Software as a Service for the Genomic Prediction of Complex Diseases

Abstract

In the last decade the scientific community witnessed a large increase in Genome-Wide Association Study sample size, in the availability of large Biobanks and in the improvements of statistical methods to model genomes features. This have paved the way for the development of new prediction medicine tools that use genomic data to estimate disease risk. One of these tools is the Polygenic Risk Score (PRS), a metric that estimates the genetic risk of an individual to develop a disease, based on a combination of a large number of genetic variants.Using the largest prospective genotyped cohort available to date, the UK Biobank, we built a new PRS for Coronary Artery Disease (CAD) and assessed its predictive performances along with two additional PRS for Breast Cancer (BC), and Prostate Cancer (PC). When compared with previously published PRS, the newly developed PRS for CAD displayed higher AUC and positive predictive value. PRSs were able to stratify disease risks from 1.34% to 25.7% (CAD in men), from 0.26% to 8.62% (CAD in women), from 1.6% to 24.6% (BC), and from 1.4% to 24.3% (PC) in the lowest and highest percentiles, respectively. Additionally, the three PRSs were able to identify the 5% of the UK Biobank population with a relative risk for the diseases at least 3 times higher than the average. Family history is a well recognised risk factor of CAD, BC, and PC and it is currently used to identify individuals at high risk of developing the diseases. We show that individuals with

LDpred2: better, faster, stronger

“…First, we provide a sparse option in LDpred2-grid which provides models that truly encourage sparsity, in contrast to LDpred1 which outputs very small non-zero effect sizes ( Janssens and Joyner, 2019 ). In practice, the sparse version of LDpred2 can drastically reduce the number of variants used in the PGS without impacting its predictive performance, as opposed to discarding the smallest effects after having run LDpred [as tested in Bolli et al (2019) ]. The second extension is LDpred2-auto, which automatically estimates values for hyper-parameters p and h 2 , which therefore does not require any validation set to tune hyper-parameters from.…”

LDpred2: better, faster, stronger

“…First, we provide a sparse option in LDpred2-grid which provides models that truly encourage sparsity, in contrast to LDpred1 which outputs very small non-zero effect sizes (Janssens and Joyner 2019). In practice, the sparse version od LDpred2 can drastically reduce the number of variants used in the PGS without impacting its predictive performance, as opposed to discarding the smallest effects after having run LDpred (as shown in Bolli et al (2019)). The second extension is LDpred2-auto, which automatically estimates values for hyper-parameters p and h 2 , therefore it does not require a validation step to choose hyper-parameters, making it an attractive option in many applications.…”

“…When the sparse option is enabled, it provides models that truly encourage sparsity, as compared to LDpred1 which outputs very small non-zero effect sizes (Janssens and Joyner 2019). It can provide sparsity in effect sizes as much as 98% for small p, while keeping predictive performance as good as non-sparse models, as opposed to if very small effects were simply discarded (Bolli et al 2019). This sparse model performs equally well as the non-sparse model, therefore we encourage users to test it in the grid and to choose it if it performs better or equally well as the non-sparse model.…”

Polygenic risk scores: from research tools to clinical instruments

“…me/) shows users where their polygenic risk score lies against a population-specific distribution of scores. Allelica provides an online service calculating polygenic risk scores [86]. In direct-to-consumer genetic testing, MyHeritage (https://www.myheritage.com/health/genetic-risk-reports) provides polygenic risk scores on four traits, 'for people who are of mainly European ancestry'.…”