MetaboT: Versatile Tools in Metabolomics

• MetaboT is a term applied to diverse computational systems in metabolomics, including an AI-driven natural language interface for knowledge graphs, a MATLAB toolbox for GEM analysis, and a simulator for metabolic computing.
• The natural language interface version employs multi-agent orchestration with explicit schema and entity resolution to convert user queries into precise SPARQL for complex datasets like ENPKG.
• Other incarnations of MetaboT support genome-scale metabolic modeling and gene–metabolite association testing, illustrating its versatility across analytic workflows and formal computational applications.

MetaboT is a name applied in the literature to several computational systems spanning metabolomics knowledge graphs, constraint-based metabolic modeling, and metabolic computing. In one usage, it denotes an AI-based agent for natural language-based interaction with metabolomics knowledge graphs, designed to translate user questions into SPARQL and execute them on the Experimental Natural Products Knowledge Graph (ENPKG) (Bekbergenova et al., 2 Oct 2025). In another, “MetaboTools” is explicitly described as being “often abbreviated in the paper as MetaboT” and refers to a MATLAB-based resource for integrating extracellular metabolomic measurements with genome-scale metabolic models (GEMs) (Aurich et al., 2016). The name also identifies a software tool for translating register-machine programs into positively controlled Metabolic P systems (MPPC) and simulating both representations under discrete-time integer semantics (Guiraldelli et al., 2015). In a related multi-omics context, IntLIM is presented with “practical guidance for ‘MetaboT’,” indicating a setting in which phenotype-specific gene–metabolite association testing can be embedded in a broader metabolomics workflow (Siddiqui et al., 2018).

1. Nomenclature and scope

In the cited literature, “MetaboT” does not denote a single canonical system. It labels distinct tools with different mathematical formalisms, data models, and target tasks.

Usage in the literature	Core function	Source
MetaboT	AI-based agent for natural language interaction with metabolomics knowledge graphs	(Bekbergenova et al., 2 Oct 2025)
MetaboTools, “often abbreviated in the paper as MetaboT”	MATLAB-based toolbox and protocol for GEM analysis from extracellular metabolomics	(Aurich et al., 2016)
MetaboT	Translator, simulator, and comparator for register machines and MPPC systems	(Guiraldelli et al., 2015)
MetaboT context in IntLIM	Host setting for embedding phenotype-specific gene–metabolite association analysis	(Siddiqui et al., 2018)

A recurrent source of confusion is therefore terminological rather than methodological. The knowledge-graph MetaboT is centered on RDF, ontologies, entity resolution, and SPARQL generation. The MetaboTools usage is centered on COBRA-compatible GEMs, extracellular exchange constraints, and phenotype stratification. The metabolic-computing MetaboT is a formal compiler and simulator for MPPC grammars. These systems share a metabolomics or metabolism-related orientation, but they solve different classes of problems and operate under different computational assumptions.

2. MetaboT as a natural-language interface to metabolomics knowledge graphs

The 2025 MetaboT system addresses a practical barrier in mass spectrometry metabolomics: the difficulty of querying semantically rich knowledge graphs without expertise in RDF modeling, SPARQL syntax, and ontology design (Bekbergenova et al., 2 Oct 2025). Its stated goal is to convert natural-language questions into ontology-compliant SPARQL queries, execute them on a metabolomics knowledge graph, and return structured results as CSV with an optional human-readable summary or visualization. The system is demonstrated on ENPKG, which integrates LC–MS/MS features, structural annotations, chemical classes, plant taxonomy, bioactivity, and provenance for a large plant dataset of 1,600 plant extracts.

The central technical problem is that accurate SPARQL requires both valid ontology predicates and correct entity IRIs. LLMs can hallucinate these identifiers and properties if they are not supplied explicitly. MetaboT therefore combines schema injection with explicit entity resolution before query generation. This design makes the system “natural language-based” at the interface level while keeping the executable query layer grounded in the ontology and identifier system of the target graph.

