Utilização de modelos de machine learning para prever resultados de reações químicas com aplicações ambientais

Abstract

Volatile organic compounds (VOCs), emitted by natural and anthropogenic sources, participate in complex photochemical processes in the atmosphere, influencing air quality and climate by forming oxidants and secondary aerosols. Understanding their chemical transformations and potential environmental impacts remains a relevant challenge for the development of monitoring and control strategies. In this work, chemoinformatics tools and machine learning models were applied to simulate the formation of oxidation products from molecules present in diesel fuel and to evaluate their toxicological potential. Specific atmospheric reactions involving hydroxyl radicals (OH), ozone (O3), and atomic oxygen (O) were modeled in a Python environment using the RDKit library. From the generated structures, molecular descriptors were calculated and used to build classification models for mutagenicity prediction, based on Ames test data. Among the algorithms evaluated, the Extra Tree Classifier achieved better performance than the deep neural network implemented in Keras, demonstrating greater robustness to database imbalance. The results show that integrating computational simulation of atmospheric reactions with predictive modeling constitutes a consistent strategy for investigating and screening compounds derived from fossil fuels from an Environmental Chemistry perspective.