Accuracy of Gene Scores when Pruning Markers by Linkage Disequilibrium

Abstract

Objective: Gene scores are often used to model the combined effects of genetic variants. When variants are in linkage disequilibrium, it is common to prune all variants except the most strongly associated. This avoids duplicating information but discards information when variants have independent effects. However, joint modelling of correlated variants increases the sampling error in the gene score. In recent applications, joint modelling has offered only small improvements in accuracy over pruning. We aimed to quantify the relationship between pruning and joint modelling in relation to sample size. Methods: We derived the coefficient of determination R² for a gene score constructed from pruned markers, and for one constructed from correlated markers with jointly estimated effects. Results: Pruned scores tend to have slightly lower R² than jointly modelled scores, but the differences are small at sample sizes up to 100,000. If the proportion of correlated variants is high, joint modelling can obtain modest improvements asymptotically. Conclusions: The small gains observed to date from joint modelling can be explained by sample size. As studies become larger, joint modelling will be useful for traits affected by many correlated variants, but the improvements may remain small. Pruning remains a useful heuristic for current studies.

Neurobehavioural Correlates of Obesity are Largely Heritable

“…To test the robustness of PPS-s, we first tested the impact of not pruning and applying p-value cutoffs. In a pruned PPS, features are omitted that a) correlate above criterion to another feature and b) have lower correlation with BMI than the other feature (75). In a PPS with p-value cutoff, features are omitted that have an above-criterion uncorrected p-value when correlated with BMI Neither pruning nor a p-value cutoff improved the predictive ability of the PPS-s (see SI Results).…”

A flexible and parallelizable approach to genome‐wide polygenic risk scores

“…In a comparative NEWCOMBE ET AL. | 731 study they performed similarly, both outperformed simple approaches based on pruning and p-value thresholding, which is to be expected if there are multiple causal variants in LD (Dudbridge & Newcombe, 2015).…”

“…Both LDPred and lassosum account for LD using genetic correlation estimates from external reference data. In a comparative study they performed similarly, both outperformed simple approaches based on pruning and p ‐value thresholding, which is to be expected if there are multiple causal variants in LD (Dudbridge & Newcombe, ).…”

Making Optimal Use of and Extending beyond Polygenic Additive Liability Models

“…Dudbridge and Newcombe [4] address a common question when generating multigenic risk scores, i.e. whether pruning should be performed or not for markers in high linkage disequilibrium (LD).…”

“…whether pruning should be performed or not for markers in high linkage disequilibrium (LD). They demonstrate that pruning is useful and loses little information in currently available sample sizes, while full consideration of all genetic markers with adjustment for the LD structure may offer more clear advantages when sample sizes become larger, in the millions of participants [4] . Such sample sizes are already feasible and they will probably become more common in the future with the advent of multiple biobanks, as well as widespread genome testing in routine healthcare or even among healthy people, even though the practical value of these behemoth datasets is still unclear.…”