Predicción de la toxicidad de líquidos iónicos utilizando los descriptores moleculares ECFP y ACSF en conjunto con algoritmos de aprendizaje máquina

• Castro-Vázquez, Arnulfo ^[2] ; García-Guadarrama, Reyna ^[1] ; Gallo Estrada, Marco Tulio ^[1]
  1. [1] Tecnológico Nacional de México

     Tecnológico Nacional de México

     México

     
  2. [2] Autonomous University of Ciudad Juárez, Institute of Engineering and Technology. Chihuahua, Mexico
• Localización: Nova scientia, ISSN-e 2007-0705, Vol. 16, Nº. 32, 2024 (Ejemplar dedicado a: Ahead of print, 16(32), 2024)
• Idioma: español
• Títulos paralelos:
  • Toxicity prediction of ionic liquids using ECFP and ACSF molecular descriptors in conjunction with machine learning algorithms
• Enlaces
  • Texto completo (pdf)
• Resumen

  • This work describes the process of toxicity prediction of ionic liquids, specifically toxicity with respect to the IPC-81 rat cell line. We studied 355 molecular structures of ionic liquids, whose three-dimensional geometry is encoded by means of symbol strings such as Simplified Molecular Input Line Entry System (SMILES) language. The feeding of the input data to the different machine learning models requires that the geometrical and near atomic neighbor information of each ionic liquid be mapped or transformed to numerical vector notation (xi) using the following two molecular descriptors: Atom-Centered Symmetry Functions (ACSF), and Extended Connectivity Fingerprints (ECFP). Three machine learning algorithms: Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR) and Kernel Ridge Regression (KRR) were used to build the predictive mathematical regression model relating the input values xi to the response value represented by the logarithm of the mean effective concentration (yi = logEC50) in the toxicity assessment, using the coefficient of determination (r2) as a metric of the degree of fitness. The results obtained indicated that the ECFP combination with a radial distance of 6 atomic neighbors in conjunction with the KRR algorithm provides the best average fit with r2=0.8602±0.032, and with respect to the ACSF molecular descriptor the best average fit was obtained with the XGBoost algorithm with r2=0.8029±0.055.