BioGITOM is an advanced ontology matching (OM) system tailored for the biomedical domain, aiming to achieve precise ontology alignment and foster semantic interoperability. It integrates Sentence-SapBERT for generating domain-specific semantic embeddings from biomedical concept definitions, capturing fine-grained lexical and contextual nuances. These embeddings are further refined through a hybrid structural encoder that combines Graph Isomorphism Networks and Graph Transformers to jointly model local and global dependencies within ontology graphs. To eficiently handle large-scale biomedical ontologies, BioGITOM employs a FAISS-based Approximate Nearest Neighbor (ANN) retrieval strategy for fast and scalable candidate selection. This synergistic design enables the system to robustly manage the complexity and heterogeneity of biomedical data. In the OAEI 2025 Bio-ML benchmark, BioGITOM ranked first in two tasks (NCIT-DOID and OMIM-ORDO) and second in three tasks (SNOMED-NCIT (Pharm), SNOMED-NCIT (Neoplas), and SNOMED-FMA (Body)), achieving the best overall performance with an average F1-score of 0.804, outperforming ten competing systems.
CEUR Workshop
ISSN1613-0073
BioGITOM to capture both local structural regularities and long-range semantic dependencies within ontology graphs.
To eficiently scale to millions of biomedical entities, BioGITOM employs a FAISS-based L2
Approximate Nearest Neighbor (ANN) [
BioGITOM is a specialized OM system developed to ensure semantic interoperability across heterogeneous biomedical knowledge sources. Its main goal is to enable consistent integration and alignment of diverse ontological frameworks, supporting large-scale data sharing and reasoning in biomedical research. By leveraging domain-aware semantic encoding and structural learning, BioGITOM provides an efective solution to the challenges posed by the diversity of biomedical terminologies and ontology structures.
BioGITOM leverages an advanced combination of neural and graph-based techniques to achieve
highprecision matching by jointly exploiting the semantic and structural dimensions of biomedical concepts.
The principal components are summarized below [
For this evaluation, BioGITOM was deployed in its standard configuration, without any task-specific adaptations or parameter tuning. This setup highlights the system’s intrinsic versatility and robustness, as it achieved consistently high performance across benchmarks without additional customization.
BioGITOM is a fully developed OM system, now publicly available for research and benchmarking purposes. The complete source code, pretrained models, and documentation can be accessed through the oficial GitHub repository: https://github.com/lias-laboratory/biogitom.
The evaluation of BioGITOM on the OAEI 2025 Bio-ML track covers five benchmark datasets. The following subsections summarize its performance and comparative analysis.
As reported in Table 1, BioGITOM delivers the best overall results on the OMIM–ORDO dataset. It achieves a precision of 0.845 and a recall of 0.736, leading to an F1-score of 0.787. These figures demonstrate the model’s capacity to generate accurate correspondences while maintaining a high coverage of relevant matches.
Although the OMIM–ORDO dataset has limited structural depth (maximum concept depth = 2), BioGITOM efectively leverages rich semantic representations to overcome this constraint and deliver highly reliable alignments.
On the NCIT–DOID dataset (Table 2), BioGITOM achieves the highest overall performance, ranking ifrst among all participating systems. It attains an F1-score of 0.918, with a precision of 0.924 and a recall of 0.911, reflecting a balanced trade-of between accuracy and completeness. While LogMapLt reaches the highest precision (0.976) and LogMapBio the highest recall (0.959), BioGITOM stands out for its stability across metrics and its ability to align structurally and semantically diverse ontologies with remarkable consistency.
Performance on the SNOMED–FMA dataset (Table 3) highlights BioGITOM’s strong recall capacity. Although BERTMap and BERTMapLt lead in precision (0.970), BioGITOM achieves a precision of 0.829 combined with the highest recall (0.748) among all systems. This combination produces an F1-score of 0.787, confirming the model’s eficiency in capturing a large proportion of valid correspondences across structurally extensive biomedical ontologies.
For the SNOMED–NCIT (Pharm) task (Table 4), BioGITOM displays a consistent balance between precision and recall. While LogMapLt reaches the highest precision (0.994), BioGITOM secures the best recall (0.779) and an F1-score of 0.786. These outcomes emphasize the model’s adaptability in identifying a broad spectrum of correct mappings and its reliability in managing heterogeneous biomedical ontologies.
