A brief comparison of polygenic risk scores and Mendelian randomisation

Abstract

Mendelian randomisation and polygenic risk score analysis have become increasingly popular in the last decade due to the advent of large-scale genome-wide association studies. Each approach has valuable applications, some of which are overlapping, yet there are important differences which we describe here.

Main text

There is often misconception surrounding the differences and similarities between polygenic risk score analysis and Mendelian randomisation, and when one method should be applied over the other. In this article we briefly describe what polygenic risk scores (PRS) and Mendelian randomisation are, their respective strengths and limitations, whether PRS and Mendelian randomisation (MR) are equivalent, and as such, whether we can use these methods interchangeably.

What are polygenic risk scores (PRS)?

Polygenic risk scores (sometimes also referred to as genetic risk scores) estimate an individual’s genetic predisposition to a trait (e.g., LDL-cholesterol) or disease (e.g., type-2 diabetes) [1]. A PRS is usually calculated using individual-level genotypes and data from genome-wide association studies (GWAS). An unweighted PRS simply reflects the sum of an individual’s risk alleles. Unweighted PRS do not take into account the relative magnitude of effect of each genetic variant on the trait of interest. Weighted PRSs are the sum of an individual’s risk alleles, weighted by the effect sizesreported in published GWAS (e.g., log(beta) or beta coefficient). The number of variants to include in a PRS depends on the intended application, whether it is to assess causality or prediction, as more variants is better for the latter, but this increases the chances of including pleiotropic variants. Table 1 outlines some of the potential applications of PRS. For more extensive details on PRS methods, see [1, 2].

Table 1 A non-exhaustive list of potential applications of PRS and MR

What is Mendelian randomisation?

Mendelian randomisation uses SNPs (i.e., single nucleotide polymorphisms – SNPS –) or common genetic variants as instrumental variables (IVs) for an exposure of interest, rather than using the observed phenotype, to examine whether the exposure (or liability to an exposure if it is binary) has an effect on an outcome of interest [4]. MR exploits the unique properties of common genetic variants and the fact that genes are randomly allocated from parents to offspring during gamete formation [5]. As such, MR exploits Mendel’s laws of ‘Independent Assortment’ and ‘Segregation’. In practice, an MR has three assumptions that need to be upheld for it to be valid: 1) robustness of association between SNPs and the exposure to be instrumented, 2) no association (horizontal pleiotropy) between the SNPs for the exposure and the outcome that does not go via the exposure (Table 1), and 3) the SNP-outcome relationship is unconfounded. MR can be performed in both individual-level and summary-level (i.e. genome-wide association study summary statistics) data settings [6], which each have different advantages and disadvantages, summarised in Lawlor et al. [6].

PRS vs. MR: understanding their similarities and differences using an applied example: body mass index (BMI) and sleep duration

The relationship between BMI and sleep duration has been extensively investigated via epidemiological and experimental studies [7]. The first PRS study which aimed to investigate shared genetic aetiology between BMI and (self-reported) sleep duration was published in 2019 [8]. Then, a comprehensive and well powered MR study of BMI and sleep duration emerged earlier this year [9], which investigated causality between this exposure and outcome. The two studies had distinct objectives and thus, employed different approaches (e.g., the PRS study employed nine different PRS with varying numbers of SNPs, whereas the MR study used a genome-wide significant 67-SNP instrument). However, both studies reached similar conclusions, and the analyses produced comparable results, such that there was little shared genetic aetiology, and no evidence of a causal relationship between BMI and self-reported sleep duration in adults. Table 2 presents a detailed account of similarities and differences between PRS and MR, while Fig. 1 is a graphical representation of the conceptual similarities and differences between the two methods.

Table 2 Similarities vs. differences between Mendelian randomisation and polygenic risk score approaches

Fig. 1

Graphical representation of the conceptual similarities and differences between PRS and MR

Conclusions

PRS and MR both have useful applications in aetiological epidemiology. PRS are useful in the case of weak genetic instruments or smaller sample sizes, as aggregation of alleles into a score increases the variance explained in the exposure, and thus increases power. MR is useful for larger sample sizes and can also be performed on publicly available summary data. MR is the preferred method for identifying and correcting for potential bias due to horizontal pleiotropy as methods are more widely developed. PRS are typically more flexible in their potential applications.