AI for Metabolomics: Identifying the Unknown Molecules Inside Living Systems

Every living organism contains thousands of small molecules that help regulate biological processes. These molecules, known as metabolites, participate in energy production, cellular communication, immune responses and numerous biochemical reactions.

Scientists can measure many of these molecules using advanced analytical technologies such as mass spectrometry. However, detecting a molecular signal does not always reveal which chemical compound produced it. A biological sample may contain thousands of molecular features, yet many cannot be confidently assigned to known metabolites.

Artificial intelligence is helping researchers address this problem. AI for metabolomics combines machine learning, deep learning and computational chemistry to interpret mass spectrometry data, predict molecular properties and prioritize possible identities for unknown compounds.

These approaches are opening new opportunities in biomarker discovery, precision medicine, drug development and microbiome research. In this article, we explore how AI-based metabolite identification works, why unknown molecular features remain difficult to characterize, and how emerging computational tools are transforming metabolomics research.

What Is Metabolomics?

Metabolomics is the comprehensive study of small molecules, or metabolites, present within cells, tissues, organisms and biological fluids. Unlike genomics, which investigates genetic information, or proteomics, which examines proteins, metabolomics provides information about molecules involved in ongoing biochemical processes.

Metabolites may include:

• Amino acids and derivatives
• Organic acids
• Sugars and carbohydrates
• Lipids and fatty acids
• Nucleotides
• Vitamins and cofactors
• Hormones and signaling molecules
• Microbial metabolites
• Plant secondary metabolites

Changes in metabolite concentrations can reflect biological responses to disease, nutrition, medication, environmental conditions or microbial activity. For example, researchers studying metabolic disorders may investigate differences in circulating amino acids, organic acids and lipid profiles. Similarly, microbiome researchers can investigate molecules produced or modified by gut microorganisms.

Metabolomics provides a way to examine these molecular changes and connect them with biological pathways.

What Is AI Metabolomics?

Traditional metabolomics analysis involves several computational steps, including signal processing, feature detection, database searching and statistical analysis. AI can support these workflows by learning relationships among molecular structures, mass spectral patterns and biological measurements.

Common applications include:

Why Are So Many Metabolites Still Unknown?

Mass spectrometry can detect an enormous variety of molecular signals, but confidently identifying the compounds behind those signals is considerably more difficult. In untargeted metabolomics, researchers investigate a broad range of molecular features without defining every target compound in advance. Many detected features remain unidentified or only partially annotated.

1. Incomplete Reference Databases

Metabolite databases contain information about characterized chemical compounds, but the chemical diversity of biological systems is greater than the reference information currently available. An unknown metabolite may not have an experimentally characterized reference spectrum in the selected database. Even if its chemical structure is recorded, suitable spectral data may be missing, which limits conventional database matching.

2. Similar Molecular Structures

Some metabolites have identical molecular formulas but different chemical structures. These compounds are called structural isomers. Others differ only in their three-dimensional arrangement. Because related molecules can produce similar mass spectra, distinguishing them may require additional experimental evidence.

3. Complex Mass Spectrometry Signals

Biological samples contain many compounds with different abundances and chemical properties. Mass spectrometry signals may include isotopes, adducts, fragments and signals from co-eluting compounds. Consequently, one detected signal does not always represent one unique metabolite. Researchers must first determine which signals are analytically meaningful before attempting molecular identification.

4. Limited Fragmentation Information

Tandem mass spectrometry provides structural clues by measuring how selected ions fragment. However, some compounds produce insufficient or ambiguous fragmentation patterns. A spectrum may support several candidate structures without uniquely identifying one. This is where computational prediction and AI-assisted annotation can help narrow the possibilities.

How Does Mass Spectrometry Identify Metabolites?

Mass spectrometry is one of the most widely used analytical technologies in metabolomics. It measures ions according to their mass-to-charge ratio, commonly written as m/z. When combined with chromatography, it allows researchers to investigate complex molecular mixtures.

LC-MS: Liquid Chromatography–Mass Spectrometry

Liquid chromatography separates compounds based on their interactions with the chromatographic system. Mass spectrometry then detects ions associated with those compounds. An LC-MS analysis may generate information such as:

• Mass-to-charge ratio
• Retention time
• Signal intensity
• Isotope patterns
• Potential adduct information

These measurements help researchers detect and compare molecular features.

