For copies of the poster or questions: email@example.com
Please visit our page in silico Endocrine Disruption Screening service if you are interested in working with us.
There is a need for new approach methodologies (NAMs) to identify potential endocrine disrupting chemicals involved in key events of endocrine pathways including binding to receptor proteins. In 2018 a guidance describing how to perform hazard identification for endocrine-disrupting properties by following the scientific criteria which are outlined in EU 2017/2100 and EU 2018/605 for biocidal products and plant protection products, respectively, was published by ECHA and EFSA. In this guidance computational approaches are proposed as line of evidence for endocrine activity assessment. An in silico screening workflow which follows this guidance is described in this study.
It employs 108 freely available and commercial models covering 27 receptors: 62 (Q)SARs, 19 rule-based profilers, 18 models of receptor interactions and 6 ToxCast pathway models. It addresses EATS (Estrogen, Androgen, Thyroid, Steroidogenesis) and “other” modalities as well as the Mode of Action agonist, antagonist or binding.
The issues with this process is that each model has its own terminology for results, definition of whether the prediction is within the applicability domain as well as different specificity and model reliability.
In order to reach a consensus outcome, it was first necessary to devise an ontology from the output of the different models into a normalised input. The model specificity, model reliability and type of model requires scoring to provide a weighting to the predictions. For example, the results of the profilers are primarily used to confirm the results of the QSAR as these approaches are known to give high rate of false positives. The molecular modelling and the in vitro ToxCast results are rated according only to the reliability of the models as the applicability domain is not an issue.
All the models are applied and the results are combined with an algorithm that considers existing experimental data, the reliability of the model and the uncertainty of each prediction. The predictions are then combined by receptor and an overall conclusion for each receptor is given. When models with the same level of uncertainty disagree an expert assessment of the nearest neighbours is carried out to get to a final conclusion. Positive predictions are used to give indication on the mechanism of action for the endocrine disruption.
As this workflow is very labour intensive, a KNIME workflow was coded to automate the data normalisation, weighting process and report generation. Two examples are demonstrated: butylparaben, correctly predicted as active towards the estrogen receptor, and triclosan, correctly predicted to be active towards the androgen receptor.