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

Similarities

 • Both MR and PRS exploit results from GWAS (summary statistics)

 • Both can be performed using individual- or summary-level data, however most applications of PRS apart from estimating shared genetic aetiology requires individual level data

 • Both MR and PRS can be used to estimate an effect of liability to an exposure on an outcome

 • PRSs can be utilised within in a one-sample MR framework

 • Provided heterogeneity is low, and the PRS is scaled to the exposure, MR and PRS should give approximately the same answer (see examples of PRS and MR studies of BMI-sleep duration below)

 • Both MR and PRS rely on the R² (variance explained) as a metric of total strength of the instrument [10]

Differences

 • PRS and one-sample MR combine all SNPs into a score, whereas summary level MR is done on a per SNP basis and meta-analysed

 • Methods for examining and correcting for bias due to horizontal pleiotropy are better developed for MR than for PRS. It is not possible to formally detect and correct for pleiotropy using PRS, but for some PRS applications, horizontal pleiotropy does not cause bias (Table 1)

 • For a given sample size, PRS have greater power than MR; thus PRS are often useful for smaller samples. However, summary level MR usually have much larger sample sizes as a result of using large GWAS

 • PRS are generally more flexible in their applications than MR (Table 1)

 • Less likely to suffer weak instrument bias [11] with a PRS as alleles aggregated into a score and thus usually explains more variance in the exposure. This relates to average strength of the instrument, which is estimated using the F-statistic (should be > 10 for an instrument of good average strength) [11]