Polygenic risk scores in pharmacogenomics: methodological challenges, current applications, and perspectives for clinical implementation

Abstract

Inter-individual variability in drug response remains a major challenge in clinical pharmacology. While clinical pharmacogenetics is largely based on actionable single-gene variants, increasing GWAS evidence indicates that many drug-response phenotypes are polygenic. Polygenic risk scores (PRS) offer a framework to capture cumulative effects across pharmacokinetic, pharmacodynamic, and disease-related pathways, enabeling probabilistic risk stratification beyond monogenic models.

This review summarizes the current status of PRS in pharmacogenomics (PGx-PRS), highlighting key challenges including limited sample sizes of drug-exposed cohorts, heterogeneous phenotype definitions, ancestry-related portability issues, and incomplete representation of complex pharmacogenes such as CPY2D6. Current applications are discussed across cardiovascular, psychiatric, and oncological settings, where PRS show emerging potential for benefit-risk stratification but remain insufficiently validated for routine care.

Clinical implementation will require standardized methodologies multi-ancestry validation, integration with rare-high impact ADME (Absorption, Distribution, Metabolism, and Excretion) variants, and clearer regulatory pathways. With these advances, PGx-PRS may become clinically relevant tools for precision prescribing and drug development.

Introduction

Inter-individual variability in drug response remains a persistent and clinically significant challenge in pharmacology, as patients receiving identical medications and doses frequently exhibit divergent outcomes with respect to both therapeutic efficacy and safety [1]. To date, clinical pharmacogenetics has been predominantly grounded in well-characterized single-gene–drug interactions – most prominently involving cytochrome P450 (CYP) enzymes or other phase I or II enzymes such as the thiopurine S-methyltransferase (TPMT) – for which actionable prescribing recommendations have been established by expert consortia such as the Dutch Pharmacogenomics Working Group (DPWG) and the Clinical Pharmacogenetics Implementation Consortium (CPIC) [2 – 6].

Paradigmatic examples of such monogenic associations with established clinical relevance include dihydropyrimidine dehydrogenase (DPYD) variants predicting severe fluoropyrimidine toxicity and the pharmacogenomic interaction between irinotecan or atazanavir and UGT1A1 polymorphisms causing neutropenia or affecting bilirubin metabolism [7 – 10]. These examples illustrate how specific germline variants in drug-metabolizing enzymes or transporters can confer substantial and predictable changes in drug disposition or toxicity risk, thereby enabling genotype-guided prescription within defined gene–drug pairs. For a comprehensive overview see Lauschke & Ingelman-Sundberg 2026 [11].

While these monogenic associations have demonstrated clear clinical utility, they collectively account for only a fraction of the observed inter-individual variability in drug response. Increasing evidence from genome-wide association studies (GWAS) indicates that numerous pharmacological phenotypes – including treatment efficacy, dose requirements, and susceptibility to adverse drug reactions – exhibit a polygenic architecture shaped by the cumulative effects of many genetic variants of individually small effect, acting across pharmacokinetic, pharmacodynamic, and disease-related pathways [12, 13] (Figure 1). Accordingly, this review focuses primarily on common-variant, GWAS-based approaches.

Figure 1:

Pharmacogenetics vs. pharmacogenomics.

Classical examples include the genetic determinants of warfarin dose variability, where variants in VKORC1 and CYP2C9 explain a substantial proportion of inter-individual variation but additional loci contribute to residual variability [14, 15]. However, despite this biological rational and genetic associations, prospective randomized controlled trials evaluating genotype-guided dosing have produced inconsistent results, and clinical implementation remains limited and geographically variable. This illustrates that even well-established genetic-predictors do not necessarily translate directly into routine clinical practice. Similarly, GWAS have identified multiple loci influencing clopidogrel response and platelet reactivity, extending beyond the well-known CYP2C19 variants [16]. For adverse drug reactions, genome-wide analyses of statin-associated myopathy have revealed variants in SLCO1B1 as major contributors while suggesting additional polygenic influences on toxicity risk [17, 18]. In psychiatric pharmacotherapy, large-scale studies of antidepressant treatment response and antipsychotic efficacy likewise demonstrate highly polygenic architectures with individually modest effect sizes distributed across numerous loci [19 – 21]. Collectively, these findings provide a conceptual rationale for moving beyond single-variant pharmacogenetic models toward polygenic frameworks capable of capturing the broader genomic contribution to drug response variability.

