Active Learning Improves Ionization Efficiency Predictions and Quantification in Nontargeted LC/HRMS.

Area of Science:

• Analytical Chemistry
• Computational Chemistry

Background:

• Liquid chromatography electrospray ionization high-resolution mass spectrometry (LC/ESI/HRMS) is crucial for nontargeted screening (NTS).
• Interpreting LC/ESI/HRMS data is challenging due to limited chemical standards and variable chemical responses.
• Machine learning (ML) models can predict ionization efficiency (IE) for chemical quantification but struggle with data outside their training set.

Purpose of the Study:

• To evaluate active learning (AL) strategies for improving IE prediction in ML models.
• To enhance the accuracy of chemical quantification in NTS by optimizing data acquisition.
• To explore methods for expanding training datasets efficiently within a limited labeling budget.

Main Methods:

• Four AL approaches (clustering-based, uncertainty-based, mix, anticlustering) and a random baseline were compared for IE prediction.
• The study focused on acquiring informative data points to enlarge the training set for ML models.
• Quantification accuracy was assessed for natural products in Alpinia officinarum before and after AL-driven training set expansion.

Main Results:

• A significant drop in root-mean-square error (RMSE) for IE prediction (up to 0.3 log units) was observed after a single AL iteration.
• Clustering-based AL showed the least RMSE reduction, while uncertainty-based AL was less practical with larger sample sizes per iteration.
• Expanding the chemical space through AL improved quantification accuracy for natural products from a fold error of 4.13× to 2.94×.

Conclusions:

• Active learning is essential for efficient chemical space exploration and improving ML-based IE prediction.
• The choice of AL strategy impacts the efficiency of data acquisition and prediction accuracy.
• Updating training set chemical space coverage is critical for enhancing quantification accuracy in complex samples.