICEA, BRCAPRO, BCRAT and IBIS for breast cancer and Framingham calculator for CAD (27,31,92,102). Results have been mixed with some studies suggesting improved risk predictions with inclusion of PRSs into the models (27,77,92) and others reporting minimal changes (95,103,104).

Efforts are being made to standardize the development and reporting of PRSs (1,14), and recently, the PGS Catalog was developed (www.pgscatalog.org). The PGS is an open database of PGSs and the relevant metadata required for accurate application and evaluation. This valuable resource tracks the development, application and evaluation of the predictive performance of published PRSs. Nevertheless, guidelines for developing PRSs will be dependent on the genetic architecture of the disease, the availability of GWAS summary data and discovery and target populations (3).

Clinical utility

Given the broad nature of clinical applications for PRS, assessments of utility should be context- and disease-specific. Broadly speaking, the clinical utility of PRS can be divided into four categories: (i) informing population screening programs, (ii) refining risk for individuals undergoing genetic testing for monogenic risk genes, (iii) guiding therapeutic interventions and (iv) facilitating diagnosis and predicting health outcomes (Box 1). The next section will review the ways in which PRS may be used in clinical practice and the evidence for clinical utility using breast cancer as an example (Table 1). Measures of clinical utility include whether the test improves health outcomes relative to existing assessments, cost-benefit analysis, exploration of the risk and benefits of offering the test, and assessments of resources (facilities, personnel and educational materials).

Box 1:

The use of PRS for breast cancer

Arguably, breast cancer is one of the most well-researched diseases for the utility of PRSs (6). There are currently hundreds of variants associated with breast cancer susceptibility, which reach a genome-wide level of significance (128–130). GWAS have also been used to identify 32 variants associated with different tumor subtypes (e.g. tumor grade and estrogen receptor (ER) status) (131). Studies have assessed the utility of breast cancer PRS in the following domains:

• Risk-stratified population screening programs (19,25–29,36): PRS identifies women at higher genetic risk who reach the threshold for population screening in the UK at a younger age (red line, equivalent to 2.4% risk for a woman at age 47 years who is eligible for population screening) (See Fig. 2) (28). A combined risk that includes PRS, family history, breast density and other risk factors is estimated to identify ~13% of the population at moderate or high risk of developing breast cancer (27).

  

• Figure 2. Cumulative breast cancer risk by PRS percentiles. (Reproduced from Mavaddat et al. (28): djv036 with permission).

• Refining risk assessment for breast cancer families with no pathogenic variants (PVs) in monogenic risk genes (30–33,132,133): Less than 25% of the familial relative risk for breast cancer is explained by PVs in monogenic risk genes (134). PRS identified women at genetically high risk of breast cancer who tested negative for monogenic risk genes (32). Higher PRS measures were reported in women who tested negative for BRCA1/2 mutations compared to mutation carriers (30).

• Modifying penetrance of monogenic risk genes (17,34,35): Incorporating PRS into genetic testing for high- and moderate-risk genes can improve accuracy of risk estimation and aid risk management decisions for women (e.g. uptake prophylactic mastectomy and risk-reducing medication). A PRS based on variants associated ER-negative disease displayed stronger association to overall breast cancer risk for BRCA1 PVs carriers than ER-positive (34).

• Differentiating between disease subtypes (19,28,31,36–38,40): Present PRSs are a more accurate predictor of ER-positive disease than ER-negative. This information may inform decisions on uptake of chemoprevention with Tamoxifen, which is a well-established risk-reducing strategy for ER-positive disease (135).

• Predicting diagnostic outcomes (37–39): Higher PRS are associated with increased risk of contralateral disease. This information may aid treatment decisions for women with a personal history of breast cancer (e.g. bilateral mastectomy versus conservative surgery). Interestingly, a high PRS is also associated with more favorable prognoses including being more likely to be diagnosed during routine screening, having ER-positive disease, a smaller tumor size and being less likely to present with distant metastasis.

Informing population screening programs

Currently, most population screening programs utilize age as the sole criterion for determining when screening should commence. However, recent studies have demonstrated successful population risk stratification utilizing the combination of PRS, family history, clinical, demographic and lifestyle risk factors (Box 1) (20,26,28,51,52,73,136–138). This personalization of population screening has the potential to reduce disease morbidity and mortality, by targeting high-risk individuals for more intensive screening. Evidence of the added value of polygenic information is also seen in prostate cancer studies, which report an increased positive predictive value of prostate-specific antigen among men with higher PRS (73). However, consideration needs to be given to the acceptability of risk-based surveillance, with previous studies suggesting a reluctance to reduce screening and poorer adherence in lower risk groups (139–142). Studies have reported a health economic benefit to PRS risk-stratified screening for prostate cancer (20,86) and breast cancer (47), but not for colorectal cancer population screening (66,67). Nevertheless, Naber et al. (66) estimated that a cost benefit for colorectal cancer screening may be achieved with a 0.05 increase in the AUC value, a 30% reduction in testing costs or greater adherence to screening programs.