Accordingly, there is a growing scientific and clinical need to extend pharmacogenetic testing beyond single-gene paradigms and to advance the development and implementation of polygenic risk score (PRS)-based strategies tailored specifically to pharmacological phenotypes [22]. However, several methodological, scientific, and clinical challenges remain to be addressed before such approaches can be translated into routine clinical application. Against this background, the present article provides (i) a critical overview of the methodological challenges inherent to PRS development in the context of pharmacogenomics, (ii) a synthesis of current applications of PRS across therapeutic areas, and (iii) a framework for the requirements underlying responsible integration into routine clinical care.

Methodological challenges in PRS development in the context of pharmacogenomics

The development of PRS for pharmacogenomics applications involves distinct methodological challenges compared to disease-oriented PRS, most notably with respect to the selection and interpretation of appropriate data sources. A central issue is the frequent reliance on disease-derived GWAS summary statistics as an indirect proxy for pharmacological phenotypes. Although such datasets offer large sample sizes – often exceeding one million participants – and correspondingly robust statistical power, the extrapolation of disease-associated variants may systematically overlook drug-specific genetic effects that are mechanistically unrelated to disease susceptibility per se. For example, a coronary artery disease (CAD) PRS does not necessarily capture the genetic determinants of statin efficacy or statin-induced adverse effects, which are determined more by pathways of lipid metabolism, hepatic transport, and skeletal muscle toxicity than by the risk of atherosclerotic disease itself [23, 24]. Similarly, a genetic predisposition to schizophrenia is not equivalent to predictors of response to clozapine treatment, in which immune, metabolic, and drug-metabolism pathways play a central role – pathways that differ mechanistically from the underlying cause of disease [25]. These examples illustrate the fundamental conceptual distinction between gene–disease associations and gene–drug interactions.

A further critical limitation concerns the weighting of variant effect sizes: estimates derived from disease GWAS may not translate to pharmacological outcomes, particularly when treatment response reflects a composite of pharmacokinetic, pharmacodynamic, and disease-modifying biological processes. Although direct pharmacogenomic GWAS conducted in drug-exposed cohorts represent a more analytically specific alternative they remain comparatively underpowered due to limited sample sizes, heterogeneous phenotype definitions, and practical challenges in capturing adherence patterns, dosing history, and co-medication effects across study populations [26, 27].

Population structure introduces additional complexity into PRS construction. Differences in linkage disequilibrium (LD) patterns between ancestral groups can result in the tagging of non-causal proxy variants rather than true causal loci, and the pronounced overrepresentation of individuals of European-ancestry – often exceeding 80 % in many datasets and frequently even higher in pharmacogenomic GWAS cohorts – substantially limits PRS portability and predictive accuracy in genetically diverse populations [28].

An additional limitation of particular relevance to pharmacogenomics is the inadequate representation of structurally complex pharmacogenes within GWAS-derived datasets. Key pharmacokinetic loci such as CYP2D6 – encoding the most polymorphic drug-metabolizing enzyme in humans – exhibit copy number variation, hybrid gene arrangements, and haplotype-level star-allele diversity that are not reliably detected by conventional SNP microarrays or accurately inferred through LD-based imputation, thereby leading to the systematic underrepresentation of major pharmacokinetic determinants within conventional PRS frameworks [29, 30].

Addressing this technical gap necessitates the integration of complementary genotyping and sequencing strategies. Conventional microarrays, while cost-effective and scalable for genome-wide discovery, are insufficient as the sole genotyping strategy for pharmacogenomic PRS which require accurate representation of complex ADME (Absorption, Distribution, Metabolism, and Exrection) genes. Short-read next-generation sequencing (NGS) enables higher-resolution variant detection than microarrays and has improved the pharmacogenetic characterization of structurally complex loci; however, it remains limited in its ability to resolve repetitive regions and ambiguous haplotype phasing. Long-read sequencing technologies, such as those provided by Pacific Biosciences (PacBio) and Oxford Nanopore Technologies (ONT), enable substantially improved resolution of structural variants, copy number variation, and phased haplotypes, thereby enabling more comprehensive characterization of complex pharmacogenetic loci such as CYP2D6 [31 – 33]. Integrating long-read sequencing data with genome-wide association approaches may therefore represent a methodological prerequisite for clinically meaningful polygenic pharmacogenomic models.

Finally, the choice of PRS construction methodology must account for the hybrid genetic architecture of drug response. Conventional disease-oriented algorithms such as LDpred2 [34] or PRS-CS [35] assume highly polygenic architectures without incorporating pharmacological prior knowledge, whereas PGx-adapted approaches – including PRS-PGx-Bayes and mtPRS-PCA – attempt to integrate established ADME gene knowledge, pathway-level information, and multi-trait covariance structure to better reflect drug-specific biology [36, 37].

