The contribution of genetic variants to disease depends on the ruler

Abstract

Our understanding of the genetic basis of disease has evolved from descriptions of overall heritability or familiality to the identification of large numbers of risk loci. One can quantify the impact of such loci on disease using a plethora of measures, which can guide future research decisions. However, different measures can attribute varying degrees of importance to a variant. In this Analysis, we consider and contrast the most commonly used measures - specifically, the heritability of disease liability, approximate heritability, sibling recurrence risk, overall genetic variance using a logarithmic relative risk scale, the area under the receiver-operating curve for risk prediction and the population attributable fraction - and give guidelines for their use that should be explicitly considered when assessing the contribution of genetic variants to disease.

Figure 1. Different measures of genetic effects on disease

A number of different measures can be used to assess how much known genetic factors contribute to the overall genetic variation in disease. These include: a. heritability, b. sibling relative risk, c. log relative risk genetic variance, d. area under the receiver operating curve (AUC), and e. population attributable fraction. These measures have their bases in traditionally distinct disciplines such as quantitative genetics and epidemiology, which have recently begun to coalesce. While the latter were originally developed to address different questions, they are presently being repurposed to assess how much genetic variation cab be explained. We compare these measures via simulation and applications.

Figure 2. Empirical evaluation of measures of genetic effects

Comparison of heritability, approximate heritability, sibling relative risk, log relative risk genetic variance, and area under the curve (AUC) explained across a range of complex disease architectures. The measures are calculated for a single causal variant with risk allele frequency (RAF) = 0.01, 0.10, 0.25, 0.50, 0.75, and 0.99 and genetic relative risk (RR) ranging from 1.0 to 3.0 (assuming multiplicative model). The overall disease risk is assumed = 0.01, and the total sibling relative risk = 5, which gives an overall genetic heritability on the liability scale = 0.55 and a maximum AUC = 0.95. The percentage of heritability, sibling risk, and logRR genetic variance explained is quite modest for low RRs and small RAF, but as these increase the measures start to materially differ. Heritability is always one of the smallest measures, and is overestimated by the approximate heritability as the RR increases. The sibling relative risk and AUC are generally the largest measures for lower RAFs.

Figure 3. Application of measures to four diseases

Comparison of commonly used measures for assessing the impact of known risk variants on four diseases: a. breast cancer (65 variants), b. Crohn’s disease (143 variants), c. rheumatoid arthritis (36 variants), and d. schizophrenia (32 variants). The measures are: heritability explained; approximation of heritability explained; sibling recurrence risk explained; logRR genetic variance explained; and the proportion of area under the curve (pAUC). Each line corresponds to an individual risk variant, indicating the percentage of each measure (e.g., total variability) it explains. Lines are different colors depending on the relative risk (estimated by the odds ratio, OR) for each variant. The percentage axes are on a squared scale.

Figure 4. Aspects of disease heritability: known, hiding, and missing

A growing proportion of the total heritability estimated from family studies can be explained by known variants detected in existing genome-wide association studies (bottom). This is one of the key measures considered here. The remaining heritability can be broken into that which is ‘hiding’ versus ‘still missing’. The hiding heritability can be estimated from genome-wide arrays using the Genetic Relatedness Estimation through Maximum Likelihood (GREML) model. The still-missing heritability is that which may remain even after genome-wide association studies, reflecting for example genetic different architectures (e.g., rare variants). Note that the total heritability may be biased upward due to confounding by non-additive genetic or non-genetic factors.