Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

In Environmental pollution (Barking, Essex : 1987)

Semivolatile organic compounds (SVOCs) in air can react with hydroxyl radicals (OH), nitrate radicals (NO3) and ozone (O3). Two questions regarding SVOC reactivity with OH, NO3 and O3 in the gas and particle phases remain to be addressed: according to the existing measurements in the literature, which are the most reactive SVOCs in air, and how can the SVOC reactivity in the gas and particle phases be predicted? In the present study, a literature review of the second-order rate constant (k) was carried out to determine the SVOC reactivity with OH, NO3 and O3 in the gas and particle phases in ambient and indoor air at room temperature. Measured k values were available in the literature for 90 polycyclic aromatic hydrocarbons (PAHs), polychlorinated biphenyls (PCBs), organophosphates, dioxins, di(2-ethylhexyl)phthalate (DEHP) and pesticides including pyrifenox, carbamates and terbuthylazine. PAHs and organophosphates were found to be more reactive than dioxins and PCBs. Based on the obtained data, quantitative structure-activity relationship (QSAR) models were developed to predict the k value using quantum chemical, molecular, physical property and environmental descriptors. Eight linear and nonlinear statistical models were employed, including regression models, bagging, random forest and gradient boosting. QSAR models were developed for SVOC/OH reactions in the gas and particle phases and SVOC/O3 reactions in the particle phase. Models for SVOC/NO3 and SVOC/O3 reactions in the gas phase could not be developed due to the lack of measured k values for model training. The least absolute shrinkage and selection operator (LASSO) regression and random forest models were identified as the most effective models for SVOC reactivity prediction according to a comparison of model performance metrics.

Wei Wenjuan, Sivanantham Sutharsini, Malingre Laeticia, Ramalho Olivier, Mandin Corinne


Machine learning, QSAR model, Random forest, Reactivity, Regression, SVOC