to the more private alleles that drive some rarer subtypes of diabetes [11,12]. With the notable exception of the HLA region (which has the major impact on T1D risk), most of these common variants have only modest effects on individual predisposition: the biggest effects for T2D modulate risk by no more than 40% per allele and most have much smaller effects [9,10]. However, in combination, the impact of this variation can be more profound [9,13]. In the most recent GWAS for T2D, the entire set of associated variants so far detected explains around 20% of the overall variation in disease risk [9], in Europeans at least (comparable analyses in non-European populations are limited by the sample sizes available for study). Estimates of the heritability of T2D vary widely [14,15] around a median of 40%, suggesting that around half the genetic contribution to the variation in risk can be quantified for each individual. Estimates of the heritability of T1D are higher [16] and a greater proportion of that genetic risk can be captured using existing approaches. Ongoing efforts to further characterize the genetic basis of both major subtypes of diabetes – through detecting significant associations at variants that have escaped detection because they are too rare, or have small effects – will increase the proportion of individual genetic predisposition that can be directly measured. The steadily expanding list of genetic variants delivered by these successive waves of genetic discovery has delivered novel mechanistic insights into disease pathophysiology. ADVANCE ARTICLE: Endocrine Reviews Downloaded from https://academic.oup.com/edrv/advance-article-abstract/doi/10.1210/er.2019-00088/5535575 by 81225740 user on 24 July 2019 ADVANCE ARTICLE Endocrine Reviews; Copyright 2019 DOI: 10.1210/er.2019-00088 4 Some of these have led to an understanding of the major processes contributing to disease risk, such as the role of islet-specific as well as immunological processes with respect to T1D risk [17,18] or the relative impact of defects in insulin secretion and action for T2D [19]. Other studies have attempted to dissect the detailed molecular, genomic and physiological events that mediate risk at individual loci [9,20-23]. These efforts can have direct translational impact, for example through the identification of novel therapeutic targets, or biomarkers that track disease progression. In this review, however, we focus on a different route from human genetics to translation, one that derives estimates of an individual’s predisposition to diabetes and its subtypes (in the form of polygenic scores) from the patterns of individual genetic variation at sites known to influence diabetes predisposition. 2. The concept of polygenic scores The idea of grouping genetic variants to capture the aggregate genetic risk for a given disease is not new. An early promise of genetic discovery in complex (polygenic) conditions was to predict clinical outcomes. It was recognized that, in contrast to classical Mendelian diseases, where the presence of a specific mutation was deterministic and typically heralded the eventual onset of disease (contingent on penetrance), the genetic risk for complex, multifactorial diseases is probabilistic and most appropriately used as a predictor that quantified a discrete increment in overall risk [24]. This is because for complex human traits, the overwhelming majority of associated genetic variants exert modest effects, and the ability of any individual variant to influence clinical outcomes is small. The obvious approach is to sum the effects of risk alleles associated with a given condition, to generate an aggregate estimate of genetic risk. This approach was justified by the observation that early genetic associations seemed to work in an additive fashion, with little or no evidence of epistasis. This concept was pioneered for age-related macular degeneration, the first disease for which GWAS proved successful [25] and had also been employed in T2D for the three reproducible genetic associations that had emerged from the pre-GWAS era [26]. This concept can be easily expanded from the disease arena to quantitative traits. Here, rather than expressing “risk” (which connotes the deleterious burden of illness), the aim is to capture the overall variance in a trait conferred by the set of genetic variants grouped into a composite score. Examples include circulating levels of a specific metabolite or the inherited predisposition toward a behavioral pattern, where the deleterious connotations ascribed to the term “risk” no longer apply. In this review, therefore, we favor the use of the term “polygenic score” as a more inclusive general descriptor. The initial uses of polygenic scores deliberately focused on the inclusion of individual genetic variants for which the evidence for association was robust. This occurred as a “route correction” to the historical trend whereby the proliferation of candidate gene studies and the adoption of liberal statistical significance thresholds had led to the publication of many genetic associations which later proved irreproducible, and likely represented false positive findings [27]. In the GWAS era, such high-likelihood variants had to achieve genome-wide significance, based on a widely-accepted threshold of p14.4 had 11.0% risk of developing multiple