MS/MS: Tandem Mass Spectrometry

Tandem mass spectrometry provides additional structural information. A selected precursor ion is fragmented, generating smaller product ions. The resulting MS/MS spectrum contains a pattern of fragment masses and intensities, which can provide clues about the chemical structure of the original compound. For example, molecules belonging to related chemical families may share characteristic fragmentation patterns.

How AI Helps Identify Unknown Metabolites

AI-based approaches are becoming increasingly valuable for interpreting molecular signals that conventional database searches cannot resolve.

1. Predicting Molecular Structures from Mass Spectra

Machine-learning models can learn relationships between molecular structures and MS/MS fragmentation patterns. When given an unknown spectrum, these models may estimate molecular properties or rank candidate structures. Some methods predict molecular fingerprints describing the presence or absence of structural features, helping prioritize chemical structures consistent with the observed spectrum. However, a predicted structure remains a candidate until sufficient supporting evidence is available.

2. Identifying Molecular Similarities

Two compounds with related chemical structures may produce similar fragmentation patterns. Computational methods can compare mass spectra and group related signals. AI models can learn spectral representations that capture more complex relationships than simple peak matching, which may help researchers identify molecular families even when exact reference matches are unavailable.

3. Predicting Chemical Classes

Sometimes researchers cannot establish the exact identity of an unknown metabolite, but they may still determine whether it belongs to a broader chemical category. For example, a computational model might suggest that a compound belongs to a class of lipids, flavonoids or other natural products. This provides useful information for further investigation.

4. Generating Candidate Molecular Structures

Some computational methods explore possible molecular structures using predicted formulas, molecular fingerprints and fragmentation evidence. Generative approaches can propose structures that may explain an observed spectrum, which is especially relevant when a suitable candidate is missing from a reference database. Nevertheless, generated structures must be evaluated against chemical feasibility, available analytical evidence and appropriate experimental measurements.

5. Supporting Metabolite-Based Biomarker Discovery

Machine learning also helps researchers investigate relationships between metabolite measurements and biological conditions. Algorithms can help identify patterns, classify samples and prioritize features associated with a particular outcome, supporting biomarker research. However, identifying a statistically significant metabolomics feature does not establish its identity or prove that it is a clinically useful biomarker.

AI Tools Used in Metabolomics Research

Researchers can use a combination of AI-based platforms, spectral databases and conventional metabolomics software.

These tools perform different functions and should not be considered interchangeable.

SIRIUS and CSI:FingerID

SIRIUS supports molecular formula determination and analysis of tandem mass spectrometry data. Its integrated CSI:FingerID approach uses machine learning to predict molecular fingerprints and search molecular structure databases. Researchers can use these predictions to rank possible molecular structures for further evaluation.

CANOPUS

CANOPUS predicts compound classes from fragmentation-derived molecular information. This is particularly useful when a metabolite lacks a confident database structure match. For example, researchers may obtain a likely chemical class even when the exact compound remains unresolved.

GNPS2

GNPS2 supports molecular networking, which connects spectra according to fragmentation similarity. Related spectra may form molecular families, allowing researchers to explore chemical relationships across a dataset. The platform includes workflows for feature-based molecular networking and downstream interpretation. Molecular networking is a computational approach; it is not automatically equivalent to AI-based molecular identification.

MetaboAnalyst

MetaboAnalyst provides tools for metabolomics data processing, visualization, statistical analysis and biological interpretation. It supports methods such as principal component analysis, heatmaps, statistical testing and machine-learning classification. Researchers can also explore pathway enrichment and biomarker analysis using appropriate metabolomics data.

DreaMS

DreaMS is a deep learning framework developed to learn useful molecular representations directly from large collections of tandem mass spectra. Its research demonstrates how self-supervised learning can support spectral similarity analysis and metabolomics annotation. The approach is particularly interesting because it can learn from large amounts of spectral data without requiring every training spectrum to have a known molecular structure.

Recent Research Advances in AI Metabolomics

Several important developments illustrate how computational metabolite identification is evolving.