Collectively, these considerations underscore that PRS development in pharmacogenomics is not a straightforward extension of disease genetics but requires purpose-built data integration strategies aligned with the exposure-dependent nature of pharmacological phenotypes and the mechanistic principles of clinical pharmacology [38, 39].

Current applications of polygenic risk scores in pharmacogenomics

To date, the application of PRS in pharmacogenomic research has been explored most extensively in cardiovascular medicine, where large, well-phenotyped cohorts and randomized clinical trial datasets have enabled comparatively robust analyses. In the context of lipid-lowering therapy, several studies have demonstrated that a CAD-derived PRS can stratify the magnitude of benefit from statin therapy in primary prevention. It has been shown that individuals in the highest PRS strata exhibit greater relative risk reduction than those at lower genomic risk [40, 41]; providing one of the strongest pieces of evidence to date for the clinical utility and benefit of a pharmacogenomic PRS. Similar approaches have been applied to PCSK9 inhibition, where post-hoc analyses of major outcomes trials suggest that a PRS may distinguish between absolute and relative treatment benefits, supporting a potential role for genomic risk stratification in identifying patients most likely to derive a net clinical benefit [42, 43]. Beyond lipid-lowering therapies, exploratory analyses have examined PRS-informed response to antiplatelet therapy, including clopidogrel [16], and to beta-blockers in heart failure [44], frequently employing SNP-by-treatment interaction models to identify genetic modulation of therapeutic effect independent of baseline disease risk.

In psychiatry, PRS applications have focused primarily on predicting treatment response and adverse effects, although the evidence base remains less consistent across studies. For antipsychotic pharmacotherapy, PRS derived from GWAS in schizophrenia have been investigated in relation to clozapine dosing requirements and treatment-resistant schizophrenia, findings across independent cohorts have been heterogeneous reflecting the inherent difficulty of disentangling disease genetic liability from pharmacologically relevant drug-response biology [45 – 47]. Studies evaluating PRS for antidepressant response to selective serotonin and serotonin-norepinephrine reuptake inhibitors (SSRIs/SNRIs) – including multimodal approaches integrating clinical and electroencephalographic (EEG) biomarkers– have generally yielded modest predictive performance [48, 49]. Lithium response represents one of the more mature psychiatric PRS applications. Analyses from the International Consortium on Lithium Genetics (ConLiGen) have demonstrated statistically significant associations, though with limited explained variance (approximately 2 – 3 %; [50]). More recent multimodal approaches combining PRS with clinical and demographic factors have shown improved predictive performance for lithium response [51], suggesting that integrated models may better capture the compelex genetic architecture underlying treatment outcomes while also illustrating both the promise and current ceiling of polygenic prediction in neuropsychiatric pharmacotherapy.

Beyond cardiovascular and psychiatric medicine, emerging investigations are exploring PRS-guided strategies in oncology – particularly for the prediction of chemotherapy toxicity and response to targeted molecular therapies [52]. This is highly relevant given the substantial inter-individual variability in cancer treatment tolerability beyond established monogenic predictors such as DPYD for fluoropyrimidines and UGT1A1 for irinotecan. Clinically important toxicities such as taxane-induced peripheral neuropthy, platinum-associated ototoxicity, and anthracycline-related cardiotoxicity likely involve broader polygenic contributions that are not captured by single-gene testing alone. Early studies suggest that PRS-based models may contribute to impoved toxicity stratification in these contexts, although prospective validation and standardized phenotype definition remain limited—for example, a PRS derived from genome-wide loci was correlated with paclitaxel-induced peripheral neuropathy severity in a chemotherapy-exposed cohort, illustrating the potential utility of germline polygenic prediction for adverse drug reactions [53]. In pain medicine, polygenic models are also being applied to characterize inter-individual variability in opioid efficacy and chronic pain trajectories [54].

Collectively, the existing literature indicates that PGx-PRS research is transitioning from proof-of-concept studies towards disease-specific implementation scenarios. Nevertheless, the field continues to be characterized by heterogeneous study designs, modest effect sizes, and limited prospective validation. These limitations reinforce the need for integrative models that combine genome-wide polygenic scores with established monogenic pharmacogenetic markers and detailed clinical phenotyping. In particular, such frameworks should explicitly incorporate rare high-impact loss-of-function or decreased-function variants in key ADME genes (e.g., DPYD, TPMT, CYP2D6, CYP2C19), which may individually exert larger effects than genome-wide PRS but are often not adequately captures by conventional GWAS-based approaches.

