A review of methods and software for polygenic risk score analysis

C+T

The Clumping and Thresholding (C+T) method is a widely used approach for calculating PRS. It identifies genome-wide significant variants and groups them based on linkage disequilibrium (LD), excluding those in strong LD with an index variant that has the lowest p-value in each group, this process helps to identify independent genetic variants associated with a trait (Wray, Goddard & Visscher, 2007; Euesden, Lewis & O’reilly, 2015). The method operates on the premise that only a few single nucleotide polymorphisms (SNPs) have non-zero effects on the trait. Genetic variants are first clumped based on LD and then filtered based on their p-values to derive polygenic scores (Kim et al., 2023; Mak et al., 2017).

Breast cancer polygenic risk scores for non-European populations, underrepresented in genetics studies, were developed using the C+T method (Ho et al., 2022). It was also used to compare different polygenic profiling methods for Alzheimer’s disease risk (Leonenko et al., 2021). This method was utilized to assess the race-specific susceptibility of SNPs to AS in the Taiwanese population, as well as to examine the connection between SNPs associated and HLA-B27 with ankylosing spondylitis (AS) susceptibility (Ko et al., 2022). A meta-analysis used this method to determine the influence of PRS on the risk of coronary artery disease (Agbaedeng et al., 2021). A PRS for autoimmune Addison’s disease was constructed and evaluated using the C+T method (Aranda-Guillé et al., 2023). Using the UK Biobank dataset, polygenic risk scores for elevated intraocular pressure, a risk factor for glaucoma, were constructed with the C+T method as described in Gao, Huang & Kim (2019).

SCT

The Stacked Clumping and Thresholding (SCT) is an extension of C+T that allows more flexibility in selecting SNPs based on four criteria: p-value threshold, LD window size, LD correlation threshold, and imputation accuracy (Privé et al., 2019). SCT generates PRSs for different settings of these criteria and then selects the optimal ones using a penalized regression approach on the validation data (Privé et al., 2019).

LASSO

The Least Absolute Shrinkage and Selection Operator (LASSO) is a method used in regression analysis that performs both variable selection and regularization. The goal of LASSO is to achieve the smallest possible sum of squared errors, with the condition that the total absolute value of the coefficients does not exceed a predetermined threshold (Ribbing et al., 2007). It is used in machine learning and statistics to select variables in a model by shrinking some of the coefficients to zero. This method helps to prevent overfitting by decreasing the number of variables included in the model. LASSO can be used for PRS development. In this context, LASSO is used as a variable selection technique to select the most important genetic variants for inclusion in the PRSs. Addressing the insufficient representation of non-European communities in genetics studies to develop breast cancer polygenic risk scores utilizing the LASSO method as indicated in Ho et al. (2022).

Lassosum

Similar to LASSO, and it applies a LASSO penalty to nullify the effect sizes of genetic variants. It further prunes genetic variants exhibiting linkage disequilibrium and applies a threshold to the remaining variants based on their p-values (Mak et al., 2017). The Lassosum approach was reported in several PRS studies. An assessment was made on the race-specific susceptibility of single nucleotide polymorphisms to ankylosing spondylitis (AS) in the Taiwanese population. The association between human leukocyte antigen (HLA)-B27 and susceptibility SNPs for AS in Taiwan was explored. Polygenic risk scores were used to analyze genetic variations in predicting the development of AS using LassoSum as indicated in Ko et al. (2022). A meta-analysis investigated how PRSs affect the likelihood of developing coronary artery disease by employing the Lassosum method (Agbaedeng et al., 2021). A reference-standardized framework assessed the predictive value of several polygenic risk score methodologies, including Lassosum (Pain et al., 2021).

SBLUP

The Super Genomic Best Linear Unbiased Prediction (SBLUP) technique adjusts SNP effect magnitudes by utilizing an external LD reference panel. This process transforms the ordinary least squares estimates of SNP into nearly optimal linear unbiased predictions (Ren et al., 2021). The SBLUP utilizes a Bayesian framework to calculate the magnitude of genetic variants effects. This method presumes that the distribution of these effects is normal, centering around zero, with their variance inversely proportional to the number of variants in the score. The SBLUP technique provides greater precision in the construction of PRS compared to other methods (Slunecka et al., 2021; Robinson et al., 2017).

Bayesian methods

Machine learning methods

The computation of PRS often relies on simple linear models which might not fully encompass the intricate interdependencies involved between phenotypes and genotypes (DeWan, 2018; Aschard, 2016). Therefore, machine learning methods that can account for non-linearities and interactions among genetic variants are of interest for improving the accuracy and interpretability of PRS.

A proposed approach combines an ensemble method for selecting SNPs with Gradient-Boosted Trees (GBT) to account for the non-linear and interactive influences of SNPs on phenotypes. When a PRS is included as a feature within an extreme gradient boosting model, there is a notable enhancement in the explained variance percentage relative to the conventional linear PRS model, as observed across nine complex phenotypes within a diverse ancestral group from the UK Biobank (Elgart et al., 2022).