The ENPKG ontology modeled in the paper includes entities such as raw materials, lab extracts, LC–MS analyses, LC–MS feature lists, LC–MS features, MS/MS spectra, SIRIUS annotations, molecular identifiers, NPCClass chemical classes, ChEMBL biological targets, and Wikidata taxa. Representative relationships include ns1:has_wd_id, ns1:has_lab_process, ns1:has_LCMS, ns1:has_lcms_feature_list, ns1:has_lcms_feature, and ns1:has_sirius_annotation, with numerical annotation properties such as ns1:has_zodiac_score and ns1:has_cosmic_score. This ontology-centric design is foundational to MetaboT’s query generation strategy.

3. Multi-agent architecture and query execution

MetaboT is implemented as a multi-agent system built with LangChain and LangGraph, with task decomposition across specialized agents rather than a single monolithic prompting step (Bekbergenova et al., 2 Oct 2025). The Entry Agent classifies a message as a new question or a follow-up and can analyze user-provided files via FileAnalyzer. The Validator Agent checks whether a question is in scope for the knowledge graph and, for plant questions, uses PlantDatabaseChecker to confirm taxon presence in the curated ENPKG plant list. The Supervisor Agent acts as the coordinator, extracting key entities and deciding whether normalization or identifier resolution is needed.

Entity resolution is delegated to the Knowledge Graph Agent, which invokes specialized tools. ChemicalResolver retrieves NPCClass URIs by retrieval-augmented generation over a FAISS index of KG-derived class labels. SMILESResolver converts SMILES to standardized InChIKeys via the GNPS API. TargetResolver maps biological targets to ChEMBL target IRIs through the ChEMBL API. TaxonResolver resolves taxa to Wikidata IRIs through SPARQL queries against the Wikidata endpoint. The paper emphasizes verification against the query context to avoid misresolution.

SPARQL generation and execution are handled by the SPARQL Query Runner Agent. It prepares a structured prompt containing the user question, the ENPKG schema in Turtle, and all resolved identifiers, then invokes GraphSparqlQAChain to generate and execute the query and save results to CSV. When a Wikidata subquery is necessary, as in taxon-to-compound co-annotations, the system uses WikidataStructureSearch and merges outputs with OutputMerger. If the initial query fails, the system performs one iterative refinement step and explicitly distinguishes construction error from true absence of data.

The Interpreter Agent can summarize results, generate visualizations such as bar charts in Plotly JSON, and construct MS/MS spectrum plots through SpectrumPlotter using Metabolomics Spectrum Resolver/USI links. This architecture makes follow-up questions context-aware and supports structured retrieval, narrative interpretation, and plotting within a single conversational workflow.

4. Evaluation, performance, and limitations of the knowledge-graph system

MetaboT was evaluated on 50 natural-language questions spanning ENPKG ontology classes, with domain experts creating a reference SPARQL query for each item (Bekbergenova et al., 2 Oct 2025). The benchmark included 11 low-complexity questions, 19 medium-complexity questions, and 20 high-complexity questions. Three configurations were compared: GPT-4o single-shot with schema provided but no orchestration, MetaboT with GPT-4o mini, and MetaboT with GPT-4o.

The reported overall accuracies were 8.16% for GPT-4o single-shot, 12.24% for MetaboT with GPT-4o mini, and 83.67% for MetaboT with GPT-4o. On the high-complexity subset, GPT-4o single-shot achieved 0%, whereas MetaboT with GPT-4o achieved 78.95%. Average processing times were 10.04 s for GPT-4o single-shot, 36.28 s for MetaboT with GPT-4o mini in cases where all agents ran, and 80.45 s for MetaboT with GPT-4o. Average per-query processing consumed approximately 22,240 tokens, with an average cost of about $0.06 and a reported range from$0.041 to $0.092.

