Crouching AI, hidden connections; The preprint authorsassemble and deploy an artificial intelligence (AI)-based semantics workflow to facilitate data integration across heterogeneous databases.

 

Setting the scene: What are the authors looking at?

 

The authors of this preprint investigated how to enable biomedical data integration across multimodal datasets, diverse model organisms and different biological scales; to evaluate a research question, in practice, knowledge across different databases must be integrated. [1].

 

Databases, including those applied for the biomedical sciences, can have heterogenous schemas and annotation practices; furthermore, they may often have incompatible application programming interfaces (API). These characteristics render interoperability between different biomedical databases difficult, especially in an era when biomedical research requires the integration of these heterogenous databases, for example, interconnection between human Genome Wide Association Study (GWAS) study catalogs, i.e., the GWAS Catalog (NHGRI-EBI GWAS Catalog), and associated mouse research databases, i.e. the International Mouse phenotyping Consortium (IMPC) [1].

 

Bridging the gap: bring a dreaming LLM back to scientific reality

 

Large language models (LLM) can be quite useful in bridging the gap between these varied and fragmented biomedical databases, with their syntactic and semantic heterogeneity, though hallucinations can occur. In this context, the LLM may incorrectly apply ontology identifiers, or alternatively, completely fabricate them; application of appropriate constraints is therefore required [1]. To address this, the flexibility of LLMs can be combined with the structured nature of biomedical ontologies, enabling database integration through semantic alignment, while at the same time, keeping the resulting inferences grounded in scientific reality.

 

The unknowns & the experimental questions

 

The authors studied whether this combination of LLM flexibility with ontology rigidity can effectively automate curation of biomedical databases, resolve semantic proliferation and ground the combination into scientific truth.  Through this study, the authors thus aimed to assemble and deploy an artificial intelligence (AI)–based semantics workflow to facilitate data integration across heterogeneous databases. They used this workflow to:

1. Curate biomedical data across heterogenous databases, either through utilization or de novo creation, if the term does not exist, of ontology terms, via a specialised workflow termed as the Ontology Lookup Service (OLS)
2. Expand LLM embedding practices to include all ontologies, currently present in the OLS

 

The experimental tools

 

To this end, the authors devised a computational workflow, employing multiple large language model (LLM) agents (curator agent, importer agent, ontologist agent, and all coordinated via the coordinating agent), which were in turn grounded in scientifically truthful knowledge via Model Context Protocol (MCP) servers (e.g., OLS MCP server for access to validated ontology terms and identifiers, art1 MCP server for the retrieval and parsing of Europe PMC content).

To generate vector embeddings for all OLS terms, a staggering 11 mil. Terms, for semantic alignment between databases, they employed four generalist models (llama-embed-nemotron-8b, harrier-oss-v1-27b, text-embedding-3-large, text-embedding-3-small); in this manner they calculated cosine similarity across various phenotypes and traits. They also implemented GrEBI (Graphs@EBI, Neo4j KG and MCP server), which makes use of parameterized, Cypher query templates that allows for bridging between human GWAS data and IMPC-derived, mouse ortholog phenotypes [6].

 

How the workflow works: ontologies, embeddings and semantic alignment

Ontologies for Database annotation

 

In the field of bioinformatics and computational biology, an ontology can be defined as a formal, structured framework with standardized vocabulary that characterizes and describes concepts in biological and medical sciences; it establishes specific terms, as well as the relationship between them [2] [3].

Ontologies have also been defined mathematically; the Gene ontology (GO), for example, has been defined as “directed acyclic graphs (DAG)”. In this framework, individual biological concepts/terms are the “nodes”, while the hierarchical relationships between them are the “directional edges” [4].

1. The vocabulary is standardized across different locations, in order to be identifiable by any computer program employing deterministic methods (thus referred to as “machine-readable”); the symbols employed may be symbolic or text-based but always adhere to a rigid set of rules.
2. Should the requirement be that such standardized terms be identifiable by machine-learning (ML) algorithms instead, then this text–based information must be transformed into numerical representations. These representations can include continuous vectors, matrices, or multi–dimensional tensors, which can be then further processed via mathematical or statistical operations. The transformation step can either be tokenization,normalization or embedding.