As displayed in Table 5, BioGITOM achieves competitive results on the SNOMED–NCIT (Neoplas) dataset. With a recall of 0.719, it ranks among the strongest systems in terms of coverage, despite LogMapLt attaining the highest precision (0.931). The resulting F1-score of 0.745 illustrates BioGITOM’s resilience in dealing with semantically intricate and hierarchically layered ontologies, reafirming its reliability across diferent biomedical alignment scenarios.
The experimental results clearly highlight the competitive advantage and robustness of BioGITOM compared to other top-performing systems in the OAEI 2025 Bio-ML track. One of the principal strengths of the proposed approach lies in the Graph Isomorphism Transformer (GIT) architecture, which efectively integrates local neighborhood modeling through Graph Isomorphism Networks (GINs) with global dependency capture via Graph Transformers (GTs). This hybrid design enables the model to generate contextually enriched and structurally coherent representations, allowing BioGITOM to handle the inherent complexity, heterogeneity, and hierarchical diversity of biomedical ontologies.
Another major strength is the Sentence-SapBERT encoder, which leverages domain-specific language modeling on biomedical concept definitions to capture subtle lexical and semantic nuances. The synergy between semantic encoding and graph-based structural reasoning proved crucial for achieving high recall across multiple datasets. Furthermore, the incorporation of a FAISS-based L2 retrieval mechanism ensures scalability and eficiency, allowing BioGITOM to manage millions of biomedical entities without compromising accuracy.
However, a limitation of the current version lies in its focus on equivalence mappings. While this
design ensures strong precision and balanced recall in equivalence-based tasks, it does not yet extend to
other semantic relationships such as subsumption or part-of relations, which are critical in hierarchically
rich biomedical ontologies. Addressing these relationships remains an open challenge for future work,
as also discussed in our recent publication [
Building upon these results, several enhancements are being explored to further increase the versatility and accuracy of BioGITOM. First, we aim to extend the system’s matching scope beyond equivalence to encompass hierarchical (subsumption) and associative relationships, enabling richer and more semantically expressive alignments. This expansion will make BioGITOM suitable for broader ontology integration tasks, such as knowledge graph fusion and reasoning-based biomedical data linking.
Second, we are investigating the transfer of learned concept representations into a hyperbolic
embedding space, motivated by the observation that Euclidean spaces inadequately preserve hierarchical
geometries inherent to biomedical ontologies. Hyperbolic spaces, in contrast, provide a more natural
geometric foundation for modeling tree-like and taxonomic structures, thereby reducing distortion
and improving the representation of complex ontological hierarchies. As reported in our recent
publication [
BioGITOM represents a novel and advanced approach to biomedical ontology matching, built upon a hybrid Graph Isomorphism Transformer (GIT) architecture that integrates the strengths of Graph Isomorphism Networks (GINs) and Graph Transformers (GTs). Supported by Sentence-SapBERT for domain-specific semantic encoding and a FAISS-based retrieval mechanism, BioGITOM efectively combines semantic and structural information to produce precise and scalable ontology alignments.
Experimental evaluations on the OAEI 2025 Bio-ML benchmark demonstrate that BioGITOM consistently outperforms competing systems across multiple biomedical datasets, achieving superior overall performance and confirming its robustness and generalization capability. Nevertheless, the current version focuses primarily on generating equivalence mappings, which limits its ability to capture other semantic relations such as subsumption or part-of.
Future work, will focus on extending BioGITOM to support a broader spectrum of semantic relationships and exploring hyperbolic representation spaces to better model hierarchical ontological structures. These enhancements aim to further improve the system’s versatility, interpretability, and accuracy, reinforcing its role as a comprehensive solution for large-scale biomedical ontology alignment.
During the preparation of this work, the authors used Grammarly in order to grammar and spell check, and improve the text readability. After using the tool, the authors reviewed and edited the content as needed to take full responsibility for the publication’s content. Ontologies with Graph Isomorphism Transformer, The VLDB Journal 34 (2025). doi:10.1007/ s00778-025-00943-7. [10] Y. Gu, X. Qu, Z. Wang, Y. Zheng, B. Huai, N. J. Yuan, Delving Deep into Regularity: A Simple but Efective Method for Chinese Named Entity Recognition, in: M. Carpuat, M.-C. de Marnefe, I. V. Meza Ruiz (Eds.), Findings of the Association for Computational Linguistics: NAACL 2022, Association for Computational Linguistics, Seattle, United States, 2022, pp. 1863–1873. doi:10. 18653/v1/2022.findings-naacl.143.