Statistical causal discovery in developing adverse outcome pathway.

Area of Science:

• Ecotoxicology
• Computational Toxicology
• Causal Inference

Background:

• Adverse Outcome Pathways (AOPs) development can be advanced by Statistical Causal Discovery (SCD) methods.
• Ecotoxicology data presents challenges for SCD, including assumption violations and small sample sizes.

Purpose of the Study:

• To evaluate the performance of a representative SCD algorithm, DirectLiNGAM, on diverse ecotoxicology datasets.
• To assess the utility of SCD in identifying key events and causal relationships within AOPs under realistic data constraints.

Main Methods:

• Applied DirectLiNGAM, a method for linear non-Gaussian acyclic models (LiNGAM), to bivariate dose-response, bivariate response-response, and multivariate ecotoxicology datasets.
• Investigated the impact of partially violated assumptions (linearity, non-Gaussianity) and small sample sizes on causal inference.
• Explored dataset pooling strategies to enhance statistical stability of inferred causal graphs.

Main Results:

• DirectLiNGAM accurately identified causal directions in most bivariate dose-response scenarios, even with violated assumptions.
• Response-response datasets yielded ambiguous causal directions, likely due to limited sample sizes.
• In multivariate cases, SCD graphs partially matched expert-curated structures but lacked stability; pooling datasets improved stability and accuracy.

Conclusions:

• SCD, particularly DirectLiNGAM, is a valuable tool for advancing AOP development by inferring causal relationships from ecotoxicology data.
• Addressing data limitations, such as increasing sample size through pooling, is crucial for robust causal inference in ecotoxicology.
• SCD can identify key events and causal links relevant to AOPs, even with imperfect data.