Convolutional neural network prediction of molecular properties for aerosol chemistry and health effects

Abstract

Quinones are chemical compounds commonly found in air particulate matter (PM).Their redox activity can generate reactive oxygen species (ROS) and contribute to the oxidative potential (OP) of PM leading to adverse health effects of aerosols. The quinones' OP and ability to form ROS are linked to their reduction potential (RP, measured in volts), a metric for the tendency to lose electrons in redox reactions. Here, we use convolutional neural networks (CNN) as quantitative structure-activity relationship (QSAR) models to relate the one-electron RP of quinones to their molecular structure. For CNN training and testing, a data set of more than 100,000 quinones with associated RP values derived from density functional theory calculations was encoded in simplified molecular input line entry system (SMILES). The best performing CNN model achieved a root mean square error (RMSE) of 0.115 V for an independent test data set and outperformed linear regression models fitted on common molecular descriptors (≥ 0.140 V RMSE). Augmentation methods were newly adapted or applied to support CNN training with smaller data sets, improving RMSE by up to approximately 37% for a data set of 321 molecules. Adjusted for solvent effects, the CNN-derived RP predictions showed good agreement with experimental data.Using the newly developed method, we identified a subset of atmospherically relevant quinones that are likely to have a high OP and play a role in aerosol health effects, which remains to be further elucidated by experimental studies. We suggest to use the presented machine learning approach in further investigations of atmospheric aerosol chemistry and health effects as well as other studies that require a target-oriented screening of the properties and effects of large classes of substances.

Atomic structures, conformers and thermodynamic properties of 32k atmospheric molecules

“…With climate change accelerating, humanity faces unprecedented social, ecological and economic changes 1 . While data-driven research is emerging in atmospheric science 2 – 5 , open research data is not yet as readily available as in many other fields 6 – 10 . We present our contribution to data-driven atmospheric science in form of the GeckoQ dataset that provides molecular data relevant for aerosol particle growth and formation.…”

“…Krüger et al . 5 trained deep learning models on 103,040 quinones, but did not extend their study beyond this single molecular class. Finally, Isaacman-VanWertz and Aumont 27 studied the p Sat of 182,000 atmospheric species with computationally-efficient group contribution methods, but did not apply more accurate DFT methods.…”

Accelerating models for multiphase chemical kinetics through machine learning with polynomial chaos expansion and neural networks

“…In the atmospheric sciences, NNs are used for air quality prediction, function approximation, and pattern recognition tasks (Gardner and Dorling, 1998), but their application as surrogate models for computationally expensive KMs is less well researched. Recently, popular applications of machine learning in atmospheric chemistry and physics include quantitative structure-activity relationship (QSAR) models that map molecular structures to compound properties as an alternative to time-consuming laboratory experiments or quantum mechanical calculations (Lu et al, 2021;Lumiaro et al, 2021;Galeazzo and Shiraiwa, 2022;Krüger et al, 2022;Xia et al, 2022). Holeňa et al (2010) used surrogate models in computationally costly evolutionary optimization and successfully enhanced this approach with the application of NNs.…”

Machine-Learning-Based Prediction of the Glass Transition Temperature of Organic Compounds Using Experimental Data

“…Within the past decade, the application of machine learning methods in natural sciences has experienced a rapid growth. Most models that were developed in this context are used either for the discovery of new systems , or for the prediction of a specific property. − In 2014, Alzghoul et al published a study in which machine learning algorithms (mainly support vector machines and neural networks) were used to predict the glass transition temperature of organic molecular compounds. One limitation of their work is that only a rather small data set of 71 druglike substances (mostly functionalized heterocycles) was considered.…”