Exploring the chemical space of FDA-approved molecules: A dataset-driven approach using machine learning

• Ano de defesa: 2025
• Autor(a) principal: Soares, Artur Caminero Gomes
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Aprendizado de máquina; Chemical Space; Cheminformatics; Drug Design; Druglikeness; Espaço químico; Machine Learning; Planejamento de fármacos; Quimioinformática
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/9/9138/tde-09062025-165507/

Small molecules (SM) possess remarkable chemical diversity, emphasizing the extensive chemical space (CS) they occupy. Within this vast CS, molecules with the potential to interact with biological targets hold particular significance for medicinal chemistry. The enormity and complexity of this space highlight the need for effective tools to navigate it, with the concept of drug-likeness serving as a valuable approach.This research explores drug-likeness through the creation of a dataset comprising FDA-approved molecules, which is contrasted against three negative datasets containing molecules from PDB, ZINC, and compounds with established toxicity. Machine learning (ML) models were developed using this dataset, achieving robust predictive performance. These models facilitated an in-depth analysis of the molecular descriptors most influential in distinguishing drug-like molecules.Key properties identified include established metrics such as logP, topological polar surface area (TPSA), and the counts of hydrogen bond donors and acceptors. The study also highlights the significance of topological descriptors, with e-state indices playing a critical role in characterizing individual atoms. By integrating cheminformatics with ML, this work provides valuable insights into the chemical space, offering a foundation for advancements in drug discovery and development.