The utilization of machine learning and deep learning techniques was thoroughly explored to compute polygenic risk scores from GWAS data. Random forest (RF) and support vector machines (SVM), and deep learning methods were employed to calculate weight vectors, which play a pivotal role in PRS computation (Öztornaci et al., 2023). Additionally, variable importance measurements obtained from the RF method serve as weight vectors. In all these methods, individual risk scores are derived by multiplying each SNP with its corresponding weight vector.

Peng et al. (2024) introduced a DL framework that captures intricate genetic interactions beyond additive effects. In contrast to traditional PRS models, which often assume linear relationships, DeepRisk leverages neural networks to model non-linear associations among single-nucleotide polymorphisms (SNPs). The approach described in Peng et al. (2024) has demonstrated superior performance in predicting disease risk, particularly in scenarios involving complex genetic architectures. The study (Zhou et al., 2023) introduced a neural network model that captures non-linear interactions among genetic variants, offering enhanced predictive accuracy and deeper insights into disease mechanisms. The PRS-Net, as described in Li et al. (2024), demonstrated that incorporating a lightweight geometric layer into gene-level PRSs yields reproducible and biologically interpretable improvements over both linear and black-box non-linear baselines, particularly for immune-mediated diseases and heterogeneous ancestries.

Threshold-based methods

These methods use a predefined threshold to identify variants by utilizing their p-values or the magnitude of their effects derived from GWAS summary data. Research has suggested that threshold-based methods are simple and relatively easy to implement, and widely used in practice (Privé et al., 2019; Lewis & Vassos, 2020). They can provide a straight forward interpretation of the PRS as a weighted sum of selected variants. However, research has shown that threshold-based methods may ignore variants with small effects that can collectively contribute to the PRS, they are sensitive to the choice of threshold, which can affect the predictive performance and the number of variants included (Lewis & Vassos, 2020).

Penalized regression methods

These methods use a regularization term to penalize the coefficients of the PRS, which can handle high-dimensional data effectively, they have been shown to reduce overfitting by shrinking coefficients, especially for correlated or weakly associated variants (Privé, Aschard & Blum, 2019). They also provide variable selection by shrinking less informative variables towards zero, which can improve interpretability and reduce collinearity. However, the choice of penalty parameters can be subjective and depend on the data and the trait (Pattee & Pan, 2020). The interpretation of the resulting coefficients may be challenging, especially when using non-linear penalties or complex models, they also assume a linear relationship between predictors and outcomes, which may not hold for some traits or diseases (Pattee & Pan, 2020).

Bayesian techniques

These methods use a probabilistic framework to estimate the posterior distribution of the PRS given the prior information and the data. They can incorporate various sources of information such as functional annotations and biological pathways. Also they account for uncertainty and provide credible intervals for the PRS (Ge et al., 2019; Zhou, Qie & Zhao, 2023). They have the flexibility to incorporate prior knowledge and choose the prior distribution (Ge et al., 2019). However, they can be computationally intensive, they may also require expertise in Bayesian statistics for proper implementation and interpretation of results (Song et al., 2020).

Machine learning methods

These methods use various algorithms to learn the optimal PRS from the data, such as random forests, support vector machines, neural networks, etc (Öztornaci et al., 2023). Studies have shown that they can capture complex and nonlinear relationships between variants and outcomes, which linear models may not capture, they optimize predictive performance by using cross-validation, grid search, or other techniques to tune the hyperparameters, and have flexibility in feature engineering, such as using interactions, transformations, or embeddings of variants (Squires, Weedon & Oram, 2023; Mamani, 2020). However, they may overfit the data if not properly regularized, which can reduce generalizability and robustness. Also, they can be challenging to interpret, particularly in the case of intricate models such as neural networks, which are often perceived as black boxes. They have been shown to have high computational complexity, which can limit their scalability and applicability (Mamani, 2020).

Increasing diversity and representation of data

PRS methods predominantly rely on data from individuals of European ancestry, leading to limitations in their applicability and generalizability to other ethnic groups. This lack of diversity can result in biased risk predictions and may exacerbate existing health disparities. There is a growing recognition of the importance of including diverse and representative genetic data from different populations. Initiatives are underway to collect and analyze genomic information from historically underrepresented populations, such as the Human Heredity and Health Africa (H3Africa) initiative. Developing methods that account for genetic diversity and mixed ancestry within varied groups is essential for improving the precision and reliability of PRS across all ethnic backgrounds (Zhang et al., 2023; Lam et al., 2019).

To enhance multi-ancestry prediction models, it is necessary to leverage genetic data from diverse populations to improve PRS performance. For example, the PROSPER method demonstrates improved precision and reliability showing a 70% increase in accuracy for individuals of African ancestry compared to traditional models (Zhang et al., 2024). Such approaches help correct biases inherent in European-centric models by accounting for differences in genetic architecture, including linkage disequilibrium patterns and allele frequencies (Cavazos & Witte, 2021). Incorporating ancestry-specific data into PRS algorithms is essential for improving their applicability across diverse populations (Lerga-Jaso et al., 2024).

