Estimating population level disease prevalence using genetic risk scores

Abstract

Clinical classification is essential for estimating disease prevalence in a population but is difficult, often requiring complex investigations. The widespread availability of population level genetic data makes novel genetic stratification techniques a highly attractive alternative. We propose a generalizable mathematical framework for determining the prevalence of a disease within a population using genetic risk scores. We compare and evaluate methods based on the means of the genetic risk scores distributions, the Earth Movers Distance between distributions, a linear combination of kernel density estimates of distributions, and an Excess method. We assess the impact on estimates resulting from the population size and proportion of cases to non-cases. Using less discriminative genetic risk scores still results in robust estimates of proportion. Genetic stratification techniques provide exciting research tools enabling unbiased insights into disease prevalence and clinical characteristics unhampered by clinical classification criteria.

The absence of islet autoantibodies in clinically diagnosed older-adult onset type 1 diabetes suggests an alternative pathology, advocating for routine testing in this age group

“…We next sought to determine the proportion of islet autoantibody negative participants that had genetically consistent type 1 diabetes in the different age groups. Using the method described above and in our previous paper (37), we found that genetically consistent type 1 diabetes in 93% (n=56/60, 95% CI 84-98%) of autoantibody negative childhood cases, 55% (37/67, 95% CI 43-67%) of young-adult cases and only 23% of older-onset autoantibody negative cases (34/151, 95% CI 16-30%). This suggested that in autoantibody negative cases of clinically defined type 1 diabetes, 77% (95% CI 70-84%) of older-adults, 45% (95% CI 33-57%) of young-adults and 7% (95% CI 0-16%) of children had non-type 1 diabetes (misclassified as clinical type 1 diabetes).…”

“…We next sought to determine the proportion of islet autoantibody negative participants that had genetically consistent type 1 diabetes in the different age groups. Using the method described above and in our previous paper (37), we found that genetically consistent type 1 diabetes in 93% (2/151, 95% CI 0-5%) of older-adults with monogenic diabetes (Supplementary Figure 2). This data suggests that in children all misclassified patients had monogenic diabetes while in adults the majority had type 2 diabetes.…”

IgA Nephropathy Genetic Risk Score to Estimate the Prevalence of IgA Nephropathy in UK Biobank

“…We performed an analysis of IgAN-GRS using currently available published data; it is likely that as case-control GWASs increase in size and with better genome-wide coverage, newer risk loci and possibly newer approaches to GRS generation will be identified that could improve discriminative power of an IgAN-GRS and the precision around estimates of IgA disease within UKBB and other similar datasets. 21 Our sensitivity analysis suggested that the IgAN-GRS could be driven by the 4 HLA SNPs in high LD, the remaining 10 SNPs are not shown to be associated with hematuria. This may be due to the modest power of our sample size of cases, and the modest effect size of the non-HLA variants associated with IgAN.…”

“…We performed an analysis of IgAN-GRS using currently available published data; it is likely that as case-control GWASs increase in size and with better genome-wide coverage, newer risk loci and possibly newer approaches to GRS generation will be identified that could improve discriminative power of an IgAN-GRS and the precision around estimates of IgA disease within UKBB and other similar datasets. 21 …”

“…The mean IgAN-GRS of artificial mixtures containing 10%, 25%, and 50% of IgAN cases was evaluated as a phenotype GRS in the preceding formula ( Figure 1 ). For each proportion, 1000 randomly generated mixtures were made, allowing the mean and 95% CIs around estimates 21 ( Supplementary Figure S2 ). To calculate CIs using reference data, we bootstrapped around the point estimates generating 10,000 mixtures of equivalent size to the phenotype of interest.…”