Genomic Prediction of 16 Complex Disease Risks Including Heart Attack, Diabetes, Breast and Prostate Cancer

Abstract

We construct risk predictors using polygenic scores (PGS) computed from common Single Nucleotide Polymorphisms (SNPs) for a number of complex disease conditions, using L1-penalized regression (also known as LASSO) on case-control data from UK Biobank. Among the disease conditions studied are Hypothyroidism, (Resistant) Hypertension, Type 1 and 2 Diabetes, Breast Cancer, Prostate Cancer, Testicular Cancer, Gallstones, Glaucoma, Gout, Atrial Fibrillation, High Cholesterol, Asthma, Basal Cell Carcinoma, Malignant Melanoma, and Heart Attack. We obtain values for the area under the receiver operating characteristic curves (AUC) in the range ~0.58-0.71 using SNP data alone. Substantially higher predictor AUCs are obtained when incorporating additional variables such as age and sex. Some SNP predictors alone are sufficient to identify outliers (e.g., in the 99th percentile of polygenic score, or PGS) with 3-8 times higher risk than typical individuals. We validate predictors out-of-sample using the eMERGE dataset, and also with different ancestry subgroups within the UK Biobank population. Our results indicate that substantial improvements in predictive power are attainable using training sets with larger case populations. We anticipate rapid improvement in genomic prediction as more casecontrol data become available for analysis. such as age and sex. Some SNP predictors alone are sufficient to identify outliers (e.g., in the 99th percentile of polygenic score, or PGS) with, e.g., 3-8 times higher risk than typical individuals. We validate predictors out-of-sample using the eMERGE dataset 16 (taken from the US population), and also with different ancestry subgroups within the UK Biobank population as done in 17 . Note that the disease conditions contain a mix of self reported and diagnosed conditions, described in Supplemental Section B, but we do not see any distinguishable difference in the results.Our analysis indicates that substantial improvements in predictive power are attainable using training sets with larger case populations. We anticipate rapid improvement in genomic prediction as more case-control data become available for analysis.It seems likely that genomic prediction of disease risk will, for a number of important disease conditions, soon be good enough to be applied broadly in a clinical setting 18-21 . Inexpensive genotyping (e.g., roughly $50 per sample for an array genotype which directly measures roughly a million SNPs, and allows imputation of millions more) can identify individuals who are outliers in risk score, and hence are candidates for additional diagnostic testing, close observation, or preventative intervention (e.g., behavior modification).We note the successful application of similar methods in genomic prediction of plant and animal phenotypes. Earlier studies have shown some success on complex human disease risk using much smaller datasets and a variety of methods [22][23][24] . Early work in this direction can be found in, for example 25 , (which highlights the utility of what were the...

Within-Family Validation of Polygenic Risk Scores and Complex Trait Prediction

“…As shown in earlier work [13,14], we expect the predictors to improve substantially as more data become available for training. This is conditioned on genotyping that captures a sufficient part of the predictive regions of the genome [15].…”

“…The main dataset we use for training is the 2018 release of the UK Biobank [16,17]. In previous work, predictors were trained exclusively on genetically British individuals (as identified by principal component analysis [18]), however it has been shown that predictors trained on populations filtered by self-reported ethnicity perform equivalently [14]. In this work, all predictors are trained on the set of self-reported white individuals excluding all individuals who belong to a sibling pair.…”

“…We construct linear models of genetic predisposition for a variety of disease conditions that were presented in [14] and linear models of several quantitative human phenotypes, some of which can be found in [13]. The disease condition phenotype data describes a binary case-control status which is defined either by self-report or from a clinical diagnosis.…”

“…This process is repeated 5 times to generate a set of 5 predictors so that statistical fluctuations associated with the training process (mean and variance) can be estimated. We do not report the performance metrics on the validation sets as they are quantitatively similar to that of the final test set -see [14] for an example of this.…”

“…For all traits, we make use of L1 penalized regression as described in [13,14]. Previous work has shown this to be an effective method of generating polygenic predictors [13,14].…”