The route to clinical implementation

The translation of PRS into routine pharmacogenetic and clinical practice will require several coordinated scientific, methodological, and regulatory developments. A substantial expansion of pharmacogenomics GWAS cohorts is needed to close the persistent statistical power gap between disease genetics and pharmacogenomic studies, which still typically comprise only 1,000 to 10,000 drug-exposed individuals. Closely linked to this requirement is the inclusion of globally representative, multi-ancestry cohorts to enhance score portability and to prevent the exacerbation of pre-existing health disparities attributable to ancestral imbalances in training datasets [28, 55].

Equally important is the standardization of phenotypic definitions and outcome reporting frameworks, as drug-response endpoints such as ‘treatment response’ are currently operationalized with considerable heterogeneity with respect to timing, dosing, adherence criteria, and co-medication adjustment, thereby limiting cross-study comparability. Methodological harmonization must be coupled with improved integration of electronic health record (EHR) infrastructures to enable scalable phenotyping, real-world validation, and eventual clinical deployment, alongside systematic external validation in independent and geographically diverse cohorts [56, 57].

From a clinical translation perspective, interpretability remains a key barrier: most PGx-PRS studies report relative associations rather than absolute risk metrics, lack standardized decision thresholds for clinical application, and provide uncertain incremental predictive value over established clinical predictors. Addressing these gaps will require the development of expert consensus guidelines, PRS-specific clinical action thresholds, and demonstration of additive benefit over validated clinical risk tools by prospective outcome trials. Emerging work emphasizes integrative prediction models in which PRS are embedded within existing risk frameworks – for example, combining genomic risk scores with cardiovascular prevention tools such as QRISK [58] or heart failure prognostic indices such as MAGGIC [59] – while formally evaluating whether genetic and clinical effects are best modelled additively or multiplicatively.

Finally, successful clinical implementation will depend on transparent regulatory frameworks and clearly defined evidence standards, supported by agencies such as the EMA and FDA, to ensure analytical validity, clinical utility, and equitable deployment across patient populations.

In Germany, the application of pharmacogenetic testing, including emerging PGx-PRS approaches, must also be considered in the context of the Genetic Diagnostics Act (Gendiagnostikgesetz, GenDG), which distinguishes diagnostic from predictive genetic testing, with corresponding implications for consent procedures, medical oversight, and counseling requirements. § 3 No. 7 lit. c GenDG explicitly addresses pharmacogenetics: a genetic investigation aimed at clarifying whether genetic characteristics are present that may influence the effect of a drug is classified as a diagnostic genetic test. In a concrete treatment context, i.e. when the test informs the prescription of a specific drug, this classification is immediate.

Less self-evident from the statutory wording alone is the situation of preemptive pharmacogenetic testing peformed outside a specific treatment setting, for instance in the form of a pharmacogenetic passport, in which an individual is genotyped in advance for variants that may become relevant for future prescriptions. Authorative guidance on such questions of interpretation is provided by the Commission on Genetic Testing (Gendiagnostik-Kommission, GEKO), an interdisciplinary commission established under § 23 GenDG at the Robert Koch Institute, whose guidelines concretize the recognized state of science and technology in genetic teting in a legally binding manner, and which additionally publishes activity reports reflecting its ongoing assessment of developments in the field. In its activity report for the period 2022–2024 the GEKO classified preemptive pharmacogenetic testing, such as the issuing of pharmacogenetic passports, as “preemptive diagnostics” (“präemptive Diagnostik”), thereby positioning it as diagnostic rather than predictive testing under the GenDG [60]. In practical terms, this means that pharmacogenetic testing is subject to the medical practitioner reservation (Arztvorbehalt, § 7 GenDG) and to the offer of genetic counseling upon disclosure of results (§ 10 Abs. 1 GenDG) but not the more stringent requirements applicable to predictive testing under § 10 Abs. 3 GenDG.

Future developments

Future progress in PGx-PRS will likely be driven by convergent methodological innovations and expanded clinical integration. Advances in machine learning and artificial intelligence – including deep learning architectures specifically optimized for polygenic score construction – are expected to improve the modelling of non-linear variant interactions, the contribution of rare variants, and gene-by-environment interplay that current linear models inadequately capture [61]. The integration of multi-omics layers, encompassing transcriptomics, epigenomics, proteomics, and metabolomics [62], will enable biologically informed scores that more closely reflect functional molecular consequences rather than relying solely on statistical genomic associations.

