In Silico Structure Predictions for Non-targeted Analysis: From Physicochemical Properties to Molecular Structures

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Biochemistry

Background:

• High-resolution mass spectrometry (HRMS) is crucial for non-targeted analysis (NTA) but identifies less than 5% of detected chemical features.
• Identifying unknown compounds in biological and environmental samples remains a significant challenge in NTA.

Purpose of the Study:

• To develop a computational pipeline integrating HRMS data with physicochemical properties to improve molecular structure elucidation.
• To create a "physicochemical fingerprint" using equilibrium partition ratios (Ksolvent-water) to differentiate isomers.
• To enhance the identification rate of chemical features in NTA.

Main Methods:

• Collected Ksolvent-water values for 129 partitioning systems for compounds in a human blood database.
• Utilized RDKit to compute molecular fragments and bits.
• Developed and trained an artificial neural network using physicochemical fingerprints to predict molecular descriptors.
• Searched a compound database to propose structures for detected chemical features.

Main Results:

• The artificial neural network achieved prediction success rates of 60–86% on the training set and 48–81% on the testing set for chemical structures.
• Physicochemical fingerprints demonstrated effectiveness in distinguishing between isomers.
• The pipeline successfully proposed molecular structures for detected chemical features.

Conclusions:

• Physicochemical fingerprints, when combined with HRMS data, can significantly improve the identification of compounds in non-targeted analysis.
• This computational approach offers a promising strategy to increase the number of identified compounds in complex samples.
• The developed pipeline has the potential to advance the field of metabolomics and environmental analysis.