1. MassSpecGym: Benchmarking Molecular Identification

A study presented at NeurIPS 2024 introduced MassSpecGym, a benchmark designed to evaluate computational methods for molecular identification from MS/MS data. The benchmark addresses three important challenges:

• Molecular structure generation
• Candidate molecule retrieval
• Mass spectrum prediction

MassSpecGym provides standardized datasets and evaluation methods for studying how well different approaches perform. This is important because AI models should be evaluated on their ability to generalize beyond familiar molecules rather than simply reproduce training examples.

2. DreaMS: Learning from Millions of Mass Spectra

Research published in Nature Biotechnology in 2025 introduced DreaMS, a self-supervised learning approach for small-molecule mass spectrometry. The model was trained using approximately 24 million tandem mass spectra. Instead of relying entirely on manually annotated compounds, the approach learns representations from large collections of spectral data. These representations can support downstream tasks related to spectral relationships and molecular characterization, demonstrating the potential of foundation-model approaches in metabolomics research.

3. Improving Confidence in Metabolite Identification

Recent scientific discussions have emphasized the importance of reporting metabolite identification confidence accurately. AI can help rank candidate identities, but researchers must distinguish between computational annotations and experimentally confirmed compounds. This distinction is particularly important when metabolomics results are used to support biological mechanisms, disease associations or biomarker claims.

Metabolite Identification vs Metabolite Annotation

One of the most important concepts in metabolomics is the difference between an identified metabolite and an annotated metabolite. A software tool may suggest a chemical name based on mass spectral evidence, but this does not automatically confirm the compound’s identity. The original Metabolomics Standards Initiative (MSI) framework distinguishes different levels of identification confidence.

Other confidence frameworks use more detailed levels, but the central principle remains the same.

Researchers should document the evidence supporting each annotation and avoid treating unverified predictions as established discoveries.

A Practical Workflow for AI-Assisted Metabolite Identification

Understanding the complete analytical process is essential for interpreting AI metabolomics results. The following workflow illustrates how researchers can investigate unknown metabolites using computational tools and reference information.

1. 1

   Select a Suitable Metabolomics Dataset

   Begin with a publicly available dataset containing LC-MS or LC-MS/MS measurements. Repositories such as MetaboLights and Metabolomics Workbench provide datasets that can support research and learning. Choose a dataset appropriate for the biological question.

2. 2

   Perform Data Quality Assessment

   Evaluate the available experimental metadata, sample information and quality-control results. Check whether the measurements are suitable for the intended analysis. Poor-quality input data can lead to unreliable downstream interpretations, regardless of the computational method used.

3. 3

   Detect and Align Molecular Features

   Use appropriate data-processing software to identify chromatographic and mass spectral features. Feature detection may involve identifying peaks, aligning retention times and organizing signals across samples. This step creates a feature table for further analysis.

4. 4

   Prioritize Unknown Features

   Examine features that lack confident metabolite annotations. Evaluate their analytical quality and consider whether they may represent isotopes, adducts, fragments or genuine distinct compounds. Features showing interesting biological patterns may be prioritized, provided that quality-control and statistical requirements are satisfied.

5. 5

   Analyze MS/MS Fragmentation Patterns

   Examine available tandem mass spectra and search spectral libraries for potential matches. For unresolved features, computational methods may be used to predict molecular formulas, molecular fingerprints or chemical classes.

6. 6

   Rank Candidate Structures

   Use tools such as SIRIUS and CSI:FingerID to investigate possible molecular identities. Compare available evidence, including formula plausibility, fragmentation behavior and database results. A candidate list should be interpreted as a set of hypotheses rather than confirmed identifications.

7. 7

   Investigate Biological Patterns

   Use statistical analysis and machine-learning methods to compare metabolite features across sample groups, for example between control and experimental samples. Suitable approaches include statistical testing, clustering and supervised classification. Models intended for prediction should be evaluated using appropriate validation procedures that avoid data leakage.

8. 8

   Interpret Biological Pathways

   Once sufficient metabolite identities or annotations are available, researchers can investigate their possible relationships with biological pathways, such as amino acid metabolism, lipid metabolism or energy production. However, pathway conclusions depend on the reliability of metabolite annotations and the methods used.