The paper’s error analysis identifies four recurrent failure modes: Validator misclassifying valid questions as invalid; incorrect SPARQL due to wrong properties or malformed syntax; Supervisor failure to invoke the KG Agent when entity resolution was required; and KG Agent selection of the wrong resolver tool. The mitigation strategy consists of schema and IRI injection, explicit resolver tools, one-step refinement, tool-selection guardrails, and directed orchestration logic. Limitations remain: best performance requires high-capacity LLMs; hallucinations in properties or tool routing are mitigated but not eliminated; the system is non-deterministic across runs; it currently supports single-graph querying rather than automatic federated SPARQL across endpoints; and multi-agent orchestration increases prompt size and cost. The public web application, project page, open-source Apache 2.0 repository, Docker and Conda environments, and audited runs via LangSmith position the system as a reproducible research artifact rather than only a prototype.

5. MetaboT as MetaboTools for genome-scale metabolic modeling

In the 2016 protocol paper, MetaboTools is described as being “often abbreviated in the paper as MetaboT” and is defined as a comprehensive, MATLAB-based resource for integrating extracellular metabolomic measurements with GEMs in order to predict intracellular pathway activity and stratify metabolic phenotypes (Aurich et al., 2016). It consists of three tightly connected components: a stepwise protocol, a MATLAB toolbox, and two tutorials for semi-quantitative and quantitative extracellular metabolomics integration. The implementation builds on COBRA Toolbox v2.0, fastFVA, and a linear programming solver, with CPLEX recommended and required for generateCompactExchModel.

The conceptual bridge constructed by this MetaboT is from extracellular “footprints” to intracellular flux states. Its standard constraint-based formulation follows flux balance analysis,

$$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$

with exchange bounds derived from metabolomic data. Flux variability analysis, pFBA-like penalties using the Euclidean norm parameter eucNorm, and uniform random sampling are part of the supported analytical repertoire. The workflow translates extracellular concentration changes into exchange constraints, generates contextualized submodels, and supports downstream analyses such as gene essentiality, flux splits, sampling, perturbation response planes, and model union/intersection summaries.

The protocol is organized into three stages. Stage 1 covers data and model preparation, including metabolite ID mapping through HMDB, KEGG, and ChEBI; ensuring transport and exchange reactions exist; and adding missing transporters or exchanges through quality-controlled curation such as rBioNet. Stage 2 applies constraints, either semi-quantitatively through LOD-aware directionality and slope ratios, or quantitatively by converting concentration changes to exchange fluxes. The general quantitative conversion is

$$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$

Stage 3 performs quality control, validation, and phenotype prediction, including analyzeSingleGeneDeletion, predictFluxSplits, performSampling, and performPPP.

This MetaboT is explicitly designed for large cohorts and phenotype stratification. Tutorial use cases recapitulate prior work on T-cell lines and the NCI-60 panel, including alternative energy-production strategies and TCA/ETC differences. The toolbox can also be used independently of the full integration workflow on any COBRA-compatible model, provided the expected container structures such as ResultsAllCellLines and OverViewResults are assembled.

6. MetaboT in metabolic computing and MPPC universality

The 2015 MetaboT is not a metabolomics analysis platform but a software tool attached to a formal result in metabolic computing: the constructive proof that a positively controlled class of Metabolic P systems, MPPC, is computationally universal and Turing-equivalent to register machines (Guiraldelli et al., 2015). The tool operationalizes this proof by translating register-machine programs into MPPC grammars, simulating both the source program and the synthesized metabolic system, and comparing their halting behavior and outputs.

The formal substrate is the MP grammar $G=(M,R,I,\Phi)$, where metabolites are integer-valued state variables, rules are multiset transformations, $I$ is the initial state, and $\Phi$ contains the regulators or flux functions. An MP system $\mathcal{M}=(G,\tau,\nu,\mu)$ adds timing and unit parameters. Its discrete-time evolution is governed by the Equational Metabolic Algorithm,

$x(t+1)=x(t)+Sv(x(t)),$

where $S$ is the stoichiometric matrix and $v(x(t))$ is the vector of rule fluxes. MPPC adds positivity and availability controls so that fluxes remain nonnegative, consumption never exceeds availability, and integer semantics are preserved.