Guiding therapeutic interventions

Polygenic information has been shown to identify individuals at increased disease risk (19,73,92), opening the possibility of implementing preventative or risk-reducing interventions. Such interventions can include risk-reducing medications and lifestyle changes (Table 1). Perhaps more so than any other diseases, the clinical utility of PRS-guided interventions has been commonly assessed for CAD and prophylactic use of cholesterol-lowering therapies (e.g. statins). Recent studies have demonstrated a greater benefit of statin therapy among individuals with higher PRS, suggesting that PRS may be used to inform treatment management (90,91). Despite the potential benefits, there has been mixed reports on whether PRS improves risk prediction over existing risk factors (e.g. age, gender, cholesterol levels and lifestyle factors), with some studies reporting improved accuracy (25,92,96,102), while others report minimal changes (103,104). Such differences can be attributed to risk algorithm variations between models, the predictive power of the PRS and/or the heritability of the trait (126). Questions also remain regarding the impact of PRS on health behavior and ideal age to offer PRS testing, with some arguing that greater clinical utility will be achieved by providing PRS at a younger age, thereby potentially influencing lifestyle risk factors (3,143).

Refining risk for families at high risk

Traditionally, clinical genetic services have focused on genetic testing for monogenic risk genes (e.g. BRCA1/2, PALB2, LDLR, MLH1) associated with high-to-moderate disease risk. PVs in these genes account for a small percentage of disease heritability in the wider population, with further heritability attributable to polygenic factors. The implementation of PRS in genetic services has the potential to refine risk and improve clinical management. For example, PRS may be used to calculate residual genetic risk for individuals who are negative for variants in high- and moderate-risk genes (6,32). Additionally, studies have also demonstrated a multiplicative effect of PRS on PVs in high- and moderate-risk genes across several disorders including breast cancers (17,34,35), prostate cancer (76), colorectal cancer (17, 56), and familial hypercholesterolemia (17,93,97,144,145) and melanoma (136).

Testing for PRS in clinical genetic services is at an introductory phase, with testing conducted alongside standard assessments of personal and family history. Recently, the CanRisk tool was developed to estimate breast cancer risk for women with a positive family history (27). This is the first tool to incorporate the effects of monogenic (e.g. BRCA1/2), polygenic, family history and non-genetic risk factors, such as breast density and lifestyle, to estimate disease risk. Integrating PRS with traditional monogenic risk testing promises to personalize risk prediction and ultimately facilitate management decisions for high-risk family members, such as informing uptake of prophylactic surgery or risk-reducing medication. However, the extent to which this information will alter clinical decisions and improve health outcomes is yet to be explored.

Facilitating diagnosis and predicting prognostic outcomes

For many diseases, obtaining accurate diagnoses can be challenging and time consuming. For example, it has been estimated that up to 45% of individuals with bipolar disorder are initially misdiagnosed, potentially leading to ineffective or harmful treatments (146). The potential utility of PRS in differentiating diagnoses has been evaluated, specifically distinguishing between schizoaffective and cases from others bipolar disorders cases (109) or differentiating between type 1 and type 2 diabetes (147). In addition, PRS may add prognostic value in psychiatric disorders, where a high PRS is associated with a higher risk of psychotic symptoms in patients with bipolar disorder (110–112). Such research is still in the discovery phase, and thus, to date, there is limited evidence for clinical utility. Nevertheless, if proven to be effective, PRS may support clinicians and patients to make informed treatment decisions by clarifying diagnosis and diagnostic outcomes.

Ethical, legal and social implications

Finally, the ethical, legal and social implications should be considered when assessing the utility of PRS. There have been limited studies assessing the acceptability of PRS for patients and providers and the psychological responses to receiving PRS. Early research suggests limited adverse psychological outcomes in the short-term (148), though research is needed to assess long-term psychological responses. Similarly, data on behavioral responses are scarce, with previous research suggesting limited changes in health behavior after receiving genetic risk information (149,150). Thus, there is a pressing need to develop models to communication PRS and interventions that support positive health behavior change. Future research needs to consider the different settings in which PRS is provided, with initial data suggesting differences based on disease, population and context in which this information is provided (48,119,151). Importantly, the ethical implications of PRS should also be considered. For example, several commentaries have raised concern about the validity of PRS across minority populations, leaving open the possibility for increased health disparities (15).

Future directions

To date, PRSs have primarily resided in the research realm, and while there are many questions remaining, clinical implementation has begun for some diseases (e.g. breast cancer, prostate cancer, etc.) is already taking place and is on the horizon in other areas. Nevertheless, several clinical trials are currently evaluating PRS utility, including the WISDOM trial for population breast cancer risk and GeneRisk for CAD (106,152). Such studies will provide important real-world data regarding utility. If clinical utility can be proven, then implementation challenges will need to be addressed, including communication of results, clinician education, associated costs and social implications. Ultimately, PRS development and evaluation will need to consider the population characteristics and the disease of interest.

Conclusion

Polygenic information, when integrated with other genomic and clinical risk factors, has the potential to customize healthcare to maximize clinical and public health benefits. There are several methodological and reporting elements that need to be addressed before clinical utility and implementation can be fully realized across all populations. Effective communication of this information to patients will be key to maximizing clinical benefits. The efficacy of their use will depend on the continued development of methods that exploit them, proper analysis and applications in various populations, appropriate interpretation and an understanding of their strengths and limitations.

Conflict of Interest statement. Shelly Cummings is an employee and stockholder of Myriad Genetics Inc., Salt Lake City, UT.

Funding

Aideen McInerney-Leo is funded by a National Health and Medical Research Council (NHMRC) Early Career Fellowship (ID 1158111). UQDI is located in the Translational Research Institute, which is supported by a grant from the Australian Government.