Causal graphical models in systems genetics: A unified framework for joint inference of causal network and genetic architecture for correlated phenotypes

Elias Chaibub Neto, Mark P. Keller, Alan Attie, Brian S. YandellThe Annals of Applied Statistics 2010 · DOI

Abstract

Causal inference approaches in systems genetics exploit quantitative trait loci (QTL) genotypes to infer causal relationships among phenotypes. However, poorly estimated genetic architectures can compromise these inferences. We present QTLnet, a statistical framework designed to jointly infer a causal phenotype network and its associated genetic architecture. By leveraging the natural randomization of alleles during meiosis, QTLnet uses genotypes as unidirectional anchors to orient causal relationships between phenotypes. We validate this approach through simulation studies and an application to mouse liver gene expression data, demonstrating its ability to distinguish direct from indirect genetic effects.

Introduction

Understanding the flow of information from genetic variation to complex traits is a central goal of systems genetics. Traditional methods often treat the mapping of Quantitative Trait Loci (QTLs) and the inference of phenotype networks as separate, sequential steps. This "siloed" approach frequently leads to circularity: an incorrect phenotype network can bias QTL mapping, and vice versa. A common artifact of this separation is the appearance of "genetic hotspots"—genomic regions seemingly controlling numerous traits—which often represent a single upstream effect propagating through a network.

The QTLnet Framework

To address these challenges, we propose QTLnet, which treats the phenotype network and genetic architecture as a single, joint inference problem. The model utilizes Homogeneous Conditional Gaussian Regression (HCGR) to integrate discrete QTL genotypes and continuous phenotypic data into a unified graphical model.

A complex joint network of QTLs and phenotypes illustrating the integrated modeling of genetic architecture and trait interactions.

The core of the QTLnet approach is the simultaneous estimation of edges between QTLs and phenotypes ($Q \to Y$) and edges between phenotypes themselves ($Y_i \to Y_j$). By exploring the space of possible graph structures using a Markov Chain Monte Carlo (MCMC) algorithm and Bayesian Model Averaging, the framework identifies the most probable causal "wiring diagram" of the biological system.

Breaking Markov Equivalence with Causal Anchors

A fundamental problem in causal inference is distribution equivalence: without additional information, it is often impossible to distinguish between $Y_1 \to Y_2$ and $Y_2 \to Y_1$ based solely on observational data. QTLnet solves this by using biology as a "causal anchor." Because genotypes are determined at meiosis and are not influenced by phenotypes, the direction of the arrow from a QTL to a phenotype is fixed ($Q \to Y$).

QTL anchors allow for the orientation of causal arrows between phenotypes that would otherwise be distribution equivalent.

As illustrated in the figure above, the introduction of a QTL node $Q$ acting on a phenotype $Y_2$ breaks the symmetry of the phenotype chain. Different orientations of the arrows between $Y_1, Y_2,$ and $Y_3$ result in different conditional independence structures (e.g., creating or removing colliders), making them statistically distinguishable.

Results and Validation

We validated QTLnet using both simulated and real-world datasets:

  • Simulations: In "strong-signal" scenarios involving a 5-phenotype, 4-QTL network, QTLnet correctly identified the true causal model as the most probable structure in 84.2% of runs. It significantly outperformed traditional sequential methods in orienting causal arrows.
  • Mouse Liver Data: Applying QTLnet to a chromosome 2 hotspot in mouse liver gene expression data (14 transcripts), the model identified Il16 as a primary upstream regulator. This suggests that the observed hotspot is largely an indirect effect of Il16 variation, providing a clear, testable hypothesis for biological follow-up.

Conclusion

QTLnet provides a robust mathematical foundation for untangling complex biological networks. By jointly modeling genetic architecture and phenotype interactions, it reduces false positives in QTL mapping and provides a clearer picture of causal flow. While computationally intensive, the framework represents a significant step toward more accurate "wiring diagrams" of life.

Summary updated

Abstract

Causal inference approaches in systems genetics exploit quantitative trait loci (QTL) genotypes to infer causal relationships among phenotypes. The genetic architecture of each phenotype may be complex, and poorly estimated genetic architectures may compromise the inference of causal relationships among phenotypes. Existing methods assume QTLs are known or inferred without regard to the phenotype network structure. In this paper we develop a QTL-driven phenotype network method (QTLnet) to jointly infer a causal phenotype network and associated genetic architecture for sets of correlated phenotypes. Randomization of alleles during meiosis and the unidirectional influence of genotype on phenotype allow the inference of QTLs causal to phenotypes. Causal relationships among phenotypes can be inferred using these QTL nodes, enabling us to distinguish among phenotype networks that would otherwise be distribution equivalent. We jointly model phenotypes and QTLs using homogeneous conditional Gaussian regression models, and we derive a graphical criterion for distribution equivalence. We validate the QTLnet approach in a simulation study. Finally, we illustrate with simulated data and a real example how QTLnet can be used to infer both direct and indirect effects of QTLs and phenotypes that co-map to a genomic region.
BibTeX
@article{neto2010causal,
  author = {Elias Chaibub Neto and Mark P. Keller and Alan Attie and Brian S. Yandell},
  title = {{Causal graphical models in systems genetics: A unified framework for joint inference of causal network and genetic architecture for correlated phenotypes}},
  journal = {The Annals of Applied Statistics},
  year = {2010},
  doi = {10.1214/09-aoas288},
}
Figure