A major driver of improved accuracy and generalizability in PRS development is the increasing availability of large-scale biobank databases. Due to the complexity of the human genome, large datasets are critical for identifying associations between genetic variants and complex traits (Raben et al., 2023). Biobank resources support the development, validation, and application of PRS by providing extensive training data and, crucially, multi-ancestry samples for cross-population evaluation (Tsuo et al., 2024; Thompson et al., 2022). Major global initiatives including the UK Biobank, All of Us (AoU), China Kadoorie Biobank, Biobank Japan, deCODE Genetics, the Estonian Biobank, and Lifelines in the Netherlands are helping to address the historical underrepresentation of non-European populations in genetic research (Ju et al., 2022). These resources provide the statistical power needed to reduce false positives, identify novel variants, and refine estimates of single nucleotide polymorphism (SNP) effect sizes (Raben et al., 2023; Ju et al., 2022). With advancements in analytic tools and machine learning algorithms, biobank databases are making PRS construction increasingly accessible to a broader range of researchers (Sakaue et al., 2020; Du et al., 2023).

Using explainable AI for improving the interpretability of PRS

ML-based methods can be used to learn the optimal PRS from the data. However, the lack of transparency of ML’s prediction could lead to a poor generalization on datasets when a model learns to predict on irrelevant features. The explainability of MLs is crucial in healthcare as the consequences of a wrong prediction in diagnostics may cause life-changing decisions for a patient (Elgart et al., 2022). Recently, explainable artificial intelligence (XAI) has been widely used in the literature to overcome the lack of insight of ML-based models in healthcare and medical diagnosis systems (Zhang, Weng & Lund, 2022). XAI reveals the decision patterns of ML-based models, which helps medical practitioners understand the logical reasoning for the model’s prediction (Zhang, Weng & Lund, 2022). XAI is a promising research direction that requires more attention from the PRS research community.

Incorporating environmental and lifestyle factors

PRS methods typically focus solely on genetic components and may not fully capture the multifactorial nature of complex traits and diseases. However, lifestyle and environmental factors—such as exercise, diet, and exposure to toxins—play significant roles in disease risk but are often overlooked in traditional PRS analyses. Integrating these non-genetic variables into PRS models can enhance their predictive power and relevance. For instance, including data on smoking behavior, socioeconomic status, and geographic location can improve risk stratification and facilitate more personalized interventions. Advanced statistical approaches, such as gene-environment interaction modeling, are being developed to better capture the interplay between genetic and environmental influences on health outcomes (Wang et al., 2021; Koch et al., 2023).

Recent advances demonstrate the potential of integrating such factors. For example, incorporating 109 exposome variables—including tobacco use, education, and others—into cardiovascular disease risk prediction using machine learning increased the area under the curve (AUC) to 0.82 (Shahbazi & Nowaczyk, 2025). Additionally, the use of Internet of Things (IoT) devices enables real-time integration of lifestyle and environmental data such as diet and air quality, supporting adaptive health interventions for hyper-personalized medicine (Tan et al., 2025). Moreover, social determinants of health (SDoH) have shown significant associations with disease outcomes, particularly in high-risk environments, further highlighting the importance of integrating socio-environmental context into PRS applications (Guare et al., 2024).

Evaluating clinical and public health implications

PRSs show great potential for predicting health risks, but their implementation also raises important ethical, societal, and legal concerns. The integration of PRS into clinical practice requires a thoughtful approach to issues such as informed consent, data privacy, protection, and equitable access to genetic services and interventions. It is essential to rigorously evaluate both the effectiveness and broader implications of PRS across diverse populations and clinical settings. Studies assessing the utility of PRS in guiding preventive strategies, screening programs, and therapeutic decisions are vital for shaping evidence-based healthcare policies.

Public understanding and health literacy around genetic risk are equally important. Enhancing awareness can help ensure informed decision making and reduce the risk of misinterpretation or misuse of genetic information in clinical and personal contexts.

Ethical considerations must be addressed, particularly when individuals are assigned high-risk scores, which may lead to psychological distress or stigmatization. In addition, disparities in access to genetic testing, risk assessment tools, and personalized interventions can exacerbate existing health inequalities (Andreoli et al., 2024). Clear communication of genetic findings and proper implementation supported by robust informed consent processes are essential for integrating PRS into personalized medicine and improving shared decision-making between patients and healthcare providers (King & Bishop, 2017).

In summary, addressing these challenges and leveraging the opportunities presented by advancements in genomics, data science, and healthcare delivery will be essential for realizing the complete promise of PRS in enhancing health outcomes and minimizing health inequities among different groups (Koch et al., 2023; Simona et al., 2023; Lewis & Vassos, 2017). The criticality of generating precise PRS for various complex traits and illnesses is paramount. PRS offers a gauge of an individual’s genetic susceptibility to a complex trait or illness, indicating the probability of manifesting a specific trait or illness grounded on one’s genetic makeup. PRS analysis aims to pinpoint individuals at heightened disease risk by analyzing genetic variations alongside clinical factors. Thus, the more precise the PRS, the more effectively we can pinpoint disease risks and devise preventative measures.