Development of Machine Learning models for molecular toxicity predictions

The increasing use of Machine Learning (ML) in the drug and food industry is undeniable and it is important to underline how the use of in silico toxicology predictions of any substance which can get in contact with the human body helped understanding the effect of chemicals on it and the environment. A wide variety of both European and worldwide programmes and guidelines underline the importance of efficient toxicology predictions and identify in silico methods and computational toxicology as valuable predictive tools, due to the wide amount of data that can be processed as well as the lower costs and waiting time for obtaining valuable data on chemicals testing. The rise in ML methods apt at predicting adverse outcome is nowadays especially focusing on addressing damage to organs such as liver and kidneys as well as mutagenic effects on cells. The aim of the present work is the development of two different models, one able to efficiently predict the mutagenic effect and the other the hepatotoxicity of different molecules. The compound sets used for model development were selected among many published ones, gathered in the free online platform VEGA. These datasets, split in training and test sets, have been used both for training the models and for predicting the molecules of the external test sets in order to assess the efficiency of the developed methods. The possibility to create new and more efficient ML methods using the right combination of descriptors and classifiers was examined, with particular attention to the statistical values indicating the number of the correctly predicted compounds. The assessment of the developed models was made comparing the obtained relevant statistical data with the ones reported in the published methods