Predictive Power
Below, we show how well our polygenic scores predict traits in a completely independent “test set” of
UK Biobank participants of European ancestry who are siblings.
None of these individuals were used to develop the prediction models, and they are not closely related to anyone in the training set.
(Specifically, they are at least fourth-degree relatives from anyone in the training set).
This lets us measure how accurate our predictors are for new individuals, such as yourself.
The siblings in this test set are, on average, about 75 years old (most are in their mid 60s to early 80s).
Sibling pairs are a powerful way to test whether our predictors capture genetic (biological) effects.
Siblings share their parental background and much of their environment but differ in about half of their DNA.
Any differences between siblings that we can predict from their difference in DNA are likely driven by direct genetic effects.
We demonstrate our ability to predict, from DNA alone, which sibling expresses a trait (e.g., type 2 diabetes or red hair),
or, in the case of quantitative phenotypes,
which sibling has a higher value (e.g., taller or higher BMI).