The reconstruction of genome-scale metabolic models (GEMs) is a complex and laborious process that depends significantly on expert manual curation. It involves integrating data from diverse sources, such as biochemical databases and scientific literature, which often contain inconsistencies due to the lack of standardized representations for metabolites and reactions. Since current solutions cannot fully resolve these discrepancies, domain experts have to manually identify and correct them, which is a time-consuming task. This paper proposes an ontology-based approach to streamline the reconstruction of GEMs. This work proposes developing a GEM ontology to formally represent both GEM structures and the expert knowledge used during reconstruction. In the future, this ontology can be integrated into the reconstruction workflow through an application that takes draft models as input and produces enhanced models by incorporating additional information from biochemical datasets. This approach is expected to reduce manual curation efort and consequently simplify the overall reconstruction process.
Metabolism is defined as a series of chemical reactions that occur continuously within living organisms
to sustain life, particularly those associated with energy production and growth [
Over the years, several computational resources have been developed to support the reconstruction of metabolic models, including databases, tools, and ontologies. However, these resources present significant limitations, especially in unifying data from multiple sources [ 2, 3, 4]. Moreover, none of the available ontologies (e.g., ChEBI [5], GO [6], SBO [7]) comprehensively represent GEMs or are actively applied in GEM reconstruction, as they contain little information about the complex relationships between reactions, metabolites, and genes – which are essential for integrating data during GEM reconstruction. Consequently, experts must rely on their domain knowledge to manually resolve data inconsistencies, suggesting that the necessary semantic information exists but is not explicitly represented in current databases and ontologies.
Based on these premises, this work proposes the development of a new ontology for representing GEMs with two main objectives: (1) to enable the integration and reconciliation of models across diferent datasets, and (2) to facilitate the quality assessment of GEMs through the use of logical inferences to identify (and potentially repair) inconsistencies in models. The following sections present the GEM reconstruction process and its main challenges, along with the proposed approach to address them.
GEM reconstruction is a complex process that involves integrating data from diverse sources, conducting thorough literature reviews, performing manual curation, running mathematical simulations, and validating results through biological experiments. Each of these activities need to be carefully conducted in order to ensure the accuracy and quality of the resulting model. Thiele and Palsson [8] proposed a detailed five-stage protocol for the construction high-quality GEMs, summarized in Figure 1. The reconstruction process begins with the creation of a draft model, derived from the organism’s genome and biochemical datasets. This initial model is then refined by experts, supported by computational tools that assess model quality and simulate the organism’s metabolic behavior. Finally, the curated model, along with documentation of the reconstruction process, is compiled and published.
Integrating information on metabolites and reactions is fundamental to GEM reconstruction. Although various eforts have been made over the years to establish representation standards for biochemical data, a universal consensus has not yet been reached [4]. For instance, while the chemical structures of metabolites can be represented as SMILES strings [9], this format does not establish a unique representation for each molecule1 [2] – water, for example, can be represented as [OH2], [H]O[H], or simply O2. The InChI system [10] addresses some of these ambiguities by providing a more detailed representation of molecular structures. However, certain molecules representable in SMILES may lack proper encoding in InChI, and the optional nature of some fields in the InChI format can result in incomplete representations in certain contexts [2]. Beyond variations in metabolite naming, the representation of reactions adds another layer of complexity to the reconciliation of biochemical data. The multiple ways of expressing reactions – ranging from chemical equations [3] to more complex representations such as reaction graphs, hypergraphs, or stoichiometric matrices [11, 12, 13, 14] – hinder data comparison across diferent datasets.
In GEMs, each reaction is associated with genes and enzymes through GPRs (gene-protein-reaction)
rules, along with other attributes essential for simulating organism behavior. SBML remains the
dominant format for representing GEMs, as its design facilitates model exchange, reuse, and supports
simulation-related properties [15]. While its flexible design allows core capabilities to be extended,
it lacks a formal specification for mandatory fields, which results in inconsistently populated fields,
containing incomplete or inappropriate information [16]. Furthermore, even when the same data
sources are used, GEM reconstructions can difer due to variations in methods, algorithms, and expert
decisions throughout the process [
Some approaches, such as MNXref [2] and MetRxn [3], were designed to provide a unified, reliable
data source by implementing iterative reconciliation processes that utilize identifiers, names and
crossreferences to resolve ambiguities. Nevertheless, these methods may overlook crucial factors such
as reaction directionality, compartmentalization, mass- and charge-balance, potentially leading to
1While Isomeric SMILES addresses some ambiguities, multiple representations remain possible [2].
2Although this last representation may seem unusual, it is adopted by ModelSEED (https://modelseed.org/biochem/compounds/
cpd00001) and PubChem (https://pubchem.ncbi.nlm.nih.gov/compound/Water#section=SMILES)
inaccurate results. Additionally, internal and external inconsistencies in names and identifiers, often
found in biochemical databases [
Ontologies have proven to be powerful tools for addressing challenges related to data integration
and system interoperability [
The ontology most directly associated with GEM reconstruction is the Systems Biology Ontology
(SBO) [7], which defines terms related to Systems Biology, including physical entities (e.g., metabolites,
genes, biomass) and processes (e.g., biochemical reactions). SBO terms can be used to annotate tags
in SBML files, facilitating the integration of models originated from diferent sources [
Initially, we conducted a literature review to understand the GEM reconstruction process, existing solutions, and available databases and tools. Subsequently, in collaboration with systems biology experts, the project scope and objectives were established, and competency questions (Table 1) were formulated based on the activities typically performed during GEM reconstruction (e.g., comparing metabolites and reactions, evaluating the metabolic network). Based on these questions we identified the core concepts and relationships of the ontology and designed an initial domain model (Figure 2).
