MRSL: a causal network pruning algorithm based on GWAS summary data

Lei Hou, Zhi Geng, Zhongshang Yuan, Xu Shi, Chuan Wang, Feng Chen, Hongkai Li, Fuzhong XueBriefings in Bioinformatics 2024 · DOI

TL;DR

This paper introduces MRSL, a novel algorithm that leverages Mendelian Randomization (MR) and graph theory to infer causal networks from publicly available GWAS summary data. MRSL efficiently prunes spurious causal links from a preliminary network, achieving higher accuracy and significantly faster computation times compared to existing methods, and has been successfully applied to a large-scale biological dataset.

What Problem Does This Paper Address?

Understanding the complex web of causal relationships between biological traits and diseases is a fundamental challenge in biomedical research. While Mendelian Randomization (MR) has emerged as a powerful tool to infer causal links using genetic variants as instrumental variables, existing methods often struggle to construct comprehensive causal networks. Specifically, inferring conditional causal structures (direct effects between variables, accounting for other factors) from purely observational data, especially when dealing with large-scale genome-wide association studies (GWAS) summary statistics, remains difficult. Current approaches often require individual-level data, are computationally intensive, or rely on strong assumptions like causal sufficiency (no unmeasured confounders among all variables). This paper aims to bridge this gap by developing an algorithm that can learn conditional causal networks efficiently and accurately using only readily available GWAS summary data.

Methods

The core of MRSL is a two-step process that refines a "marginal causal graph" (representing all pairwise causal effects) into a "conditional causal graph" (representing direct causal effects).

  1. Marginal Causal Graph Generation: This initial graph is constructed using pairwise bidirectional MR or by summarizing existing MR studies.
  2. Network Pruning (MRSL Algorithm):
    • Topological Sorting: The algorithm first computes a topological sort of the nodes in the marginal causal graph. This ordering is crucial for efficient processing.
  3. MR-Separation: Instead of traditional d-separation (which assumes no unmeasured confounders), MRSL introduces "MR-separation." This novel criterion uses Multivariable Mendelian Randomization (MVMR) to test for conditional independence between two variables, explicitly accounting for potential unmeasured confounding. MVMR estimates the direct effect of one variable on another while adjusting for the genetic associations of other relevant variables.
  4. Sufficient Separating Sets: MRSL considers three candidate sets of variables to adjust for in MVMR: (1) all nodes on open paths between the two variables, (2) a minimal sufficient adjustment set plus mediators, and (3) all variables except the two of interest and colliders.
  5. Iterative Pruning: The pruning process is iterative. If MR-separation holds for an edge in the marginal graph (meaning the direct causal effect is likely zero), that edge is removed. This step is repeated until the graph structure stabilizes, ensuring that spurious edges are pruned and direct causal relationships are identified.

Datasets

The paper utilizes publicly available GWAS summary statistics from the UK Biobank for its applied example. Specifically, it uses data for 26 biomarkers and 44 ICD10-defined diseases from approximately 337,198 individuals of European ancestry. The simulations were conducted using generated data based on various graph structures and parameter settings.

Key Insights and Ideas

  • MR-Separation for Robust Causal Discovery: The introduction of MR-separation, which leverages MVMR to handle unmeasured confounding, is a key innovation. This allows for more reliable causal inference in complex biological systems where unobserved confounders are common.
  • Leveraging GWAS Summary Data: The ability to perform causal network inference using only summary statistics is a significant advantage, as this data is widely accessible and computationally less demanding than individual-level data.
  • Topological Sorting for Efficiency: The use of topological sorting significantly speeds up the pruning process, making MRSL practical for large networks.
  • Pruning vs. Building: MRSL focuses on pruning spurious edges from an initial marginal graph, rather than building a graph from scratch. This approach relies on the accuracy of the initial marginal graph but is efficient and less prone to introducing false positives.

Key Results

  • Performance: In simulations, MRSL achieved up to a 2-fold higher F1 score compared to eight other competitive methods.
  • Speed: MRSL demonstrated a remarkable 100-fold speedup in computing time compared to other methods.
  • Application: When applied to 26 biomarkers and 44 ICD10-defined diseases from the UK Biobank, MRSL identified a conditional causal graph with 192 edges, uncovering many biologically plausible causal links and several novel associations supported by existing literature.

Limitations and Future Work

  • Dependence on Marginal Graph: The accuracy of MRSL is dependent on the quality of the initial marginal causal graph. If true edges are missing from this initial graph, MRSL cannot recover them.
  • Invalid/Weak IVs: While MRSL can converge even with invalid or weak instrumental variables (IVs), it may converge to an incorrect causal graph. The paper suggests incorporating more robust MR methods (like MR-Egger, weighted median, etc.) into MRSL to mitigate these issues.
  • Data Homogeneity: Similar to standard MR, the GWAS summary data for different phenotypes should ideally come from a homogenous population.
  • Future Directions: The authors suggest extending MRSL to incorporate more sophisticated MR methods, address issues like measurement error and selection bias, and explore its application in disease diagnosis and other complex biological systems.

Summary updated

BibTeX
@article{hou2024mrsl,
  author = {Lei Hou and Zhi Geng and Zhongshang Yuan and Xu Shi and Chuan Wang and Feng Chen and Hongkai Li and Fuzhong Xue},
  title = {{MRSL: a causal network pruning algorithm based on GWAS summary data}},
  journal = {Briefings in Bioinformatics},
  year = {2024},
  doi = {10.1093/bib/bbae086},
}
Figure