In parallel, the concept of dynamic PRS – incorporating longitudinal phenotypic and exposure data, including age, comorbidity trajectories, and cumulative drug exposure – may enable time-dependent prediction of drug response and toxicity risk. Multi-ancestry methodological frameworks will be essential to ensure cross-population transferability, with trans-ethnic fine-mapping improving causal variant resolution and population-specific calibration strategies preventing performance disparities across clinically diverse patient groups [55].

Expanded clinical applications are envisioned through integrated risk frameworks that jointly model disease susceptibility and drug-response prediction, enabling simultaneous estimation of therapeutic benefit, optimal drug selection across competing therapeutic options, and individual susceptibility to adverse drug reactions within a unified algorithmic platform. Emerging therapeutic areas – including immunotherapy, gene-based therapies, and personalized vaccination strategies – may derive particular benefit from PGx-PRS approaches given their pronounced inter-individual response variability and high per-treatment cost.

A further key direction is precision dosing: PRS-guided dose titration, integrated with pharmacokinetic/pharmacodynamic (PK/PD) modelling and therapeutic drug monitoring (TDM), could shift pharmacogenomics from categorical prescribing decisions towards continuous, individualized drug exposure optimization. Notably, TDM is already well established in clinical psychiatry (e.g., clozapine and tricyclic antidepressants) [25] and transplantation medicinve (e.g., tacrolimus and ciclosporin) [63], providing an existing infrastructure for dose optimization and longitudinal monitoring. PGx-PRS may therefore have a particularly direct translation pathway in these fields complementing established TDM frameworks with pre-treatment genomic risk stratification and individualized dose targets.

From a translational and regulatory perspective, growing discussion concerns the embedding of PGx-PRS into drug development pipelines not only as companion diagnostic tools at the marketing authorization stage, but also at earlier developmental phases to enrich trial populations, mitigate toxicity signals, and improve the probability of clinical success. The recent approval of the anti-amyloid monoclonal antibody lecanemab illustrates how genetic information can inform pharmacogenetic risk stratification. Lecanemab is approved for patients with early Alzheimer’s disease and confirmed amyloid pathology, but treatment safety is influenced by APOE ε4, the strongest common genetic risk factor for late-onset Alzheimer’s disease. In the phase III CLARITY-AD trial, ε4 carriers—particularly homozygotes—showed markedly higher rates of amyloid-related imaging abnormalities (ARIA) compared with non-carriers, while efficacy appeared broadly similar across genotype groups. This led to regulatory recommendations for pre-treatment APOE genotyping to support clinical decision-making and safety management [64].

Importantly, this case also highlights why PRS may provide added clinical value beyond monogenic stratification. Although APOE status captures a major component of ARIA susceptibility, it does not fully explain inter-individual differences in treatment response or toxicity risk. A PRS framework could integrate multiple common risk variants across relevant pathways (e.g., neuroinflammation, vascular integrity, amyloid processing), potentially enabling a more comprehensive and clinically actionable prediction models to guide patient selection and monitoring intensity.

Conclusion

PRS shift pharmacogenetics to pharmacogenomics by capturing the polygenic architecture of drug response, treatment resistance, and adverse drug reactions more comprehensively than single-variant approaches. Early evidence – particularly from cardiovascular medicine and emerging psychiatric applications – demonstrates the potential of PGx-PRS to refine benefit–risk stratification, guide dosing decisions, and identify patients at higher likelihood of therapeutic response or drug toxicity. However, important challenges remain, including limited pharmacogenomic GWAS sample sizes, lack of methodological standardization, uncertain cross-ancestry transferability, and insufficient prospective evidence of clinical benefit. Progress will require larger multi-ancestry datasets, harmonized analytical frameworks, systematic integration with clinical and pharmacokinetic predictors, and prospective validation in clinical practice and drug development settings. With these advances, PGx-PRS carry the potential to transition from exploratory research tools into clinically actionable decision-support tools for precision prescribing.

Biographies

Dr. Carina M. Mathey

Dr. med. Martin Coenen

Prof. Dr. Dr. Ingolf Cascorbi

Dr. rer. nat. Per Hoffmann

Affiliations

¹Institute of Human Genetics, University of Bonn Medical School and University of Bonn, Germany

²University of Bonn, University Hospital Bonn, Institute for Clinical Chemistry and Clinical Pharmacology, Bonn, Germany

³Institute of Experimental and Clinical Pharmacology, University Hospital Schleswig-Holstein, Campus Kiel, Germany