Bayesian Stacking of Quantitative Adverse Outcome Pathways Isolates Primary Drivers of Liver Steatosis

Abstract

The quantification of adverse outcome pathways (qAOPs) involving multiple converging pathways remains a fundamental challenge in toxicology. This difficulty is worsened by small in vitro datasets with missing values and heteroscedasticity, which limit the estimation of non-linear response functions and the robust estimation of uncertainty. To overcome these limitations, we developed a framework for estimating a Bayesian qAOP Ensemble (Bayesian qAOPE) that combines heteroscedastic dose-response functions with continuous splines and linear models. By combining multiple converging non-linear pathways through Bayesian stacking, the most meaningful pathways are identified, and strongly correlated responses are naturally accounted for. Applying our approach to a hepatic steatosis dataset, we demonstrate by exact Bayesian leave-one-out cross-validation that our framework yields an exceptional predictive performance suitable for the derivation of a probabilistically weighted Benchmark Dose (BMD). Our framework is highly adaptable and provides a chemical-agnostic, mathematically founded blueprint for integrating multiple converging pathways in a Bayesian network.