Abstract: Objective To quickly evaluate the potential neurotoxicity of levornidazole for injection and its main impurities so as to provide reference for safer clinical medications. Methods The modeling data were prepared using online databases. The artificial neural network and support vector machine algorithm were used to establish two classification prediction models with different principles for interactive rapid evaluation. The characteristic domain was used to explore the causes of toxicity. Results The false positive and false negative results of prediction of models numbered ANN-3 and SVM-2 were less than 4.0%, and these two models were highly sensitive, accurate and robust. The new models were used to predict and evaluate the potential neurotoxicity of levornidazole for injection and its three main impurities, and the reliability exceeded 90%. Analysis of the characteristic domain showed that hydroxyl substitution in the structure increased molecular polarity, which might account for its potential neurotoxicity. Conclusion Levononidazole and its main impurities show potential neurotoxicity, so we need to keep close watch on the safety of the injection in clinical use. This study is expected to provide reference for rapid identification and evaluation of the toxicity of other antibiotics and their impurities.

Key words: computational model, database technology, rapid evaluation, levornidazole for injection, potential neurotoxicity of impurities