BioGITOM is an advanced ontology matching (OM) system developed for the biomedical domain, designed to meet the increasing need for precise ontology alignment and data interoperability. By integrating Graph Isomorphism Networks and Graph Transformers, BioGITOM produces enriched concept embeddings that combine both structural and semantic information. This hybrid model enables the system to accurately identify correspondences between concepts across various ontologies, effectively addressing the challenges presented by the complexity and diversity of biomedical data. BioGITOM demonstrated outstanding performance in the Bio-ML benchmark tasks, ranking as the top system in all three tasks and outperforming eight competing methods.
Recent advances in learning methods, particularly those utilizing Deep Learning (DL) and
Graph Neural Networks (GNNs), offer a more powerful means of extracting meaningful entity
representations for OM [
In this paper, we present BioGITOM, a novel OM system specifically designed for the
biomedical domain. BioGITOM enhances concept matching by integrating both semantic and
structural information. It leverages BioBERT to extract semantic features and employs a Graph
Isomorphism Transformer (GIT) model, which combines Graph Isomorphism Networks (GINs)
[
BioGITOM is a specialized OM system developed to address the increasing complexity and heterogeneity in biomedical ontologies. Its core purpose is to ensure effective integration and alignment of disparate ontologies, which is essential for improving data interoperability in biomedical research. BioGITOM is particularly designed to manage the unique challenges posed by the biomedical field, where ontologies often differ in structure and semantics. By combining advanced graph-based techniques, BioGITOM is able to produce more accurate mappings between concepts from different ontologies, thereby supporting enhanced data sharing and collaboration across systems.
BioGITOM leverages a sophisticated set of techniques to deliver high-precision OM by
integrating both structural and semantic features of biomedical concepts. The system employs
a hybrid Graph Neural Network (GNN) model, combining the strengths of Graph Isomorphism
Networks (GINs) and Graph Transformers (GTs) to handle complex biomedical data. Below is a
detailed breakdown of the specific techniques used:
1. Preprocessing: This module prepares the raw ontology data for processing. It reads input
files in OWL (Ontology Web Language) format, creates RDF (Resource Description
Framework) graphs, and extracts concept labels and synonyms. By doing so, it generates
a rich set of terms and relationships for further processing.
2. Concept Name Encoder: BioGITOM leverages BioBERT [
structural embeddings for each concept, capturing both fine-grained and broad context information about how concepts relate to each other in the ontology graph.
Gating Aggregator: The Gating Aggregator is responsible for merging the semantic and
structural embeddings generated by the Concept Name Encoder and GIT, respectively.
This is done through a gated mechanism [
For this evaluation, BioGITOM was applied in its standard configuration without any taskspecific modifications. This approach demonstrates the system's inherent versatility and robustness, as it was capable of achieving high performance without the need for additional customization.
The results underscore BioGITOM's effectiveness and generalizability across different OM tasks within the biomedical domain, highlighting its potential as a reliable tool for diverse applications.
BioGITOM is currently in the development phase and has not yet been released to the public. A public release is planned once the core development is finalized, ensuring that the system is fully functional and ready for broader use in OM tasks.
BioGITOM’s results for OAEI 2024 are summarized in the following sub-sections:
Table 2 shows that BioGITOM performs exceptionally well on the DOID-NCIT dataset, achieving a precision of 0.944 and an F1 score of 0.913. While BioGITOM's recall value (0.884) is slightly lower than the highest recall of 0.959 achieved by LogMapBio, its overall performance remains highly competitive, demonstrating a strong balance between accuracy and recall.
As shown in Table 3, BioGITOM delivers outstanding performance on the SNOMED-FMA (Body) dataset, achieving the highest precision (0.962), recall (0.886), and F1 score (0.923) among all competing methods.
As shown in Table 4, BioGITOM demonstrates strong performance on the SNOMED-NCIT (Pharm) dataset, achieving the highest precision (0.983). However, its recall is relatively lower at 0.713, leading to an F1 score of 0.827. Despite this, BioGITOM secures the second position overall in this dataset, reflecting its high accuracy in producing correct mappings while acknowledging room for improvement in capturing a greater number of relevant matches.
The experimental results highlight the significant advantages of BioGITOM compared to other highly ranked systems. A key strength of our approach lies in the Graph Isomorphism Transformer (GIT) model, which effectively generates contextually relevant representations, enabling the system to handle the complexities of biomedical ontologies. This capability is especially valuable when working with intricate and heterogeneous ontological structures.
However, a limitation of our current approach is its focus solely on generating equivalent mappings. This narrow focus does not fully address other types of semantic relationships, such as subsumption, which may be crucial in certain OM tasks.
To enhance the system’s versatility and performance, we are actively investigating ways to expand the range of matching relationships that BioGITOM can handle, moving beyond simple equivalences to include subsumption, and other relevant relationships.
Additionally, we are exploring the transfer of concept representations into a hyperbolic
integration space. This shift is motivated by the limitations of Euclidean space for hierarchical
ontologies, where distortions can occur. Hyperbolic space is better suited for preserving
hierarchical structures, and we believe that this transformation will significantly improve
BioGITOM’s accuracy and representation of complex ontological relationships [
BioGITOM is a novel approach for biomedical OM that leverages a hybrid Graph Neural Network model, GIT, integrating the strengths of Graph Transformers (GTs) and Graph Isomorphism Networks (GINs).
Experimental results show that BioGITOM consistently outperforms competitive methods across most of the evaluated datasets, underscoring its strong ability to produce highly accurate mappings. However, the system currently focuses exclusively on generating equivalent mappings. To address this limitation, we are actively working on extending BioGITOM to handle a broader range of matching relationships, such as subsumption, which will enhance the system’s versatility and applicability in more complex ontology matching scenarios.
We would like to extend our heartfelt thanks to Jérôme Euzenat for his efforts and support in enabling the submission of our system to the OAEI. His contributions to the initiative are greatly appreciated, and we are grateful for the opportunity to participate.