MetaboKG: An Analysis-centric Knowledge Graph Framework for Untargeted Metabolomics

This paper introduces MetaboKG, an analysis-centric knowledge graph framework that integrates fragmented untargeted metabolomics data from public repositories and GNPS molecular networks into a unified, traceable, and semantically rich structure to enable reproducible SPARQL exploration and reuse of biochemical knowledge.

The Gist

Imagine the world of untargeted metabolomics as a massive, global library of chemical fingerprints. Scientists use special machines (mass spectrometers) to scan samples from soil, plants, human blood, and oceans, generating billions of data points about the tiny molecules inside them.

The problem? This library is currently a mess.

The Problem: A Library with Scattered Books

Right now, this data is scattered across different warehouses (repositories).

• Warehouse A stores the raw chemical scans.
• Warehouse B stores the lists of what those scans might be (annotations).
• Warehouse C stores the notes on where the sample came from (e.g., "a leaf from a tree in France").

These warehouses speak different languages, use different filing systems, and often don't talk to each other. If a researcher wants to know, "What chemicals are found in French trees, and how confident are we about that?" they have to manually dig through spreadsheets, cross-reference different websites, and try to glue the information together. It's like trying to solve a puzzle where the pieces are in different boxes, and the picture on the box lid is missing.

The Solution: MetaboKG (The Universal Translator)

The authors of this paper built MetaboKG, a new framework that acts like a super-intelligent librarian and a universal translator.

Instead of leaving the data in scattered spreadsheets, MetaboKG takes all these messy files and reorganizes them into a single, giant Knowledge Graph. Think of a Knowledge Graph not as a list, but as a massive, interconnected web of sticky notes. Every piece of information (a chemical, a machine setting, a location) is a note, and the relationships between them are the strings connecting the notes.

Here is how they built it, using three main tools:

1. The "Provenance" Trail (The Receipt)

In the old system, if you found a chemical name, you often didn't know exactly how it was found or which machine made the measurement.
MetaboKG attaches a digital "receipt" to every single piece of data. Using a standard called PROV-O, it tracks the entire journey:

• Where did this sample come from?
• Which machine scanned it?
• Which software analyzed it?
• Who did the work?

This ensures that if you find a result, you can trace it back to its exact origin, just like following a receipt back to the store and the specific cashier who helped you.

2. The "Universal ID" (The Passport)

One of the biggest headaches in science is that the same chemical might be called "Compound X" in one file and "Molecule Y" in another.
MetaboKG introduces a Universal Annotation Identifier (UAI). Think of this as a passport for every chemical finding.

• Even if the data comes from different sources or is added later, this passport links them all together.
• It allows the system to say, "Ah, this result from Study A and that result from Study B are actually talking about the exact same thing, even though they were processed differently."
• This makes it possible to add new data to the library without breaking the connections to old data.

3. The "Semantic Map" (The Dictionary)

The system uses a set of agreed-upon dictionaries (ontologies) to translate everything into a common language.

• If one scientist says "soil" and another says "dirt," MetaboKG knows they mean the same thing in the context of the environment.
• It connects chemical names to biological names (like "human" or "bacteria") and environmental names (like "ocean" or "forest") using a shared vocabulary.

What Can You Do With It? (The "Competency Questions")

The authors tested their new library by asking four tough questions that were previously very hard to answer. Because everything is now connected in the graph, the answers come out instantly via a simple search (SPARQL):

1. The Context Check: "Do the chemicals we found actually match the samples we think they came from?"
   • Result: Yes. The system successfully linked chemical findings back to their specific biological and environmental origins.
2. The Quality Check: "How good are these chemical matches across different machines?"
   • Result: The system can instantly filter results to show only high-confidence matches, regardless of which machine or lab produced them.
3. The Classification Check: "Do different naming systems agree on what a chemical is?"
   • Result: The system can compare different ways of categorizing chemicals (like "family tree" vs. "chemical structure") and see where they overlap.
4. The "Where Else?" Check: "In what other types of samples has this specific chemical been found?"
   • Result: The system can scan the entire global library to tell you, "This chemical has been seen in 50 different studies, mostly in marine environments and human blood."

The Bottom Line

MetaboKG doesn't just store data; it connects it. It turns a chaotic pile of spreadsheets and isolated files into a coherent, searchable web.

By keeping the "receipts" (provenance) and using a "passport" system (Universal IDs), it allows scientists to explore relationships between chemicals, environments, and biological organisms that were previously hidden because the data was too fragmented to see. It's the difference between having a pile of loose puzzle pieces and having the picture on the box, with every piece already snapped into place.

Technical Summary

Technical Summary: MetaboKG

Problem Statement
Untargeted metabolomics generates vast amounts of tandem mass spectrometry (MS/MS) data and computational annotations. While public repositories like GNPS/MassIVE, MetaboLights, and the Metabolomics Workbench have expanded, the field faces a critical fragmentation issue. Data remains scattered across heterogeneous repositories with differing submission logics, file formats, and metadata practices. Although initiatives like Pan-ReDU have improved metadata harmonization, the analytical layer remains disjointed: spectra, features, workflow outputs, annotations, confidence scores, and contextual metadata are often trapped in tabular artifacts (e.g., CSVs) or isolated repository exports. This fragmentation prevents the traceable reuse of annotations, hinders cross-study comparison, and obscures biochemical relationships that require the integration of chemical data with biological and environmental context. Current semantic approaches often focus on curated collections or specific downstream biological interpretations, leaving a gap in the systematic integration of heterogeneous public untargeted metabolomics data with their native analytical outputs and provenance.

