An interpretable molecular descriptor for machine learning predictions in atmospheric science.

Area of Science:

• Atmospheric Chemistry
• Computational Chemistry
• Machine Learning

Background:

• Machine learning for aerosol formation and chemistry is hindered by inadequate molecular descriptors for atmospheric compounds.
• Existing descriptors struggle with large, highly oxidized organic molecules prevalent in the atmosphere.
• Interpretable models are especially limited by dictionary-based descriptors tied to specific molecular substructures.

Purpose of the Study:

• To introduce ATMOMACCS, a novel interpretable molecular descriptor designed for atmospheric compounds.
• To evaluate the performance of ATMOMACCS against traditional descriptors in predicting key physicochemical properties.

Main Methods:

• Developed ATMOMACCS by combining MACCS fingerprint keys with SIMPOL-inspired motifs.
• Utilized kernel ridge regression models to assess descriptor performance on six atmospheric compound datasets.
• Performed feature analysis to understand descriptor-property relationships.

Main Results:

• ATMOMACCS demonstrated superior performance compared to the RDKit topological fingerprint.
• Achieved significant error reductions in predicting saturation vapor pressures (up to 43%), equilibrium partition coefficients (up to 9%), glass transition temperatures (22%), and enthalpies of vaporization (61%).
• Identified key molecular features influencing different physicochemical properties, highlighting descriptor generalizability.

Conclusions:

• ATMOMACCS is a versatile and interpretable molecular descriptor effective for atmospheric compounds.
• The descriptor enhances the accuracy of machine learning models for predicting atmospheric compound properties.
• Provides insights into structure-property relationships relevant to atmospheric chemistry and physics.