across the various axes is likely to be more informative regarding disease presentation and clinical course. In accordance with the “palette” analogy, each of these processes can be considered to be represented by a particular base color (red, blue, yellow etc): for any given individual, risk along each axis would be captured by the saturation of the relevant base color, and their overall profile of T2D-predisposition visualized in terms of the mix of those colors which results when they are combined. This “palette” model is consistent with current understanding of the pathogenesis and the genetic architecture of T2D. Over the past decade, T2D-associated variants have been shown to modulate T2D risk through diverse mechanisms: some increase T2D risk through an impact on obesity (e.g. FTO), others reduce insulin sensitivity (e.g. PPARG, IRS1) whilst others compromise insulin secretion, either through direct effects on islet function (e.g. KCNJ11) or development (e.g. HNF1A), or indirectly through impact on incretin signalling (e.g. GLP1R) [82]. The various classes of T2D therapeutics operate through the same range of mechanisms to reverse the diabetic phenotype or control its glycemic consequences. The weight of evidence indicating that the genetic contribution to T2D predisposition mostly arises from common variants of limited individual effect [11,12] emphasizes the need to think in terms of a gradation of polygenic risk across individuals, rather than a classification based around rigid, discrete subtypes [3]. As well as providing a framework for capturing the mechanistic basis of T2D heterogeneity, this model also offers an approach to understanding how an individual’s particular genetic profile contributes to their progression from normal metabolic health towards the diabetic state. 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 13 In 2010, Voight et al. were first to demonstrate that patterns of genetic association across diabetes-related quantitative traits could be utilized to annotate T2D-risk loci with respect to their physiological impact, analyses which highlighted the predominant role played by variants influencing insulin secretion [19]. This approach was further developed by Dimas and colleagues [83] to perform a systematic analysis of the relationships between 36 T2Drisk alleles and a range of glycemic measures including indices of insulin secretion and insulin resistance gathered in nondiabetic individuals. Scott and colleagues extended this approach to a larger set of 93 T2D-risk alleles and included BMI and lipid measures in their clustering in addition to glycemic traits [44]. Three main patterns of multi-trait association emerged from this analysis, two of them reflecting defects in insulin secretion and insulin action respectively, and a third characterized by obesity and dyslipidemia. One major limitation of the unsupervised hierarchical “hard” clustering approach used in these papers [44,83] is that it requires each variant to be assigned to a single cluster, based on the questionable assumption that each variant can only be involved in one pathophysiological pathway. Access to an expanded range of large-scale quantitative trait association data (from largescale GWAS efforts within global consortia such as GIANT [anthropometric traits], MAGIC [continuous glycemic traits] and GLGC [lipids]) plus advancements in clustering algorithms have enabled a new wave of variant clustering analyses [20,38]. These described efforts to aggregate GWAS data from more diverse sets of T2D-related quantitative traits and employed more sophisticated “soft” clustering techniques [84,85] to pick out clusters of T2Dassociated variants with similar patterns of impact across the suite of phenotypes. These soft clustering approaches explicitly allow for the possibility that a variant influences more than one process. Mahajan et al. [20] deployed a C-means clustering approach across GWAS data from 10 T2D-related quantitative traits for a set of 94 T2D association signals that emerged from a T2D-GWAS of ~450K individuals, identifying 6 variant clusters (based on a threshold of 80% for cluster membership). Udler et al. [38] employed a complementary soft clustering approach - Bayesian nonnegative matrix factorization - to a partly overlapping set of 94 T2Drisk variants, gathering GWAS data from 47 diabetes-related traits, and identifying five clusters. Reassuringly, despite these differences, the clusters identified by both were broadly similar (Table 2). The variants within each of the genetic clusters can be used to generate “partitioned” polygenic scores that capture the genetic contribution to each intermediary process. Each of these clusters (and the pPS generated therefrom) can be assigned mechanistic labels based on the observed patterns of GWAS effects: for example, a cluster which features T2D risk alleles most clearly associated with decreased fasting insulin, can, on the basis of known pathophysiological relationships, be attributed to reduced insulin secretion. On this basis, two of the clusters were associated with an adverse impact on beta-cell function, three were characterized by insulin sensitivity