Increasing Accessibility of Bayesian Network-Based Defined Approaches for Skin Sensitisation Potency Assessment

Skin sensitisation is a critical adverse effect assessed to ensure the safety of compounds and materials exposed to the skin. Alongside the development of new approach methodologies (NAMs), defined approaches (DAs) have been established to promote skin sensitisation potency assessment by adopting and integrating standardised in vitro, in chemico, and in silico methods with specified data analysis procedures to achieve reliable and reproducible predictions. The incorporation of additional NAMs could help increase accessibility and flexibility. Using superior algorithms may help improve the accuracy of hazard and potency assessment and build confidence in the results. Here, we introduce two new DA models, with the aim to build DAs on freely available software and the newly developed kDPRA for covalent binding of a chemical to skin peptides and proteins. The new DA models are built on an existing Bayesian network (BN) modelling approach and expand on it. The new DA models include kDPRA data as one of the in vitro parameters and utilise in silico inputs from open-source QSAR models. Both approaches perform at least on par with the existing BN DA and show 63% and 68% accuracy when predicting four LLNA potency classes, respectively. We demonstrate the value of the Bayesian network's confidence indications for predictions, as they provide a measure for differentiating between highly accurate and reliable predictions (accuracies up to 87%) in contrast to low-reliability predictions associated with inaccurate predictions.