• applicability domain
• in silico models
• machine learning
• molecular toxicity

The integration of artificial intelligence and machine learning (ML) techniques is becoming increasingly common in computational toxicology and drug development. These methods are recognized as powerful tools for evaluating the safety of compounds, particularly in lead optimization and ADMET analysis. Additionally, they align with the 3Rs framework, which promotes the replacement, reduction, and refinement of animal testing. In this context, my thesis project aimed to develop ML models able to predict the toxicity of chemical compounds related to six different toxicity endpoints: respiratory toxicity, embryotoxicity, nephrotoxicity, mitochondrial toxicity, neurotoxicity, and thyroid toxicity. For each endpoint, 21 different ML models, generated by combining five molecular representations with six different ML algorithms, were developed and tested. After identifying the most reliable model for each endpoint, a consensus prediction strategy was then employed to identify a potential optimal combination of models able to increase the predictive performance. The last part of the project was focused on the development of a novel strategy to determine the applicability domain (AD) of our model and define the chemical space within which predictions can be considered as reliable, thus reducing extrapolation errors and ensuring that the predictions are robust and meaningful. By defining the applicability domain, we aimed to further improve the reliability of our models and enhance their practical utility in toxicological assessments and chemical safety evaluations.