Quantitative structure activity relationship–based prediction of acute exposure guideline levels for aliphatic compounds

The rapid advancement of modern industry has led to a substantial escalation in the probability of accidents involving hazardous chemicals across the life cycle of products, including storage, use, and disposal. The increasing recognition of the impact of toxic chemicals on human health underscores the growing significance of conducting research on toxicity indicators. In this study, the acute exposure guideline levels (AEGL) of aliphatic compounds were predicted using the quantitative structure activity relationship (QSAR) method. We collected and organized a sample set of 90 aliphatic compounds from the U.S. Environmental Protection Agencys' database. The molecular structures of these compounds were graphed, and a genetic algorithm was used to select eight feature molecular descriptors as input variables for our models. We developed individual models, namely, gradient boosting decision tree (GBDT), extreme gradient boosting (XGBoost), and extremely randomized trees (ERT), to forecast AEGL values. Subsequently, we used a voting regressor (VR) model to combine these three models and address any potential limitations in their predictions. Furthermore, we conducted a comparative analysis of the performance of these four models on both training and testing datasets, with the VR model demonstrating superior performance. In the VR model, the R2 values for the training set and test set were 0.940 and 0.951, respectively. The root-mean-square error values for the training set and test set were 0.321 and 0.143, respectively. The mean absolute error values for the training set and test set were 0.155 and 0.104, respectively. Williams plots were utilized for characterizing the model's applicability domain. The results demonstrate that a majority of the data points fall within this domain, affirming the suitability of the established model for predicting samples within it. This study uses QSAR methodology to establish GBDT, XGBoost, ERT, and VR models for AEGL prediction, thereby providing robust theoretical and technical support for constructing a toxicity index system.