AI model identifies over 500 toxic chemicals in e-liquids, revealing hidden dangers of e-cigarettes

In a recent review published in scientific report, The authors' group used graph convolutional neural networks (NNs) to predict and analyze pyrolysis products in e-liquid flavors and correlated them with mass spectrometry (MS) data to assess potential health risks. did.

study: Predicting the health risks of e-cigarettes by predicting flavor pyrolysis reactions using neural network models. Image credit: iama_sing/Shutterstock.com

background

Inhaling nicotine has long had a negative impact on public health. Considered a safer alternative, vape e-liquids have evolved from simple compositions to include many flavor additives and now often exceed nicotine levels.

This change has particularly appealed to young people, raising concerns about long-term health effects and the re-normalization of nicotine use.

The 2019 outbreak of e-cigarette-related lung injuries linked to additives such as vitamin E acetate highlights 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 within e-liquids when heated and inhaled.

Overview of e-liquid flavor chemicals

This study investigates 180 flavor chemicals identified in e-liquid usage around the world, selected based on existing literature. Analysis of their chemical structures reveals that they are diverse, including 66 esters, 46 ketones/aldehydes, 26 aromatics/heterocycles/carbocycles, 27 alcohols/acetals, and 15 carboxylic acids/amides. functional groups have been revealed. This diversity suggests a wide range of possibilities for different pyrolysis reactions.

Further structural analysis considered properties such as molecular weight and polarity, and 3D chemical space visualization showed a moderate diversity driven primarily by molecular weight, surface area, and rotational flexibility. The average molecular weight is 146.2, indicating that it is a generally volatile chemical.

Workflow for e-liquid flavor risk assessment

Risk assessment of 180 liquid flavors included a workflow that integrated MS data with NN, a prediction of pyrolysis reactions.

Chemical structures were first converted to Simple Molecular Input Line Entry System (SMILES) format. A graph convolution NN model predicted pyrolysis transformations and products and correlated them with MS data detailing molecular ions, fragmentation masses, and their abundances.

Matches of NN-predicted products and MS fragments were further categorized with respect to health risk using the Globally Harmonized System (GHS). This automated process also estimated response activation energies for significant health risks and organized the data into a comprehensive list for each flavor.

Graph convolution NN model for predicting pyrolysis products

Traditional reaction prediction methods have focused on synthetic transformations involving multiple reactants. However, thermally induced pyrolysis reactions typically break down a single reactant into various products.

In this study, the Weisfeiler-Lehman neural network (WL NN) model was employed because it can predict molecular reaction centers and bond changes without the need for pyrolysis-specific training data.

The WL NN was trained using a dataset of 354,937 reactions derived from the US patent literature. This training excluded flavor molecules to prevent data leakage and ensured fair performance when predicting pyrolysis of new flavor molecules.

Implementing the WL NN model

The implementation involves converting the chemical structure into SMILES notation and then into a graphical representation where each atom is labeled with a feature vector. This vector takes into account atomic number, connectivity, valence, and other properties.

WL NN uses local and global feature vectors to predict potential bond breaking changes during pyrolysis. For example, for 2,3-pentanedione, the model identified up to 16 bond cleavage sites and predicted several possible chemical transformations at each site. Products that did not comply with chemical valence regulations were excluded.

Correlation with electron impact mass spectrometry (EI-MS) data

Experimental EI-MS data were used to confirm the NN predictions. EI-MS identifies bond breaks in molecules due to energy bombardment, similar to bond breaks in thermal pyrolysis.

For each of the 180 liquid flavors, molecular weights and fragmentation patterns were obtained from EI-MS data and compared to the pyrolysis products predicted by NN.

A significant number of agreements between NN predictions and MS data confirmed the accuracy of the pyrolysis predictions.

Data integration and health risk analysis

Integrating NN predictions with EI-MS data identified 1,169 matches across 180 flavors, demonstrating strong correlation between predicted and actual pyrolysis products. Ta.

The health risks of these matched products were assessed by obtaining the GHS classification from PubChem.

As a result of the analysis, a significant number of compounds were classified as acute toxins, health hazards, or irritants, indicating various hazards.

Prediction of pyrolysis activation energy

A directed message passing neural network (D-MPNN) was utilized to estimate the activation energy (AE) of pyrolysis reactions, focusing on those producing high-risk health products.

The obtained AE values ​​help understand the thermal conditions required for these reactions and highlight potential health hazards under typical vaping conditions. For example, analysis of acetate esters showed multiple degradation pathways with the formation of acetic acid and substituted alkenes as possible hazards.

Comprehensive report on e-liquid flavors

Detailed data for each flavor was compiled, including NN predicted response, EI-MS compliant product, and GHS classification.

This extensive dataset serves as a valuable reference for understanding the complex chemistry of e-cigarette products and provides the basis for future research and regulatory evaluation.