Objectives The release of toxic gases following an unexpected accident has a significant impact on human health, making it crucial to predict the concentration of these gases during short-term exposure. In this study, we employed Quantitative Structure–Activity Relationship (QSAR) to forecast Protective Action Criteria (PAC) for aliphatic gas compounds Methods A dataset comprising 120 aliphatic gas compounds released by the US Department of Energy (DOE) was collected and organized as sample sets, with their respective molecular structures plotted. The Gradient Boosting Decision Tree (GBDT) model, eXtreme Gradient Boosting (XGBoost) model, Extremely Randomized Trees (ERT) model, and Voting Regressor (VR) model were, respectively, constructed for the prediction of PAC. The performance parameters, including R2, MAE, RMSE, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">${Q}_{loo}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>Q</mi> <mrow> <mi mathvariant="italic">loo</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">${Q}_{ext}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>Q</mi> <mrow> <mi mathvariant="italic">ext</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation>, and others, were integrated to evaluate the models. Additionally, SHapley Additive exPlanations (SHAP) were employed to enhance the interpretability of the VR model, and Williams plots were used to characterize the application domain of the models. Results The VR model demonstrated superior performance. Specifically, the R2 values for the training set and test set in the VR model were 0.902 and 0.905, respectively, while the corresponding RMSE values were 0.419 and 0.333; as for MAE values they were 0.204 and 0.272, respectively; additionally, <InlineEquation ID="IEq3"> <EquationSource Format="TEX">${Q}_{loo}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>Q</mi> <mrow> <mi mathvariant="italic">loo</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation> of the training set was found to be 0.902 and <InlineEquation ID="IEq4"> <EquationSource Format="TEX">${Q}_{ext}^{2}$</EquationSource> <EquationSource Format="MATHML"><math> <msubsup> <mi>Q</mi> <mrow> <mi mathvariant="italic">ext</mi> </mrow> <mn>2</mn> </msubsup> </math></EquationSource> </InlineEquation> of the test set was determined as 0.940. SHAP analysis quantified the contributions of individual molecular descriptors to the VR predictions. The Williams plots showed that over 95% of data points from GBDT model, XGBoost model, ERT model, and VR model fell within this application domain range successfully validating their predictive capabilities within the application context. Conclusion The GBDT, XGBoost, ERT, and VR models were established using the QSAR method in this study to predict PAC. This not only serves as a foundation for supplementing the PAC toxicity index database but also provides robust theoretical and technical support for enhancing the PAC toxicity index system.

Predictive modeling of PAC for aliphatic gaseous compounds based on QSAR

期刊论文
发表日期：

2026年2月16日

查看全文

XiongJun Yuan,
WeiHua Zheng,
JingJie Shi,
ChangHao Chen,
Tao Huang

摘要

Objectives

The release of toxic gases following an unexpected accident has a significant impact on human health, making it crucial to predict the concentration of these gases during short-term exposure. In this study, we employed Quantitative Structure–Activity Relationship (QSAR) to forecast Protective Action Criteria (PAC) for aliphatic gas compounds

Methods

A dataset comprising 120 aliphatic gas compounds released by the US Department of Energy (DOE) was collected and organized as sample sets, with their respective molecular structures plotted. The Gradient Boosting Decision Tree (GBDT) model, eXtreme Gradient Boosting (XGBoost) model, Extremely Randomized Trees (ERT) model, and Voting Regressor (VR) model were, respectively, constructed for the prediction of PAC. The performance parameters, including R², MAE, RMSE, ${Q}_{loo}^{2}$ $Q_{loo}^{2}$ , ${Q}_{ext}^{2}$ $Q_{ext}^{2}$ , and others, were integrated to evaluate the models. Additionally, SHapley Additive exPlanations (SHAP) were employed to enhance the interpretability of the VR model, and Williams plots were used to characterize the application domain of the models.

Results

The VR model demonstrated superior performance. Specifically, the R² values for the training set and test set in the VR model were 0.902 and 0.905, respectively, while the corresponding RMSE values were 0.419 and 0.333; as for MAE values they were 0.204 and 0.272, respectively; additionally, ${Q}_{loo}^{2}$ $Q_{loo}^{2}$ of the training set was found to be 0.902 and ${Q}_{ext}^{2}$ $Q_{ext}^{2}$ of the test set was determined as 0.940. SHAP analysis quantified the contributions of individual molecular descriptors to the VR predictions. The Williams plots showed that over 95% of data points from GBDT model, XGBoost model, ERT model, and VR model fell within this application domain range successfully validating their predictive capabilities within the application context.

Conclusion

The GBDT, XGBoost, ERT, and VR models were established using the QSAR method in this study to predict PAC. This not only serves as a foundation for supplementing the PAC toxicity index database but also provides robust theoretical and technical support for enhancing the PAC toxicity index system.