Table 1.

Assessing the clinical utility of polygenic risk scores

A: Population level

The predictive ability of polygenic risk scores can be measured in research studies, where differences between cases and controls (Fig. 1) or of a continuous trait in a population are assessed. Here, the disease status or trait is pre-established, and the studies measure the extent to which this is determined by the PRS. Outcome measures from such studies include:   (1) R2 from linear regression, which quantifies the proportion of variance in a continuous trait captured by the PRS, or equivalently Nagelkerke’s R2 for logistic regression for case-control disease status.

(2) R2 on a liability scale, which transforms Nagelkerke’s R2 to reflect disease prevalence, instead of the case-control ratio of the research study [15].

(3) The area under the receiver operating characteristic curve (AUC) [16], which takes a value from 0.5 to 1. This gives an overall summary of the predictive ability of the model. It is most easily interpreted as the probability that a randomly selected case will have a higher polygenic risk score than a randomly selected control. Such models can also include risk factors such as age and sex, which will increase the AUC values above that based on PRS alone.

(4) The proportion of the population that has a k-fold increased odds (k = 2, 3, …), compared to the population disease risk.

(5) Odds ratio of disease risk conferred by a 1-standard deviation increase in PRS.

(6) Odds ratio of disease for an individual in the top PRS decile (or other quantiles) compared to individuals in a different part of the PRS distribution. The high-risk group may be compared to the lowest decile, a mid-quintile (e.g. 40–60%), or those outside the high-risk group (0–90%). Comparing the upper and lower tails maximises the odds ratio for impact but raises concerns about the arbitrariness of the quantile used.

B: Individual level

In a clinical setting, the focus is on a single person: what information does their PRS give about their risk of disease? Possible outcome measures that are relevant at an individual level include:

(a) At what percentile in the distribution of PRS does this individual lie? This is between 0 and 100%, with scores having a normal distribution.

(b) What is this person’s relative risk of disease compared to the average risk in the population?

(c) What is this person’s absolute risk of disease, and by what age [17]?