Fig. 1

Overall scheme of PredAOT. PredAOT uses the chemical structure as an input. Thereafter, the molecular fingerprint (i.e., ECFP4) is used as an input feature for models in PredAOT. PredAOT is composed of one classification model (i.e., “AOT classifier”) and two regression models (i.e., “toxic regressor” and “less or non-toxic regressor”). The AOT classifier is used for prediction as “toxic” (LD₅₀ ≤ 300 mg/kg) or “less or non-toxic” (LD₅₀ > 300 mg/kg) for a given compound. If a compound is predicted to be toxic, PredAOT predicts the LD₅₀ of the compound using the toxic regressor trained with compounds with LD₅₀ ≤ 300 mg/kg. If a compound is predicted to be less or non-toxic, PredAOT predicts the LD₅₀ of the compound using the less or non-toxic regressor trained with compounds with LD₅₀ > 300 mg/kg. All these procedures are equally applied to the AOT prediction process in mice and rats