Models of quantitative relationship “Structure – activity” in performing preliminary toxicological assessment of chemicals

Abstract

Introduction. In vivo testing of a huge number of chemical compounds is difficult from an ethical point of view, time-consuming, depends on a large number of objects of animal origin and requires large material costs for conducting experiments. Therefore, there is a need for new thinking to optimize the conduct of toxicological studies. The purpose of this study is to substantiate the possibility of using structure-activity models in the framework of a preliminary assessment of chemicals toxicity. Materials and methods. The study included three groups of chemicals including organothiophosphates, triazoles, and carbamates. The calculation of descriptors based on SMILES, the construction and validation of regression models was carried out using the tools of the Scikit-learn Version 1.2.2 library in an interactive cloud environment working with the Google Colaboratory program code. Results. When comparing a number of models for predicting oral toxicity, it was revealed that a model based on decision trees has the best predictive ability for organothiophosphates and triazoles: 70.1% and 69.5% of cases of descriptor changes led to a change in the endpoint value, respectively; a model for predicting carbamate toxicity based on a random forest explains 53.1% of the observed variance common log (1/DL50). Limitations. The study is limited to the area of distribution of the obtained mathematical models. Conclusion. As the study showed, the constructed models can explain only some part of the studied effect, therefore, models based on the structure-activity relationship should be used exclusively for preliminary assessment of the toxicity of chemicals, as a screening tool.