The problem of selecting relevant descriptors in predicting the toxicity of chemicals

Abstract

Introduction. Mathematical models are widely applicable in conducting toxicological studies and can be used to fill gaps that arise in the assessment of chemical safety. Most of the attention is paid to the study of algorithms for constructing models, rather than approaches to choosing the most informative features.
 The purpose of this study is to highlight aspects of the problem of choosing useful variables during mathematical modeling.
 Material and methods. SMILES and molecular descriptors for organothiophosphates were generated in the interactive Google Colaboratory environment based on the program code using the RDKit, Mordred software. Using the tools of the scikit-learn Ver. 1.2.2 library, features were selected by filtering and by recursive feature exclusion. The values of acute oral toxicity parameters were taken from official information sources about chemicals. The obtained models are subjected to an internal validation procedure to evaluate the performance of the models.
 Results. It should be noted that models where recursive exclusion of features was used have better characteristics than models based on descriptors selected by the filtering method. In particular, the acute toxicity prediction model for organothiophosphates based on the decision tree method with recursive exclusion of features has a high coefficient of determination (R2=0,91713), a relatively small root-mean-square error (RMSE= 0,35099), as well as high values of the cross-validation coefficient of determination (Q2LOO= 0,79756).
 Limitations. The results obtained can be used only in predicting the toxicity of the specified group of chemicals with a similar mechanism of action.
 Conclusion. The use of mathematical modeling is a promising tool for assessing the toxicity of chemicals, which has a number of features: on the one hand, it is a quick and convenient resource for screening the toxicity of substances, on the other hand, the model needs to be trained based not only on reliable research data, but also to carry out a qualitative selection procedure for signs that make a significant contribution to the functioning of the prognostic model.