Transferência indutiva de aprendizagem: uma abordagem de inteligência artificial para avaliação ecotoxicológica de produtos químicos para aves e peixes
Saved in:
| Hovedforfatter: | |
|---|---|
| Publication Date: | 2024 |
| Format: | Master thesis |
| Sprog: | por |
| Source: | Repositório Institucional da UFG |
| Download full: | http://repositorio.bc.ufg.br/tede/handle/tede/13882 |
Summary: | The extensive use of pesticides in planting, including the removal of field margins and new crop protection measures, has been connected to a significant decline in biodiversity in plants, insects, birds, and other animals in the contemporary environment. A comprehensive safety evaluation of chemicals should require toxicity assessment in both the aquatic and terrestrial test species. Due to the application practices and nature of chemical pesticides, avian toxicity testing is considered an essential requirement in the risk assessment process. In silico models have become a feasible alternative to be used as screening tools and within integrated testing methods to prevent animal testing as well as minimize expense and chemical waste as a result of an understanding of the expensive and time-consuming aspects of experimental assays. In this study, multitask quantitative-structure activity relationship (QSAR) models were developed to predict the chemical toxicity of pesticides against four avian species (Anas platyrhynchus, Colinus virginianus, Coturnix japonica, Phasianus colchicus) following OECD guidelines. Initially, a dataset comprising 683 compounds with experimental LD50 values was collected from various sources. Then, ensemble decision-forest single-task models and a multitask learning model were developed using molecular fingerprints (ECFP2). As a result, multi-task models significantly outperform the single-task approach, with acceptable average test MAE values ranging between 0.32–0.54 and r values between 0.62–0.80. Since the "black-box" nature of deep learning can make mechanistic interpretation of predictions difficult, feature contributions to toxicity were assessed following SHapley Additive exPlanations (SHAP) values. The quantitative feature contributions from the local SHAP value computations support the potency toxicity against four bird species. To be more precise, the models showed that the moieties made up of carbanium acid, chlorine, phosphate, and sp2 hybridized carbons positively weighted the anticipated potency across the three tasks. In conclusion, the method created here is a novel computational tool for determining the toxicity of pesticides to four different bird species. |