Extracting Structural Information from Physicochemical Property Measurements Using Machine Learning─A New Approach...

Area of Science:

• Environmental Chemistry
• Environmental Health
• Analytical Chemistry
• Computational Chemistry

Background:

• Non-targeted analysis (NTA) is crucial for environmental chemistry and health assessments.
• A major challenge in NTA is the scarcity of analytical standards for environmental chemicals.
• Accurate chemical identification is vital for understanding environmental contamination and health risks.

Purpose of the Study:

• To develop a novel approach for predicting molecular structures using equilibrium partition ratios.
• To integrate measurements of solvent-water partition coefficients (K_SW) with machine learning for chemical identification.
• To explore the use of K_SW as a chemical fingerprint for database searching.

Main Methods:

• Partitioning experiments were conducted using a mixture of 185 chemicals in 10 organic solvents and water.
• Equilibrium partition ratios (K_SW) were measured using liquid chromatography coupled with mass spectrometry (LC-QTOF MS and LC-Orbitrap MS).
• A machine learning algorithm converted K_SW data into molecular functional groups (RDKit fragments) for database searching.

Main Results:

• LC-QTOF MS and LC-Orbitrap MS yielded different log K_SW values, with mean absolute errors of 0.22 and 0.33, respectively.
• Prediction accuracy for RDKit fragments also varied between methods, showing mean absolute errors of 0.23 (QTOF) and 0.31 (Orbitrap).
• The developed approach demonstrated feasibility in predicting functional groups for chemical structure elucidation.

Conclusions:

• This study presents a promising new strategy for structure elucidation in non-targeted analysis.
• The integration of partitioning experiments and machine learning offers a viable alternative for compound identification.
• The method shows potential to overcome the bottleneck of unavailable analytical standards in environmental analysis.