Benchmarking Mendelian randomization methods for causal inference using genome-wide association study summary statistics AJHG (2024)

https://www.cell.com/ajhg/fulltext/S0002-9297(24)00222-2

요약:

MR 분석 방법들을 real-world dataset을 사용해 benchmarking 한 것이 특징. 
다음 세가지를 평가함 type I error control, accuracy of causal effect estimates, replicability.

 

논문을 읽은 이유:

real-world data를 사용해 benchmarking할 경우 어떻게 성능 평가의 기준이 될 정답을 알 수 있는가가 문제일텐데 이 논문에서 어떻게 아 문제를 해결했는지가 가장 궁금함

대충 말하자면, 이를 각 평가하고자 하는 것에 알맞는 exposure와 outcome을 선택하여 해결함. 

 

*replicability 정확한 의미가 뭐지? 다른 데이터셋으로 분석했을 때 replicate되는지인가? 그렇다면 그것으로부터 알 수 있는 정보는? 

 

평가 방법에 대한 더 자세한 설명:

In scenario 1 (population stratification), we conducted negative control studies using three datasets to investigate the performance of MR methods in the presence of population stratification. These negative control studies were carefully designed based on two criteria. First, the outcomes should not be causally affected by the exposures (no causal effect between trait pairs). Second, both the outcomes and exposures should be affected by population stratification. In this analysis, we chose hair color-related traits and tanning ability as negative control outcomes. These
traits are mainly determined at birth and are likely influenced by population stratification.

 

In scenario 2 (pleiotropy), we analyzed trait pairs between 11 exposures and 7 negative control outcomes. The
selected exposures included five adult behavior-related traits and six aging-related traits, while the negative control outcomes were seven childhood-related traits (Table S4). This analysis allows us to assess the influence of pleiotropy, as these trait pairs meet two important criteria: (1) there is no causal link between the trait pairs, as traits developed before adulthood are unlikely to be causally affected by traits developed after adulthood, and (2) out of the 77 trait pairs analyzed, 49 pairs exhibited significant genetic correlations at the nominal level of 0.05, suggesting that pleiotropy is a major confounding factor in causal inference.

 

In scenario 3 (Family-level confounders)

We conducted a comparison between population-based MR and family-based MR to evaluate the influence of family-level confounders.

 

In evaluating the accuracy of causal effect estimates, we examined six pairs of UK Biobank traits where each pair comprised the same trait as both the ‘‘exposure’’ and the ‘‘outcome’’ (Table S6). The UK Biobank dataset was divided equally to obtain exposure and outcome GWAS summary-level data (the UK Biobank dataset are from UK Biobank resources under application number 30186). Importantly, the true causal effects in this analysis were known to be exactly one.17,34 This design allowed us to assess the accuracy of MR methods in estimating causal effects.

To evaluate replicability and power, we focused on a positive control example involving low-density lipoprotein cholesterol (LDL-C) and coronary artery disease (CAD). We applied all the MR methods to six GWAS datasets for LDL-C obtained from five distinct studies (Table S7).

 

'By analyzing multiple GWAS dataset for the same trait, we assessed the replicability.'

<- 이거는 MR 뿐 아니라 GWAS 분석의 replicability 영향도 받는 것 아닌가? GWAS 결과가 같으면 무조건 MR 결과도 같으니까 결국 같은 건가?

GWAS 결과가 다르더라도 exposure beta 값과 outcome beta 값을 regression 했을 때 기울기가 같다면 MR 결과는 같으니까 'GWAS 결과가 같음'은 'MR 결과가 같음'에 속함... 즉, 위 대로 테스트했을 때 'GWAS 결과가 조금 달라도 MR 결과가 로 같게 나오는가'를 보는 건가