Methodology
The authors present MetaboKG, an analysis-centric knowledge graph framework designed to engineer reusable metabolomics knowledge from public repositories. The methodology is built on three coordinated pillars:

1. Semantic Modeling and Ontology Integration:

   • Provenance Core: The framework is grounded in the PROV-O ontology to explicitly model data collection, workflow activities, and agents, ensuring traceability. The authors deliberately avoided SOSA to prioritize data workflow representation over biological transformation processes.
   • Ontology Stack: The model utilizes a restricted but structured set of ontologies: MS (Mass Spectrometry) for domain concepts (extended for metabolomics features), SIO (Semantic Science) for generic relations and n-ary patterns, ChEBI for chemical entities, NCBITaxon for organisms, ENVO for environmental context, and NCIT for sample descriptors.
   • N-ary Relations: To handle complex metadata, the framework employs an n-ary relation pattern using intermediate resources. This allows values to carry semantic context (via rdf:type) rather than being simple literals, reducing graph complexity through deduplication strategies for controlled vocabularies while maintaining distinct provenance trails for each workflow.
   • Metadata Design: Moving away from the full ISA-Tab model due to alignment limitations and sparse metadata coverage, the authors adopted a custom model focusing on sample description, source organism, and experimental protocol, linked via a handcrafted sampling process.

2. Transformation Workflow:

   • Metadata Harmonization: The system ingests sample and study metadata from Pan-ReDU, which aggregates data from GNPS, MetaboLights, and NMDR. A semi-automated workflow maps uncontrolled terms to ontologies (using EBI OLS4) and parses compositional metadata (e.g., extraction solvents) into structured parameters. Organism entities are constructed as composite individuals aggregating eight contextual columns, achieving a 98.4% reduction in tabular redundancy.
   • Analytical Output Extraction: The framework processes 680 GNPS molecular networking results (both MN and FBMN) derived from 2,141 open-access papers. A Large Language Model (LLM) is employed to extract MassIVE dataset IDs and GNPS job identifiers from literature, bridging the gap between repository metadata and analytical outputs.
   • Data Ingestion: CSV outputs from GNPS are harmonized column-by-column using three strategies: direct mapping to typed nodes, literal mapping, and ontology lookup.

3. Universal Annotation Identifier (UAI):

   • To address the challenge of linking heterogeneous outputs, MetaboKG introduces the UAI, an extension of the Universal Spectrum Identifier (USI).
   • The UAI decomposes identifiers into reusable components (collection ID, file, scan, hit, feature). This supports late binding and incremental ingestion, allowing datasets and annotations to be linked post-hoc through shared identifier fragments even if metadata is sparse.
   • The system uses deterministic URIs (GlobalPrefix + human-readable concept + hash) to ensure stability and reproducibility without relying on external registries.

Key Contributions

• Analysis-Centric Framework: A novel approach that shifts focus from isolated repository exports to a unified graph structure where analytical entities, metadata, and provenance are queried jointly.
• Provenance-Preserving Transformation: A workflow that preserves links between repository exports, analytical files, spectra, features, and annotation results, explicitly modeling the "who, what, and how" of data generation.
• Semantic Model for MS/MS: A graph schema that integrates semi-structured MS/MS data with controlled vocabulary terms (ChEBI, NCBITaxon, ENVO) while maintaining explicit provenance via PROV-O.
• UAI Strategy: A mechanism for cross-workflow linkage and incremental data ingestion that overcomes the limitations of monolithic identifiers in fragmented data landscapes.
• Implementation: A functional implementation using OpenLink Virtuoso (chosen over Apache Jena for scalability with ~10 million triples per batch) and a public source code repository.

Results and Evaluation
The framework was evaluated through four Competency Questions (CQs) using SPARQL queries to test its ability to support cross-dataset exploration:

1. Contextual Integration: Successfully linked GNPS molecular annotations to harmonized sample metadata (Pan-ReDU), source organisms (NCBITaxon), and acquisition protocols, demonstrating the ability to traverse from analytical results to biological context.
2. Quality Stratification: Enabled the aggregation and filtering of annotation confidence scores (e.g., modified cosine scores, shared peaks) across different studies and instruments, treating identification descriptors as first-class entities rather than flat table columns.
3. Taxonomic Consistency: Allowed for the comparison of complementary taxonomies (ClassyFire and NPClassifier) across all collections simultaneously, identifying co-occurring chemical classes and biosynthetic pathways.
4. Reverse Metabolomics: Successfully retrieved all public studies where a specific compound (identified by InChIKey) was observed, aggregating sample types and source organisms across the integrated repositories.

The results demonstrate that graph-based integration supports traceable annotation reuse and reproducible SPARQL exploration of biochemical relationships that remain fragmented in repository-native resources.

Significance
The paper claims that MetaboKG addresses a critical gap in metabolomics by providing a FAIR-compliant (Findable, Accessible, Interoperable, Reusable) alternative to ad-hoc scripts and file transformations. It demonstrates that current semantic web languages and ontologies, when suitably combined, can cover the integration needs of semi-structured MS/MS data. By shifting the paradigm from table-based representation to a unified semantic layer, MetaboKG enables the discovery of biochemical relationships that are otherwise hidden when resources are analyzed in isolation. The authors position this as a foundational step toward a more interconnected metabolomics ecosystem, where analytical provenance and contextual metadata are central to data reuse. Future work is identified as adopting LinkML to unify the schema generation process and prevent drift between ontology, RML, and code artifacts.