In a recent review published in the Scientific Reports, a group of authors used a graph-convolutional neural network (NN) to predict and analyze the thermal decomposition products of e-liquid flavors, correlating them with mass spectrometry (MS) data to assess potential health risks.
Study: Forecasting vaping health risks through neural network model prediction of flavour pyrolysis reactions. Image Credit: iama_sing/Shutterstock.com
Background
Nicotine inhalation has long been detrimental to public health. Vaping e-liquids, seen as a safer alternative, have evolved from simple compositions to include numerous flavor additives, which now often exceed nicotine levels.
This shift has particularly appealed to younger demographics, raising concerns about long-term health impacts and the re-normalization of nicotine use.
The 2019 outbreak of vaping-related lung injuries, linked to additives like vitamin E acetate, underscores the potential risks of inhaling chemically complex e-liquids.
Further research is needed to fully understand the long-term health effects of the complex chemical interactions in e-liquids when heated and inhaled.
Overview of e-liquid flavor chemicals
This study investigates 180 flavor chemicals identified in global e-liquid use, selected based on existing literature. Analysis of their chemical structures revealed a diverse array of functional groups, such as 66 esters, 46 ketones/aldehydes, 26 aromatic compounds/heterocycles/carbocycles, 27 alcohols/acetals, and 15 carboxylic acids/amides. This diversity suggests a broad potential for varied pyrolysis reactions.
Further structural analysis considered properties like molecular weight and polarity, with a 3D chemical space visualization indicating moderate diversity, primarily driven by molecular weight, surface area, and rotational flexibility. The average molecular weight was 146.2, pointing to a generally volatile set of chemicals.
Workflow for e-liquid flavor risk assessment
The risk assessment for 180 e-liquid flavors involved a workflow that integrated NN: predictions of pyrolysis reactions with experimental MS data.
The chemical structures were initially converted into a simplified molecular-input line-entry system (SMILES) format. A graph-convolutional NN model predicted the pyrolysis transformations and products, then correlated with MS data that detailed molecular ions, fragmentation masses, and their abundances.
Matches between NN-predicted products and MS fragments were further classified for health risks using the Globally Harmonized System (GHS). This automated process also estimated reaction activation energies for significant health risks, organizing the data into a comprehensive list for each flavor.
Graph-convolutional NN model for predicting pyrolysis products
Traditional reaction prediction methods have focused on synthetic transformations involving multiple reactants. However, pyrolysis reactions, driven by heat, typically involve a single reactant breaking down into various products.
For this study, the Weisfeiler–Lehman neural network (W–L NN) model was adopted due to its ability to predict reaction centers and bond changes in molecules without requiring pyrolysis-specific training data.
The W–L NN was trained using a dataset of 354,937 reactions derived from United States (US) patent literature. This training excluded flavor molecules to prevent data leakage, ensuring unbiased performance in predicting new flavor molecules’ pyrolysis.
Implementation of the W–L NN model
The implementation involves converting chemical structures into SMILES notation, then into graph representations where each atom is labeled with a feature vector. This vector accounts for atomic number, connectivity, valence, and other properties.
The W–L NN uses local and global feature vectors to predict potential bond-breaking changes during pyrolysis. For example, in the case of 2,3-pentanedione, the model identified up to 16 probable bond-breaking sites, predicting several feasible chemical transformations for each site. Products that did not comply with chemical valence rules were excluded.
Correlation with experimental electron-impact mass spectrometry (EI-MS) data
Experimental EI-MS data was used to confirm the NN predictions. EI-MS identifies bond-breaking in molecules due to energy impact, analogous to bond-breaking in pyrolysis due to heat.
For each of the 180 e-liquid flavors, EI-MS data provided the molecular weight and fragmentation patterns, which were then compared to the NN-predicted pyrolysis products.
A significant number of matches between the NN predictions and the MS data confirmed the accuracy of the pyrolysis predictions.
Data amalgamation and health risk analysis
The amalgamation of NN predictions and EI-MS data identified 1,169 matches across the 180 flavors, indicating a robust correlation between predicted and actual pyrolysis products.
The health risks of these matched products were then assessed by obtaining GHS classifications from PubChem.
The analysis showed a variety of hazards, with a substantial number of compounds classified as acute toxins, health hazards, or irritants.
Prediction of pyrolysis activation energies
A directed message-passing neural network (D-MPNN) was utilized to estimate activation energies (AEs) for pyrolysis reactions, focusing on those producing high-risk health products.
The derived AE values help understand the thermal conditions required for these reactions, highlighting potential health hazards under typical vaping conditions. For example, the analysis of acetate esters indicated multiple degradation pathways, with the formation of acetic acid and substituted alkenes as likely hazards.
Comprehensive reporting for e-liquid flavors
Each flavor’s detailed data, including NN predicted reactions, EI-MS matched products, and their GHS classifications, was compiled.
This extensive dataset serves as a valuable reference for understanding the complex chemistry of vaping products and forms a basis for future research and regulatory assessment.