A Machine Learning and Benchmarking Approach for Molecular Formula Assignment of Ultra High-Resolution Mass...

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Environmental Science

Background:

• Ultra-high resolution mass spectrometry (UHRMS) is vital for analyzing complex mixtures like dissolved organic matter (DOM).
• Accurate molecular formula assignment is challenging yet crucial for interpreting UHRMS data.
• Traditional methods often rely on heuristics and manual tuning, limiting efficiency and adaptability.

Purpose of the Study:

• To develop and evaluate machine learning (ML) models for enhanced molecular formula assignment in UHRMS.
• To compare the performance of ML approaches against traditional methods using curated datasets.
• To provide a benchmark dataset and code for future research in ML-based formula assignment.

Main Methods:

• Application of k-nearest neighbors (KNN) algorithm trained on curated UHRMS datasets of DOM.
• Evaluation of mass accuracy influence (0.15-1 ppm) on model performance.
• Utilized Decision Tree Regressor (DTR) and Random Forest Regressor (RFR) models.

Main Results:

• ML models assigned 43% more formulas than traditional methods (5796 vs 4047).
• Model-Synthetic achieved a 99.9% assignment rate, annotating twice as many formulas (8,268 vs 4047).
• DTR and RFR models demonstrated high formula-level accuracies of 86.5% and 60.4%, respectively.

Conclusions:

• Machine learning approaches significantly increase the number and accuracy of molecular formula assignments from UHRMS data.
• These ML models offer a more robust and efficient alternative to traditional methods for complex mixture analysis.
• The study provides valuable resources (dataset and code) to advance UHRMS data interpretation in environmental science, metabolomics, and petroleomics.