9. 9

   Validate Important Findings

   Candidate metabolite identities should be confirmed using suitable analytical evidence, including authentic standards when required. Potential biomarkers should be evaluated in independent datasets and appropriate biological studies. Computational predictions are most valuable when they guide reproducible experimental investigation.

Example: Investigating an Unknown Metabolite

Imagine researchers comparing metabolomics profiles from two groups of biological samples. During LC-MS analysis, they detect a molecular feature that differs between the groups. The feature has a measurable mass-to-charge ratio and retention time, but no confident chemical identity. A traditional database search produces several possible matches.

The researchers then examine its MS/MS fragmentation pattern and use computational annotation tools to evaluate candidate structures. An AI-assisted method predicts that the feature may belong to a particular chemical class, and further comparison identifies several candidate compounds consistent with the available spectral evidence.

At this point, the researchers have generated a useful molecular hypothesis, but they have not yet demonstrated which exact compound is present. Additional analytical evidence is required to confirm the identity. This example shows how AI can transform an unresolved molecular feature into a more focused research question, and why AI-assisted metabolite identification should support, rather than replace, experimental validation.

Applications of AI Metabolomics in Biological Research

Challenges and Limitations of AI Metabolomics

Despite progress, several scientific challenges remain.

• Incomplete reference information: many compounds lack suitable experimental reference spectra.
• Structural ambiguity: isomeric compounds may produce highly similar mass spectra.
• Data variability: different instruments, acquisition methods and laboratory conditions can affect spectral characteristics.
• Limited biological generalization: models trained on particular datasets may perform less reliably on unfamiliar compounds or experimental conditions.
• False annotations: high computational scores do not guarantee correct molecular identification.
• Biomarker validation: a feature that separates samples within one dataset may not perform equally well in an independent population.

Addressing these challenges requires reliable analytical data, transparent computational methods and appropriate experimental validation.

The Future of AI in Metabolomics

AI metabolomics research is moving toward more integrated methods for interpreting complex molecular data. Promising developments include:

• Foundation models trained on large mass spectral datasets
• Improved molecular representation learning
• Generative methods for candidate structure prediction
• Better integration of spectral and chemical information
• Explainable AI for molecular annotation
• Multi-omics integration with genomics, transcriptomics and proteomics
• More reliable confidence assessment for metabolite identification

However, the objective is not simply to assign names to more signals. The greater scientific goal is to generate reliable molecular knowledge that improves our understanding of biological systems.

Frequently Asked Questions About AI Metabolomics

Conclusion

AI for metabolomics is creating new opportunities to investigate the unknown molecules present within living systems. By combining mass spectrometry, machine learning, molecular prediction and computational analysis, researchers can explore chemical information that conventional approaches may leave unresolved.

Developments such as DreaMS, CSI:FingerID, CANOPUS and molecular networking demonstrate how modern computation supports more detailed interpretation of complex metabolomics datasets. These advances are relevant to disease biomarker discovery, precision medicine, drug development, microbiome research and plant biotechnology.

However, a computationally predicted identity must not be confused with experimentally confirmed metabolite identification. The future of metabolomics will depend on combining powerful computational methods with reliable experimental measurements and transparent scientific reporting.

For students, researchers and biotechnology professionals, learning how to analyze and interpret metabolomics data is an important step toward understanding the molecular processes that shape living systems.

Scientific References and Further Reading

1. Bushuiev et al. (2025). Self-supervised learning of molecular representations from millions of tandem mass spectra using DreaMS. Nature Biotechnology. Read the research
2. Bushuiev et al. (2024). MassSpecGym: A benchmark for the discovery and identification of molecules. NeurIPS 2024. Read the research
3. MetaboAnalyst 6.0 (2024). Towards a unified platform for metabolomics data processing, analysis and interpretation. Nucleic Acids Research. Read the research
4. SIRIUS 6 Documentation. Computational metabolite annotation, CSI:FingerID, CANOPUS and MSNovelist. Explore documentation
5. GNPS2 Documentation. Feature-based molecular networking for LC-MS/MS metabolomics. Explore documentation
6. What’s in a name? Metabolite identification: challenges and pitfalls in untargeted metabolomics. Metabolomics. Read the research