QSAR Models for Predicting ERPG Toxicity Index of Aliphatic Compounds

As the awareness of the impact of toxic chemicals on human health continues to grow, research on toxicity indicators has gained increasing significance in scientific discourse. The quantitative structure activity relationship (QSAR) method is utilized in this article to predict the toxicity indicators of emergency response planning guidelines (ERPG) for aliphatic compounds. A collection of 70 aliphatic compounds was gathered and organized from the database provided by the American industrial hygiene association (AIHA). The molecular structures of these compounds were depicted, and a genetic algorithm was employed to select 10 key molecular descriptors as input variables for the model. Subsequently, multiple linear regression (MLR), artificial neural network (ANN), and support vector machine (SVM) models were constructed to accurately predict ERPG. The MLR model demonstrated correlation coefficients (R²) of 0.723 and 0.748 for the training and test sets, respectively. In contrast, the ANN model exhibited R² values of 0.812 and 0.831 for the training and test sets, while the SVM model showcased R² values of 0.833 and 0.858. In terms of internal validation parameters ( \(Q_{{\rm{loo}}}^{\rm{2}}\) ), the MLR model achieved a value of 0.723, while its external validation parameter ( \(Q_{{\rm{ext}}}^{\rm{2}}\) ) was measured at 0.733. Similarly, the ANN model yielded \(Q_{{\rm{loo}}}^{\rm{2}}\) values of 0.812 and \(Q_{{\rm{ext}}}^{\rm{2}}\) value of 0.715, whereas the SVM model attained \(Q_{{\rm{loo}}}^{\rm{2}}\) and \(Q_{{\rm{ext}}}^{\rm{2}}\) values of 0.833 and 0.836. The analysis of comprehensive performance parameters, including R² values and validation parameters, demonstrates that the SVM model surpasses both the MLR and ANN models in terms of predictive performance. The application domain of the models was characterized using Williams plots, indicating that the majority of data fell within the specified range for application suitability. As a result, these well-established models can be effectively utilized to accurately predict samples within their specific application domain. This study employs QSAR method to establish MLR, ANN, and SVM models for predicting ERPG. Henceforth, this research offers substantial theoretical and technical backing in toxicity indicator systems.