Construção de modelos de predição Qsa(P)R com objetivos ambientais e toxicológicos

Abstract

This dissertation aimed to develop QSAR (Quantitative Structure–Activity Relationship) and QSPR (Quantitative Structure–Property Relationship) predictive models for the assessment of toxicological and environmental properties of organic compounds of regulatory interest. The study focused on three main groups of compounds: per- and polyfluoroalkyl substances (PFAS), pesticides classified as persistent organic pollutants (POPs), and substituted diphenyl ethers, including polymethoxylated, polyhydroxylated, methoxylated polychlorinated, and hydroxylated polychlorinated derivatives. The environmental persistence, high mobility, and toxic potential of these substances highlight the need for computational tools capable of supporting risk assessment in a rapid, reliable manner while reducing the dependence on experimental testing. The first study aimed to develop QSAR models to predict oral and inhalation toxicity parameters of PFAS in rats and mice using two-dimensional molecular descriptors and partial least squares (PLS) regression. The second study focused on developing QSAR models based on partial least squares discriminant analysis (PLS-DA) to classify the contact toxicity of pesticides toward honeybees (Apis mellifera), enabling the identification of structural patterns associated with the ecotoxicological potential of these compounds. The third study consisted of developing QSPR models to predict the soil organic carbon sorption coefficient (LogKoc) of substituted diphenyl ethers. Different algorithms for estimating LogP and LogS descriptors were compared, and the application of a hybrid descriptor (LogSp), obtained by combining lipophilicity and water solubility, was evaluated. The results demonstrated that LogP-based descriptors exhibited superior predictive performance, with the MLOGP algorithm showing the best performance among the theoretical methods evaluated. Furthermore, models developed using the hybrid Original_LogSp descriptor outperformed those based solely on individual descriptors, demonstrating that the combination of lipophilicity and water solubility improves the representation of physicochemical properties associated with soil sorption. The developed models exhibited satisfactory statistical robustness, stability, and high predictive performance, meeting the recommended validation criteria for QSAR/QSPR studies. The results demonstrate the potential of these computational approaches as valuable tools for toxicological and environmental assessment, contributing to the understanding of the molecular mechanisms underlying toxicity, mobility, and persistence of organic contaminants. Therefore, this dissertation provides predictive models that can support risk assessment, chemical regulation, and the development of safer compounds while reducing the need for experimental testing and promoting more sustainable strategies for protecting human health and the environment.