MetaboT compiles register-machine instructions into metabolic templates. Registers become metabolites $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$0; instruction pointers become metabolites $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$1; each JNZ instruction introduces a comparator metabolite $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$2; and halting is represented by an absorbing HALT metabolite. [INC](https://www.emergentmind.com/topics/in-network-collective-operations-inc)(R_i) is encoded by $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$3 and pointer advance $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$4, both gated by $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$5. [DEC](https://www.emergentmind.com/topics/decision-estimation-coefficient-dec)(R_i) is encoded by $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$6 with MPPC availability control. JNZ(R_i,I_k) uses a four-rule comparator construction with regulators such as $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$7 and $$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$8. Determinism is enforced by the invariant

$$\max \text{ or } \min \; c^T v \quad \text{subject to} \quad Sv = 0,\; \ell \le v \le u,$$9

The tool includes a parser/translator, a simulator/executor, and a comparator/validator. Translation size is linear in program size: INC and DEC contribute two rules each, HALT one rule, and JNZ four rules plus one new comparator metabolite. The paper’s case study for a program computing max(R1,R2) uses 10 instructions with 3 JNZ, 3 INC, 2 DEC, and 2 HALT; the translation yields 19 metabolites and 24 rules, and both the register machine and the MPPC system reach the same final state and halt. This use of MetaboT is therefore a formal methods instrument for metabolic computation, rather than a metabolomics data-analysis environment.

7. MetaboT as a host setting for IntLIM-based multi-omics integration

In the IntLIM paper, “practical guidance for ‘MetaboT’” situates MetaboT as a broader environment into which phenotype-specific integration of transcriptomics and metabolomics can be embedded (Siddiqui et al., 2018). IntLIM itself is an R package with a Shiny interface designed for clinical and translational metabolomics–transcriptomics studies in which both data types are collected at a single time point and many metabolites are unidentified. Its core statistical question is whether the association between a gene and a metabolite differs by phenotype.

For a metabolite $$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$0, gene $$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$1, and sample $$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$2, IntLIM fits

$$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$3

with primary testing focused on $$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$4 versus $$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$5. Interaction p-values are computed for all gene–metabolite pairs and adjusted by Benjamini–Hochberg FDR. Effect size and directionality are summarized through the difference in within-phenotype Spearman correlations,

$$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$6

with a threshold of $$v_{ex,i} = \frac{\Delta C_i \times V_{medium}}{\Delta t \times N_{cells} \times cellWeight}.$$7. Significant pairs are clustered by within-phenotype correlation profiles using hierarchical clustering with Euclidean distance and complete linkage.

The package workflow comprises ReadData(), FilterData(), ShowStats(), PlotDistribution(), PlotPCA(), RunIntLim(), DistPValues(), pvalCorrVolcano(), ProcessResults(), CorrHeatmap(), PlotGMPair(), OutputData(), OutputResults(), and runIntLIMApp(). A matrix-algebra implementation enables fitting 3–7 million models in 2–6 minutes on a laptop with a 2.7 GHz quad-core Intel Core i7 and 16 GB RAM, requiring less than 3% of the time compared to iterating lm() calls. In the NCI-60 application, 3,561,360 pairs were tested and 1,009 pairs satisfied the paper’s thresholds; in breast tumor versus adjacent non-tumor tissue, 6,908,412 pairs were tested and 2,842 pairs were significant.

This relation does not make IntLIM synonymous with MetaboT. Rather, it indicates a modular analytical connection: MetaboT can serve as a broader framework in which pathway-agnostic discovery of phenotype-specific gene–metabolite associations is incorporated alongside other metabolomics capabilities. A plausible implication is that, across the distinct usages of the name, “MetaboT” recurrently denotes software intended to lower technical barriers between complex metabolic data structures and executable analytical workflows, even though the underlying computational paradigms differ substantially.