Artificial Intelligence in Mass Spectrometry-Based Metabolomics and Lipidomics for Drug Discovery and Development
4 affiliations
- 1College of Pharmacy, Seoul National University, Seoul 08826 Gwangju 61186 Republic of Korea
- 2Natural Products Research Institute, Seoul National University, Seoul 08826 Gwangju 61186 Republic of Korea
- 3Forensic Toxicology & Chemistry Division, National Forensic Service Daejeon Institute, Daejeon 34054 Gwangju 61186 Republic of Korea
- 4College of Pharmacy and Research Institute of Pharmaceutical Sciences, Chonnam National University Gwangju 61186 Republic of Korea
Abstract
Artificial intelligence applied to mass spectrometry-based metabolomics and lipidomics has advanced through three technological generations: chemometric and classical machine learning, deep learning, and self-supervised foundation models, with large language models so far entering mainly at the edges. The metabolome and lipidome lie closest to drug action, yet they have been among the slowest omics layers to adopt modern artificial intelligence. This paradox arises from their downstream position, which leaves them without a closed molecular reference and without the standardized, readily modeled data available to the genome and transcriptome. This review examines what these methods contribute to drug discovery and development. We survey applications across the drug-development pipeline, from target identification and mechanism of action deconvolution, through the prediction of drug-induced toxicity, to pharmacometabolomics, in which pre-dose metabolic phenotypes inform patient stratification and dosing. Across these metabolomic and lipidomic applications, interpretable classical machine learning remains the dominant form of artificial intelligence. The newer deep-learning and foundation models have concentrated instead on compound annotation, the cross-cutting problem of assigning confident identities to spectra, yet successive generations have mitigated but not solved it. Beyond annotation confidence, batch effects, semi-quantification, and small-cohort overfitting also remain practical constraints. Overall, targeted, hypothesis-driven metabolomic and lipidomic studies can already support drug discovery when their inferences are restricted to confidently identified metabolites, whereas open, hypothesis-free discovery remains the frontier. Near-term progress will depend less on any single algorithm than on curated spectral resources, rigorous measurement, leakage-aware validation, and honest reporting of identification confidence.
Related
- Scientific criteria and guidelines for identifying, reporting and interpreting omics data – transcriptomics, proteomics, metabolomics and lipidomics – in meat science research
- Detecting and Quantifying Variability in Mass Spectrometry-Based Metabolomics Datasets
- Artificial Intelligence for Therapeutic Peptide Discovery and Development
- High-Resolution MRT-TOF Mass Spectrometry Enables Simultaneous Carbon and Nitrogen Tracing in Stable Isotope Resolved Metabolomics
- Trends in Computational Metabolomics: A Perspective on Five Years of Software Development, Challenges, and Opportunities (2021–2025)
BibTeX
@article{yoon2026artificial,
title = {{Artificial Intelligence in Mass Spectrometry-Based Metabolomics and Lipidomics for Drug Discovery and Development}},
author = {Young Cheol Yoon and Seoyeon Choi and Cheol Woon Jung and Hendra Mahakam Putra and Young Jin Min and Sun Jo Kim and Sung Won Kwon},
journal = {Biomolecules \& Therapeutics},
year = {2026},
month = nov,
doi = {10.4062/biomolther.2026.225},
url = {https://doi.org/10.4062/biomolther.2026.225},
note = {CC BY-NC 4.0},
}
RIS
TY - JOUR
TI - Artificial Intelligence in Mass Spectrometry-Based Metabolomics and Lipidomics for Drug Discovery and Development
AU - Young Cheol Yoon
AU - Seoyeon Choi
AU - Cheol Woon Jung
AU - Hendra Mahakam Putra
AU - Young Jin Min
AU - Sun Jo Kim
AU - Sung Won Kwon
PY - 2026
DA - 2026/11/01
JO - Biomolecules & Therapeutics
DO - 10.4062/biomolther.2026.225
UR - https://doi.org/10.4062/biomolther.2026.225
AB - Artificial intelligence applied to mass spectrometry-based metabolomics and lipidomics has advanced through three technological generations: chemometric and classical machine learning, deep learning, and self-supervised foundation models, with large language models so far entering mainly at the edges. The metabolome and lipidome lie closest to drug action, yet they have been among the slowest omics layers to adopt modern artificial intelligence. This paradox arises from their downstream position, which leaves them without a closed molecular reference and without the standardized, readily modeled data available to the genome and transcriptome. This review examines what these methods contribute to drug discovery and development. We survey applications across the drug-development pipeline, from target identification and mechanism of action deconvolution, through the prediction of drug-induced toxicity, to pharmacometabolomics, in which pre-dose metabolic phenotypes inform patient stratification and dosing. Across these metabolomic and lipidomic applications, interpretable classical machine learning remains the dominant form of artificial intelligence. The newer deep-learning and foundation models have concentrated instead on compound annotation, the cross-cutting problem of assigning confident identities to spectra, yet successive generations have mitigated but not solved it. Beyond annotation confidence, batch effects, semi-quantification, and small-cohort overfitting also remain practical constraints. Overall, targeted, hypothesis-driven metabolomic and lipidomic studies can already support drug discovery when their inferences are restricted to confidently identified metabolites, whereas open, hypothesis-free discovery remains the frontier. Near-term progress will depend less on any single algorithm than on curated spectral resources, rigorous measurement, leakage-aware validation, and honest reporting of identification confidence.
SN - 1976-9148
SN - 2005-4483
ER -