 

Semantic alignment

 

Through the process of semantic alignment, a machine-readable ontology term (e.g., a string label in the Experimental Factor Ontology [EFO], an ontology developed by the EMBL–EBI) will be processed into a large language model (LLM) embedding (feature vectors), i.e., a dense numerical vector, which can then be readable by machine–learning (ML) algorithms. Using this representation, mathematical operations can be then carried out:

• In this manner, scientists can evaluate how closely related two concepts are; for example, through calculation of the cosine similarity between two embedding vectors, they can align and connect two non-identical terms, across different databases, based on their conceptual proximity [1] [5] [6].

 

Table summarising the key findings and applications of the study

 

Building and grounding the framework

Evaluating performance

Biological applications

Infrastructure enabling data integration 

 

Why this work is interesting

 

The work by McLaughlin et al., 2026, presents a valuable, and useful toolkit, employing multi-agent AI curation, semantic embeddings, and a unified knowledge graph (KG) implementation, through which incompatible biological and medical databases can be coordinated to allow for a unified response to a specific research question [1]. Nowadays, many research projects, and research questions involve the combined use of more than one database, oftentimes spanning different species; a unified layer of ‘translation’ between those databases is sure to make this a lot easier.

 

Questions to the authors:

 

• On training data contamination: In order to prevent the LLM from training on ontology terms generated by the AI-based workflow, you mention that you plan to tag these terms with the editor property ‘AI agent‘. However, given the fact that LLM development regularly involves data scraping (i.e., the automated extraction of information from publicly available data):
  • Is application of an internal annotation property enough, or do you think additional barriers that affect automated data extraction should be implemented, especially given the fact that often, automated processes/scripts that scrape metadata tags may be involved in the processes applied in neural network training?

 

References

 

[1]. McLaughlin, J., Puig-Barbe, A., Ibrahim, A., Pava, D., Pendlington, Z. M., Matentzoglu, N., Sollis, E., Foreman, A., Wilson, R., Gomez, F. L., Harris, L., Adeleye, Y., Kaur, S., Meldal, B., Smedley, D., Parkinson, H. (2026) AI semantics for biomedical data integration [Preprint], BioRχiv, doi:  https://doi.org/10.64898/2026.08.03.742514

[2]. Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Cherry JM, Davis AP, Dolinski K, Dwight SS, Eppig JT, Harris MA, Hill DP, Issel-Tarver L, Kasarskis A, Lewis S, Matese JC, Richardson JE, Ringwald M, Rubin GM, Sherlock G. (2000) Gene ontology: tool for the unification of biology. Nature Genetics. 25(1):25-9. doi: 10.1038/75556.

[3]. Alghamdi, S.M, Schofield, P.N., Hoehndorf, R. (2022) Contribution of model organism phenotypes to the computational identification of human disease genes. Disease Models & Mechanisms. 15 (7): dmm049441. doi:  https://doi.org/10.1242/dmm.049441

[4]. Yang, X., Li, J., Lee, Y., Lussier, Y.A. (2011) GO-Module: functional synthesis and improved interpretation of Gene Ontology patterns, Bioinformatics, 27 (10): 1444–1446, doi: https://doi.org/10.1093/bioinformatics/btr142

[5]. Kulmanov, M., Smaili, F.Z., Gao, X., Hoehndorf, R. (2021) Semantic similarity and machine learning with ontologies, Briefings in Bioinformatics, 22 (4): bbaa199, doi: https://doi.org/10.1093/bib/bbaa199

[6]. Smaili, F. Z., Gao, X., Hoehndorf, R. (2019) OPA2Vec: combining formal and informal content of biomedical ontologies to improve similarity-based prediction, Bioinformatics, 35 (12): 2133–2140, doi: https://doi.org/10.1093/bioinformatics/bty933

 

 

Read preprint 1 Star (No Ratings Yet)