A systematic framework for assessing the clinical impact of polygenic risk scores

Abstract

The estimate of an individual's genetic susceptibility to a disease can provide critical information when setting screening schedules, prescribing medication and making lifestyle change recommendations. The polygenic risk score is the predominant susceptibility metric, with many methods available to describe its construction. However, these methods have never been comprehensively compared or the predictive value of their outputs systematically assessed, leaving the clinical utility of polygenic risk scores uncertain. This study aims to resolve this uncertainty by deeply comparing the maximum possible, currently available, 15 polygenic risk scoring methods to 25 well-powered, UK Biobank derived, disease phenotypes. Our results show that simpler methods, which employ heuristics, bested complex, methods, which predominately model linkage disequilibrium. Accuracy was assessed with AUC improvement, the difference in area under the receiver operating curve generated by two logistic regression models, both of which share the covariates of age, sex, and principal components, while the second model also contains the polygenic risk score. To better determine the maximal utility of polygenic risk scores, straightforward score ensembles, which bested all methods across all traits in the training data-set, were evaluated in the withheld data-set. The score ensembles revealed that the accuracy gained by considering a polygenic risk score varied greatly, with AUC improvement greater than 0.05 for 9 traits. Many additional analyses revealed widespread pleiotropy across scores, large variations between assessment statistics, peculiar patterns amongst phenotype definitions, and wide ranges in the optimal number of variants used for scoring. If these many variable aspects of score creation can be well controlled and documented, simple methods can easily generate polygenic risk score that well predict an individual's future liability of certain diseases.

Global biobank analyses provide lessons for developing polygenic risk scores across diverse cohorts

“…The ORs between the top 10% and bottom 10% were more heterogeneous between biobanks and also higher relative to other comparisons (e.g., top 10% vs middle and other strata). This is consistent with previous studies where OR reported between tails of the PRS distribution is generally inflated relative to those between top ranked PRS and general populations 11 . We measured the variation of OR between biobanks using the coefficient of variation of OR (CoeffVar OR , see STAR Methods).…”

Polygenic Scores for Height in Admixed Populations

“…LDpred's infinitesimal model ( Figure S7). More sophisticated approaches provide limited improvement in predictive power, require additional assumptions about LD structure or other parameters, and do not necessarily lead to substantial improvements in predictive power or transferability (Kulm et al 2020).…”

Predicting Multiple Sclerosis: Challenges and Opportunities

“…The earliest effort to predict MS using environmental and genetic data was published in 2009 ( 29 ). Since then, there have been several efforts incorporating increasingly refined genetic maps of MS susceptibility and applying this approach to novel datasets ( Table 1 ) ( 29 – 36 , 38 , 39 ). Broadly, these studies support the view that genetic risk scores (GRS) / PRS can discriminate between cases and controls.…”