Introduction

1 The science behind 23andMe's Type 2 Diabetes report White Paper 23-19 The science behind 23andMe's Type 2 Diabetes report Estimating the likelihood of developing type 2 diabetes with polygenic models Authors: Michael L. Multhaup*, Ryo Kita*, Becca Krock, Nicholas Eriksson, Pierre Fontanillas, Stella Aslibekyan, Liana Del Gobbo, Janie F. Shelton, Ruth I. Tennen, Alisa Lehman, Nicholas A. Furlotte, and Bertram L. Koelsch *these authors contributed equally to this work Updated April 2019; see the "Change Log" section for more details.

2 Introduction In 2017, the Centers for Disease Control (CDC) estimated that over 20% of type 2 diabetes (T2D) cases are undiagnosed, representing more than 7 million residents (CDC National Diabetes Statistics Report 2017, Dall et al. 2014). By 2050, the number of undiagnosed cases could be over 13 million, as T2D prevalence is projected to increase to 25-28% of the population (Boyle et al. 2010). The CDC estimates that more than 80 million current residents have prediabetes, but only 11% of them have been diagnosed by a healthcare professional (CDC National Diabetes Statistics Report 2017).

3 More than 70% of individuals with prediabetes will eventually develop T2D (Tab k et al. 2012). This high rate of progression can be mitigated. Several intervention strategies have been shown to reduce the risk of progression from prediabetes to T2D. Lifestyle interventions in people with prediabetes using either a combination of weight loss and exercise or weight loss alone lower the risk of T2D by greater than 50% (Knowler et al. 2002, Tuomilehto et al. 2001, Lean et al. 2018). Similarly, multiple therapeutics, including thiazolidinediones (DREAM et al.)

4 2006), -glucosidase inhibitors (Chiasson et al. 2002), biguanides (Metformin) (Lily et al. 2009), and 1 The science behind 23andMe's Type 2 Diabetes report GLP-1 analogs (Astrup et al. 2012), have been shown to reduce the risk of progression from prediabetes to diabetes. One in four healthcare dollars spent in the US is used to treat diabetes and its complications (American Diabetes Association 2018). Each diagnosed case of T2D costs the healthcare system more than $10,000 per year, resulting in a yearly total estimate of $240-330 billion (Dall et al.

5 2014, American Diabetes Association 2018). By contrast, the estimated annual cost per year of each case of prediabetes is estimated to be $500 (Dall et al. 2014). Therefore, early identification of individuals at high risk for T2D could improve allocation of prevention resources. More than a hundred T2D risk scores to predict clinical outcomes have been published in the medical literature, leveraging combinations of demographic variables, family history, and biomarkers as predictors with widely varying levels of accuracy (Noble et al.

6 2011). The discriminatory ability as measured by the area under the receiver operating characteristic curve (AUC) ranges from for scores that use only demographic and self-reported data (Chien et al. 2009) to over for scores that include multiple clinical biomarkers, such as triglyceride levels, HbA1c levels, and fasting plasma glucose (Rathmann et al. 2010, Guerrero-Romero et al. 2010). Genetics is a T2D risk factor that has been increasingly investigated for development of predictive models (L ll et al. 2017, Khera et al.

7 2018). The heritability of T2D has been estimated at 25-75% (Almgren et al. 2011). The predictive models based on genetics use multiple genetic markers reflecting the polygenic nature of T2D. Building on these promising findings and harnessing the unique genotypic and phenotypic scale of the 23andMe customer database, here we present a consumer-oriented T2D genetic report powered by a polygenic score (PGS) based on over 1,000 T2D-associated genetic variants. Methods Genotyping The genotyping in this study was performed as previously described (Youna et al.

8 2016). Briefly, saliva samples were used to extract DNA and genotyping was performed by the National Genetics Institute (NGI), a subsidiary of the Laboratory Corporation of America and a Clinical Laboratory Improvement Amendments (CLIA)-certified clinical laboratory. Two different Illumina BeadChip platforms were used for genotyping: the Illumina HumanOmniExpress+ BeadChip (HOEB, also known as Version 4) with a base set of 730,000 variants, augmented with 250,000 variants to obtain a superset of HumanHap550+ content as well as a custom set of 30,000 variants, and the Illumina Infinium Global Screening Array (GSA, also known as Version 5, ~640,000 variants) supplemented with ~50,000 variants of custom content.

9 Samples that failed to reach call rate were discarded. Unless otherwise specified, the methods, figures, and tables throughout the text describe the GSA platform samples, while tables specific to HOEB platform samples are found in the Supplementary Material. 2 The science behind 23andMe's Type 2 Diabetes report Due to the differences in variants assayed by each platform, we used imputed genotypes for the GWAS in this report to maximize the number of samples with data for each variant. To obtain the imputed genotypes, phasing was done using the Eagle software package (Loh et al.)

10 2016). Imputation was performed on the autosomal and X-chromosomes separately using the Minimac3 version software package (Fuchsberger et al. 2015). For the imputation panel, we combined the 1000 Genomes Phase 3 haplotypes (1000 Genomes Project Consortium et al. 2015) with the UK10K imputation reference panel (UK10K Consortium et al. 2015). For model training and evaluation of the 23andMe PGS models, we only used the variants that are genotyped in each samples respective array to stay consistent with what is available for customer reports.