Quantification of an Adverse Outcome Pathway network by Bayesian regression and Bayesian network modelling

The Adverse Outcome Pathway (AOP) framework has gained international recognition as a systematic approach linking mechanistic processes to toxicity endpoints. Nevertheless, successful implementation into risk assessments is still limited by the lack of quantitative AOP models (qAOPs) and assessment of uncertainties. The few published qAOP models so far are typically based on data-demanding systems biology models. Here, we propose a less data-demanding approach for quantification of AOPs and AOP networks, based on regression modelling and Bayesian networks. We demonstrate this approach with the proposed AOP #245 "Uncoupling of photophosphorylation leading to reduced ATP production associated growth inhibition", using a small experimental dataset from exposure of Lemna minor to the pesticide 3,5-dichlorophenol. The AOP-BN reflects the network structure of AOP #245 containing two molecular initiating events (MIEs), three key events (KEs) and one adverse outcome (AO). First, for each dose-response and response-response (KE) relationship, we quantify the causal relationship by Bayesian regression modelling. The regression models correspond to dose-response functions commonly applied in ecotoxicology. Secondly, we apply the fitted regression models with associated uncertainty to simulate 10,000 response values along the predictor gradient. Thirdly, we use the simulated values to parameterise the conditional probability tables of the BN model. The quantified AOP-BN model can be run in several directions. (1) prognostic inference: run forward from the stressor node to predict the AO level, 2: diagnostic inference: run backward from the AO node, and (3) omnidirectionally: run from the intermediate MIEs and/or KEs. Internal validation shows that the AOP-BN can obtain a high accuracy rate, when the run from intermediate nodes, and when a low resolution is acceptable for the AO. While the performance of this AOP-BN is limited by the small dataset, our study demonstrates a proof-of-concept: the combined use of Bayesian regression modelling and Bayesian network modelling for quantifying AOPs. This article is protected by copyright. All rights reserved.