Modelos de aprendizado de máquina para avaliação preditiva de toxicidade dérmica aguda de compostos químicos

• Ano de defesa: 2020
• Autor(a) principal: Hall, Steven Umualê Silva
• Orientador(a): Andrade, Carolina Horta
• Banca de defesa: Andrade, Carolina Horta, Scotti, Marcus Tullius, Neves, Bruno Junior
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/38995/0013000001spg
• Idioma: por
• Instituição de defesa: Universidade Federal de Goiás
• Programa de Pós-Graduação: Programa de Pós-graduação em Ciências Farmacêuticas (FF)
• Departamento: Faculdade de Farmácia - FF (RMG)
• País: Brasil
• Palavras-chave em Português: Toxicidade dérmica aguda; Aprendizado de máquina; QSPR; Métodos alternativos
• Palavras-chave em Inglês: Acute dermal toxicity; Machine learning; Alternative methods
• Área do conhecimento CNPq: CIENCIAS BIOLOGICAS::FARMACOLOGIA::FARMACOLOGIA GERAL
• Link de acesso: http://repositorio.bc.ufg.br/tede/handle/tede/12928

Introduction: Acute dermal toxicity is a collection of adverse effects that a substance can cause in the first 24 hours of dermal exposure to this chemical. This toxicological property using animals by estimating the lethal dose (LD50), the dose required to kill 50% of individuals of a test population. However, due to public and political pressure on issues related to animal testing, alternative methods are becoming essential to reduce the costs and number of test animals. Computational methods such as machine learning methods have been presented as a reliable alternative to animal testing for hazard assessment. Objectives: The main goal of this study was to develop machine learning models capable of predicting acute dermal toxicity of chemicals, and to make these models available in an online server for the scientific community. Methodology: Acute dermal toxicity datasets where compiled from literature and rigorously curated. Classificatory and multi-classificatory QSPR (Quantitative Structure-property Relationships) models were developed using molecular descriptors and machine learning algorithms and where then used to screen the CosIng library as a case study. Results and discussion: After data curation, 2,622 compounds were kept in the dataset (384 toxic and 2238 non-toxic). To avoid developing biased models, the dataset was balanced, and 768 (384 toxic and 384 non-toxic) were kept for the modeling. The best classificatory models were generated with random forest algorithm and achieved CCR of 79%, SE of 80%, SP of 78%. The best models were used in an integrated hierarchical strategy that obtained ACC= 74% and Recall =74%. In total 33 compounds from CosIng were correctly predicted as toxic. Conclusions: The developed QSPR models were robust and predictive and are capable of predicting acute dermal toxicity efficiently. These are thde first models developed for this endpoint which passed by a rigorous data curation and validation process, being valuable in the prediction of toxicity from new untested compounds. These models are available in a web server through the address https://stoptox.mml.unc.edu .