TL;DR
This paper introduces MR-Path, a novel latent mixture model for Mendelian Randomization (MR) that accounts for "mechanistic heterogeneity" – situations where genetic variants influence an outcome through multiple biological pathways, even if they are valid instruments. By clustering genetic variants based on their estimated causal effects, MR-Path can uncover these distinct mechanisms, offering a more nuanced understanding of causal relationships than traditional MR methods that assume a single, common causal effect.
What Problem Does This Paper Address?
Mendelian Randomization (MR) is a powerful tool for inferring causal relationships between exposures and outcomes using genetic variants as instrumental variables. A core assumption in most MR analyses is that all genetic instruments identify a single, common causal effect. However, this "effect homogeneity" assumption can be unrealistic. Genetic variants can influence outcomes through multiple biological pathways (pleiotropy), and even if these variants are valid instruments (i.e., they don't violate the core MR assumptions), they might still reflect different causal mechanisms. This "mechanistic heterogeneity" can lead to misleading or incomplete causal effect estimates. Traditional MR methods often struggle to disentangle these multiple pathways, potentially masking important biological insights.
Methods
The authors propose MR-Path, a latent mixture model designed to address mechanistic heterogeneity. The core idea is to group genetic variants (SNPs) that exhibit similar estimated causal effects, assuming these groups represent distinct biological pathways or mechanisms.
Key methodological components include:
- Latent Mixture Model: MR-Path models the variant-specific causal effects (Wald estimands) as coming from a mixture of Gaussian distributions. Each component of the mixture represents a distinct causal mechanism, with its mean ($\mu_k$) estimating the average causal effect for that mechanism and its variance ($\sigma_k^2$) quantifying within-mechanism heterogeneity.
- Monte-Carlo Expectation-Maximization (MC-EM) Algorithm: To fit the mixture model, the authors develop an MC-EM algorithm. This iterative approach handles the latent variables (cluster assignments) and estimates model parameters. The "Monte-Carlo" aspect is crucial for approximating the Q-function in the E-step, which is necessary when analytical solutions are intractable.
- Ascent-Based MC-EM: To ensure convergence and address potential issues with Monte-Carlo error, they incorporate an ascent-based strategy that dynamically adjusts the Monte-Carlo sample size.
- Modified Bayesian Information Criterion (BIC): A modified BIC is proposed for selecting the optimal number of clusters ($K$), balancing model fit with complexity.
- Error-in-Variables Regression: MR-Path uses an error-in-variables approach to estimate variant-specific causal effects, making it robust to weak instrument bias, a common challenge in MR.
Datasets
The paper utilizes publicly available summary-level data from large-scale Genome-Wide Association Studies (GWAS). Specifically, they use data for:
- HDL Cholesterol (HDL-C) and Coronary Heart Disease (CHD): Data from multiple large-scale GWAS datasets for plasma lipids (HDL-C, LDL-C, triglycerides) and CHD.
- Adiposity (BMI) and Type II Diabetes (T2D): Data from GWAS summary datasets for BMI and T2D.
These datasets are preprocessed and homogenized using a three-sample summary-data MR design.
Key Insights and Ideas
- Mechanistic Heterogeneity is Pervasive: The paper argues that assuming a single causal effect for all genetic instruments in MR is often unrealistic. Multiple biological pathways can influence an outcome, even if instruments are valid.
- Clustering Reveals Mechanisms: By clustering SNPs based on their estimated causal effects, MR-Path can identify distinct underlying biological mechanisms that influence the exposure-outcome relationship.
- Robustness to Weak Instruments: The use of an error-in-variables regression framework makes MR-Path more reliable than methods that directly model ratio estimates, especially when dealing with weak genetic instruments.
- Data-Driven Model Selection: The modified BIC allows for data-driven selection of the number of causal mechanisms, reducing the need for strong prior domain knowledge.
Key Results
- HDL-C and CHD Example: MR-Path identified two clusters for the HDL-C and CHD relationship. One cluster (larger, $\pi_1 \approx 0.82$) showed a negative effect ($\mu_1 \approx -0.343$), potentially linked to inverse associations with LDL-C and ApoB. The second cluster (smaller, $\pi_2 \approx 0.18$) showed a positive effect ($\mu_2 \approx 0.14$), possibly related to mechanisms regulating HDL particle size. Heatmap analysis of lipoprotein subfractions supported these distinct patterns.
- Adiposity and T2D Example: For the BMI and T2D relationship, MR-Path identified two clusters. A larger cluster ($\pi_2 \approx 0.88$) showed a positive effect ($\mu_1 \approx 0.77$), while a smaller cluster ($\pi_1 \approx 0.12$) indicated a very large negative effect ($\mu_1 \approx -12.4$). Further investigation suggested the large negative effect in the smaller cluster might be due to horizontal pleiotropy rather than a direct negative causal effect of adiposity.
- Simulation Studies: Numerical simulations demonstrated that the MC-EM algorithm provides accurate parameter estimates and that the approximate confidence intervals achieve good coverage probabilities. The modified BIC criterion was effective in selecting the correct number of clusters, particularly with stronger instruments.
Limitations and Future Work
- Distinguishing Mechanisms: While MR-Path identifies clusters of SNPs with similar effects, it doesn't inherently distinguish what those mechanisms are. Post-hoc analyses with external data are needed for this interpretation.
- Computational Efficiency: The MC-EM algorithm can be computationally intensive, especially for larger datasets, due to the need for a large number of Monte-Carlo samples.
- Future Work:
- Multivariate MR-Path: Extending the model to handle multiple risk exposures simultaneously.
- Correlated SNPs: Relaxing the independence assumption for SNPs to account for potential correlations.
- Variational EM: Exploring more computationally efficient inference methods like variational EM.
- Interpreting Mechanisms: Developing methods to better integrate external data for mechanism identification.
Summary updated