Statistical causal discovery in developing adverse outcome pathway

Statistical causal discovery (SCD) has the potential to advance the development and evaluation of adverse outcome pathways by inferring causal relationships directly from data. However, ecotoxicology data often have challenges for SCD applications, such as violations of SCD algorithm assumptions and small sample sizes. As a proof of concept, we applied DirectLiNGAM-a direct method for estimating a linear non-Gaussian acyclic model, a representative SCD algorithm-to three types of ecotoxicology datasets: bivariate dose-response relationships, bivariate response-response relationships, and a multivariate dataset with a known causal structure involving thyroid hormone disruption in zebrafish. A DirectLiNGAM identified correct causal directions with high statistical reliabilities in three of four bivariate dose-response cases, even when assumptions such as linearity and non-Gaussianity were partially violated. In contrast, response-response cases did not yield a single dominant direction in two of three cases, likely due to the limited sample size. In the multivariate case, some inferred graphs resembled the expert-curated causal graph but exhibited low statistical stability. By increasing an effective sample size through pooling datasets with a shared causal ordering, the inference became more stable and more closely matched the expert-curated graph. These results demonstrate the utility of SCD in identifying relevant key events and the causal relationships under realistic ecotoxicological constraints.