P18-53 AOrTA: a reliable in silico tool to predict acute oral toxicity for regulatory purposes

In the context of the 3Rs, acute oral toxicity is one of the last endpoints of REACH Annex VII which needs to be replaced by new alternative methods (NAMs). Existing in silico tools are struggling to provide models with good performance and reliable prediction in regards of QSAR Assessment Framework [1] for regulatory use. In this context, a new in silico global model named AOrTA v1.0 for Acute Oral Toxicity Alert model, was developed to predict acute toxicity estimation (ATE) for acute oral toxicity (AOT) according to the CLP classification for regulatory purpose. The dataset of the model has been constituted from AOT studies extracted from ECHA dossiers and scientific literature. Only organic monoconstituents were used to develop the model. Studies were validated individually in agreement with the OECD guidelines in order to build a model using a high-quality dataset. Dataset splitting was performed using a genetic algorithm, optimising parameters to ensure structural representativity between the training and test sets. Then, a machine learning approach was applied using structural features of the substances, and an applicability domain was defined based on structural analogues using Tanimoto similarity index. Furthermore, analogues provided to the users were identified using Tanimoto similarity index and the mechanisms of toxic action (MechoA). For a test item, the ATE classification is predicted using AOrTA as “Cat. 1 or 2”, “Cat. 3”, “Cat. 4” or “non-classified” according to CLP. Whenever a test item is predicted to be “Cat. 1 or 2”, a refinement of the prediction is performed by the model based on the closest analogues from the training set to distinguish Cat. 1 and Cat. 2 classifications for the test item. In the same way, a prediction refinement to provide a GHS category is performed for substances predicted to be non-classified according to CLP ( i.e . “GHS category 5” or “non-classified according to GHS”). Furthermore, AOrTA v1.0 is linked to local models with restrictive applicability domain for specific chemical substances ( e.g. simple ester, simple carboxylic acid, simple alcohol, thiol, …). Therefore, whenever a test item falls within a local model applicability domain, a consensus prediction of the AOT category according to CLP is provided to the user based on AOrTA global model and local models. AOrTA v1.0 was developed to assess AOT based on structural features of the substances to predict ATE with high confidence. In a near future, further local models will be developed for substances that are not well-covered by the global model in order to improve confidence and reliability of the AOT predictions.