Untargeted LC-HRMS for Pharmaceutical Impurity Screening: Towards Machine Learning-Based Feature Evaluation

Untargeted LC-HRMS for Pharmaceutical Impurity Screening: Towards Machine Learning-Based Feature Evaluation

M. Anzböck 1, M. Kinzig 2, F. Sörgel 2, O. Scherf-Clavel 1

1) Ludwig-Maximilians-Universität München, Department of Pharmacy, Clinical Pharmacy and Pharmacotherapy, Butenandtstraße 5-13, 81377 München

2) IBMP, Institut für Biomedizinische und Pharmazeutische Forschung, Paul-Ehrlich-Straße 19, 90562 Nürnberg-Heroldsberg

 

Background

Drug shortages and complex global supply chains highlight the need for comprehensive pharmaceutical quality control 1, 2. Pharmacopeial methods target predefined impurities 3 but may overlook unexpected contaminants or degradation products. Untargeted liquid chromatography-high resolution mass spectrometry (LC-HRMS) enables simultaneous detection of known and unknown compounds but generates complex datasets requiring transparent feature evaluation 4, 5. This study establishes a transparent, criteria-based LC-HRMS workflow supporting future machine learning (ML)-assisted data interpretation 5, 6.

Methods

Piperacillin was selected as a model compound because of its known β-lactam instability and supply relevance during antibiotic shortages. Chemical Reference Standard solutions were spiked with amino acids as model impurities and retention-time markers and analyzed by LC-HRMS on a quadrupole time-of-flight (qTOF) mass spectrometer in ESI positive/negative mode. Information-dependent acquisition (IDA) MS/MS data were acquired for the pilot dataset; feature-derived inclusion/exclusion lists are being implemented for subsequent measurements to improve MS/MS coverage. Data were processed using openMUCS (open Mass spectrometry-based Untargeted Compound Screening), a Python workflow detecting chromatographically coherent MS1 ion traces, blank matching, quality assessment, MS/MS assignment, neutral-loss evaluation, and scoring.

Results

openMUCS detected 3,461 chromatographically coherent MS1 ion-trace features in the sample dataset. Blank matching classified them as sample-specific (2,683), blank-associated (642), or ambiguous (136). Of the 2,683 sample-specific candidates, 205 were assigned MS/MS spectra, and 128 of these showed at least one predefined neutral-loss match. Combining blank matching with assigned MS/MS spectra prioritized 205 of 3,461 detected sample features for structural review, corresponding to a 94.1% reduction.

Conclusion

openMUCS produced a transparent, ML-ready feature table with traceable classification and scoring criteria, reducing the manually reviewable feature space by 94.1 %. Ongoing work includes iterative IDA acquisition, validation using the spiked retention-time markers, and optimization of scoring thresholds for future ML-assisted screening.

Keywords: pharmaceutical quality control, impurities, LC-HRMS, untargeted screening, machine learning

 

(1) European Medicines Agency. Recommendations to strengthen supply chains of critical medicines; European Medicines Agency, Amsterdam, 2024.

(2) World Health Organization. Quality assurance of pharmaceuticals: A compendium; World Health Organization, 2024.

(3) EDQM. Piperacillin sodium. In European Pharmacopoeia, 11th ed.; Commission, E. P. Ed.; Council of Europe, 2023; pp 5549-5553.

(4) Hollender, J.; Schymanski, E. L.; Ahrens, L.; Alygizakis, N.; Béen, F.; Bijlsma, L.; Brunner, A. M.; Celma, A.; Fildier, A.; Fu, Q.; et al. NORMAN guidance on suspect and non-target screening in environmental monitoring. Environmental Sciences Europe 2023, 35 (1), 75. DOI: 10.1186/s12302-023-00779-4.

(5) Bushuiev, R.; Bushuiev, A.; Samusevich, R.; Brungs, C.; Sivic, J.; Pluskal, T. Self-supervised learning of molecular representations from millions of tandem mass spectra using DreaMS. Nature Biotechnology 2025. DOI: 10.1038/s41587-025-02663-3.

(6) Nguyen, J.; Overstreet, R.; King, E.; Ciesielski, D. Advancing the Prediction of MS/MS Spectra Using Machine Learning. Journal of the American Society for Mass Spectrometry 2024, 35 (10), 2256-2266. DOI: 10.1021/jasms.4c00154.