Développement de modèles QSPR/QSAR pour la prédiction de propriétés physicochimiques et biologiques des substances organiques

Abstract

The physicochemical and biological properties of organic compounds play a crucial role in various fields such as drug manufacturing, development of analytical methods, and toxicity assessment. Prediction models for structure-property relationships (QSPR) and structure-activity relationships (QSAR) offer a promising solution for rapidly and cost-effectively estimating these properties. In this study, we developed QSPR and QSAR models using relevant chemical descriptors and chemometric techniques. The dataset included the structures of 472 organic molecules sourced from the PubChem and Mol-Instincts databases. 383 descriptors were calculated for each structure, with 354 obtained using the MOE software and 29 via the pkCSM server. After preprocessing the data, we selected 264 descriptors, of which 222 were used as independent variables and 26 as responses. The dataset was divided using the Kennard Stone algorithm. An initial modeling was performed using multiple linear regression (MLR) with simple descriptors that are easily accessible and straightforward to use. Various variable selection methods (forward, backward, stepwise, and genetic algorithm) were employed. The chosen responses for the chemometric models were: logP(o/w), h_logP, logS, h_logS, mr, h_mr, TPSA, Caco2 permeability, Intestinal absorption, BBB permeability, CNS permeability, Oral Rat Chronic Toxicity (LOAEL), and Minnow toxicity. MLR yielded models with R² values ranging from 0.252 to 0.987. Among the 14 studied responses, 12 models achieved R² ≥ 0.6. To improve predictions, we also explored models based on artificial neural networks (ANN). The ANN models outperformed MLR significantly, with R² values ranging from 0.394 to 0.999. Among the 14 studied responses, 12 ANN models achieved R² ≥ 0.7, and 9 models achieved R² ≥ 0.9. In conclusion, the results confirm the effectiveness of MLR and ANN models for accurate modeling and prediction of the studied organic molecule properties. These approaches offer promising prospects for rapid and cost-effective estimation of physicochemical and biological properties of organic compounds.