Although this model still requires refinement, it already provides a strong foundation for understanding the domain. The central component of the model is the MetabolicModel class, which serves as an information aggregator, encompassing properties of the organism and a list of metabolites, reactions, genes, and compartments. A BiologicalEntity, corresponds to entities that physically exist in the real world, which contains a unique ID, name, and a list of cross references including external data and metadata (CrossReference class). A Metabolite correspond to any molecule participating in metabolic reactions, either as a reactant, product, cofactor, or intermediate. In GEMs, each metabolite is located in a specific Compartment (a region within the cell, e.g., glucose in the cytosol vs. extracellular glucose are distinct). A Reaction is a biochemical transformation that converts a set of reactants into products. Reactions can be enzymatic (catalyzed by a single enzyme or by an enzymatic complex) or non-enzymatic (occurring spontaneously). An Enzyme is a special type of protein involved in the catalysis of biochemical reactions. A single enzyme may catalyze multiple reactions, and a single reaction may be catalyzed by diferent enzymes (the same applies to enzymatic complexes). Additionally, each enzyme can be encoded by one or more genes. A Gene is a DNA sequence that encodes a protein (in this context, an enzyme). Genes and reactions are linked through GPR associations, which are boolean formulas represented in the diagram through the relationships among the Gene, Enzyme,
In order to fully support concept disambiguation, and promote explainability and systems
interoperability, the current model still requires further enhancements. For instance, several terms mentioned in
the competency questions lack clear and explicit definitions, such as “same” in CQ02, “blocked reaction”
in CQ07, and “required” in CQ12. Therefore, the next step, which we are currently working on, involves
using the initial model as a foundation to create a more explicit model with the necessary
semantic information. Recent studies have demonstrated that the application of ontological unpacking
techniques, combined with the modeling language OntoUML, yields promising results by enriching
concept definitions and making implicit knowledge explicit [
The evaluation of the proposed approach must be twofold: (1) assessing the consistency and accuracy
of the GEM ontology in representing domain knowledge and (2) verifying the quality of the GEMs
generated using this ontology-based approach. For the first part, the representation can be assessed
by encoding the ontology in OWL and using automated reasoners to detect logical inconsistencies
in the knowledge base – for example, defining a reaction as occurring in one compartment while its
metabolites are located in another, or incorrectly declaring two entities equivalent when they have
diferent property values. In addition, competency questions can be translated into SPARQL queries and
used to evaluate whether the ontology can answer the questions proposed by domain experts, thereby
assessing the correctness and completeness of the representation [
The second part of the evaluation can be carried out by comparing the model generated by an
application based on the GEM ontology with manually curated models from well-studied organisms,
such as Escherichia coli (an approach commonly adopted in the literature [
To seamlessly integrate the GEM ontology into the GEM reconstruction workflow, a user-friendly software application must be developed in the future. This application should be able to load a draft model (any valid SBML file) and automatically retrieve additional data from biochemical databases (e.g., BiGG, KEGG, MetaCyc, MetaNetX) and relevant ontologies (e.g., Gene Ontology, ChEBI) to enrich the model and build a standardized representation, while the ontology would provide a formal structure for the data, facilitating data integration and logical consistency verification. The application should also generate a report summarizing reconciliation results – logging automated decisions, describing unresolved issues, providing qualitative and quantitative model evaluations, and presenting additional information to assist domain experts in refining the model, either manually or automatically, depending on the nature of the issues detected. Finally, it should also allow the experts to edit, refine, and subsequently export the model in multiple formats.
The technical details of such an application should be defined during the implementation phase. This includes selecting the most appropriate biochemical databases and ontologies for data enrichment, defining mechanisms for data access (e.g., APIs, SPARQL endpoints), and establishing strategies for data management and storage. In addition, the deployment and documentation of the application should be carefully planned to ensure long-term maintainability, reproducibility, and ease of use for both users and developers.
This work proposed an ontology-based approach to streamline the reconstruction of genome-scale metabolic models (GEMs), comprising both a GEM ontology and an application to integrate it into the reconstruction workflow. The ontology – currently under development – is expected to facilitate data reconciliation across biochemical databases and enable automated reasoning. The application is intended to centralize information retrieval from multiple datasets, thereby reducing manual efort and improving eficiency. Developing both the ontology and the application is complex and time-consuming; therefore, the work can be divided into two phases. The first phase should focus on ontology development and evaluation, while the second should focus on building an application on top of the ontology to generate enhanced models.
In addition, the comparison of GEMs remains an open challenge, as it requires the unambiguous identification of model components such as metabolites and reactions. Defining comparison criteria for metabolic models is expected to provide a novel standard for evaluating the quality of GEMs generated by diferent tools in future studies.
This work was supported by a doctoral scholarship from CAPES/PROEX and CNPq (140323/2025-2), and by funding from Inria ERABLE – University of Lyon, which enabled a research visit to the LBBE (Laboratoire de Biométrie et Biologie Évolutive) in Lyon, France, to foster collaboration between our research teams.
Special thanks to Gabriela Torres Montanaro (ICB-USP), Ariel Mariano Silber (ICB-USP), Lucas Gentil Azevedo (Cidacs/FIOCRUZ, BA), Mariana Galvão Ferrarini (Max Planck Institute for Chemical Ecology in Jena, Germany) and Marie-France Sagot (University of Lyon, France) for their valuable contributions to this work.
During the preparation of this work, the author(s) used ChatGPT (GPT-4o) and Claude Sonnet 4 in order to: Grammar and spelling check, paraphrase and reword. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.