Ontology-linked trial data turns vaccine registries into a searchable clinical asset
Clinical trial registries are rich in detail and poor in consistency. Vaccine studies on ClinicalTrials.gov often use different naming conventions for the same product, and key fields can be distributed across structures that were never designed for cross-trial analysis. This preprint addresses that fragmentation by mapping vaccine-related records into the Study-ExperimentAssay Common Data Model SEA-CDM and then layering biomedical ontologies on top of the registry data.
The work focused on 9,147 vaccine-related study records drawn from the Aggregate Analysis of ClinicalTrials.gov database. Vaccine products were linked to the Vaccine Ontology and reported adverse events were linked to the Ontology of Adverse Events. By placing those data into relational tables and a Neo4j knowledge graph, the investigators preserved links back to the original registrations while also giving the dataset a semantic structure that supports retrieval, ranking, and recommendation.
A registry entry becomes more useful when a machine can recognize that two names point to the same biologic concept.
Query expansion captured more RNA vaccine trials than exact matching alone
The first application tested semantic trial retrieval. Using VO-based query expansion, the system retrieved 41 candidate RNA vaccine trials, compared with 16 trials found by exact VO identifier matching. In practical terms, that difference reflects the limitations of literal keyword searching in a domain where a platform, product family, or antigen-specific construct may be described in more than one way across sponsors and protocols.
Exact identifier matching is precise, but it can miss trials whose descriptions are semantically related rather than textually identical. Ontology-guided expansion widens the search net without abandoning clinical meaning, allowing retrieval to move from a narrow string match toward a concept-level query. For vaccine research, where platform families and product variants often sit under overlapping nomenclature, that broader reach can matter for surveillance, evidence synthesis, and protocol benchmarking.
The authors built the expanded retrieval around the vaccine ontology rather than free-text similarity alone, which is the key methodological move. A query for a specific vaccine concept can then surface records that share biologic family membership even when their exact naming differs across study fields. The result is a more complete inventory of a vaccine class, not merely a cleaner version of keyword search.
Adverse-event ranking improved when trial context entered the model
The second application addressed adverse-event prioritization. A context-aware ranking model combined SEA-CDM features with VO-derived vaccine family information and improved recall@10 from 0.317 to 0.383 over a frequency-based baseline. The gain is modest in absolute terms, but it is meaningful because adverse-event interpretation in vaccine development depends heavily on context: platform, study design, population, and the biologic family of the product all shape what counts as expected, unusual, or clinically important.
Frequency alone tends to surface the most common terms, not necessarily the most informative ones. By incorporating structured trial variables and ontology-derived vaccine families, the model moved closer to the way a clinician or trial methodologist reads safety data—by weighing the exposure context rather than treating every reported event as interchangeable. In a vaccine portfolio, a reaction pattern associated with one family may carry different relevance when seen in another, and the model was designed to preserve that distinction.
This is where semantic integration starts to resemble clinical reasoning. A knowledge graph can connect a vaccine construct, its study setting, and the adverse-event vocabulary in a way that supports ranking by relevance instead of raw prevalence. For safety signal review, that shift can reduce the noise that comes from comparing unlike studies or overvaluing terms that are common but not discriminating.
Embedding models retrieved trial neighbors with high accuracy
The third application evaluated trial similarity. TransE and CompGCN embeddings achieved Hits@10 of 0.950 and 0.946, respectively, for retrieving trials that shared the same VO identifier within disease-specific candidate sets. Those values suggest that the graph representation captured enough semantic and relational structure to find closely related trials with strong accuracy in the tested setting.
In effect, the knowledge graph did more than index records. It encoded relationships among vaccine products, diseases, and study features in a form that embedding algorithms could exploit to identify trials with shared ontology identifiers. For investigators trying to assemble comparable protocols, find prior designs, or cluster studies by biologic class, this kind of retrieval could shorten the path from database to decision.
The candidate-set restriction matters because it reflects a realistic search problem. Trial similarity is most useful when applied within a meaningful disease or indication space rather than across the entire registry. Embedding methods work best when the graph has enough structure to distinguish closely related entities, and here the authors report strong retrieval performance within those disease-specific pools.
Design recommendation points to promise, but class-to-class performance varied
The final application was an ontology-based clinical trial design recommendation model built with several machine learning methods under a leave-one-Material-out evaluation across 23 vaccine classes. Performance varied across classes, which is not surprising in a domain where some vaccine families have abundant precedent and others have sparse or uneven representation in the registry.
A recommendation engine in this setting is only as stable as the historical design space it learns from. If one class has many comparable studies and another has few, the model will naturally perform differently across those groups. The investigators did not present this as a universal automation tool; instead, the results position the framework as a decision-support layer for trial design review, especially where prior studies can be organized by ontology rather than by inconsistent text alone.
The broader implication is methodological rather than therapeutic. Vaccine trials generate large volumes of structured and semi-structured data, but much of that information remains hard to compare across sponsors and disease settings. SEA-CDM and ontology linkage offer a way to standardize the pieces without flattening their clinical meaning.
Public registries gain value when semantic structure is built in
This preprint uses public ClinicalTrials.gov data, draws no individual-level patient data, and keeps the original registry links intact. The authors also make the code and supporting materials available through GitHub and Zenodo, which should make the framework easier to inspect and reproduce.
The central limitation is also the central boundary of the work: these are registry-derived analyses, not patient-level clinical outcomes. The models can organize what trials say about vaccines, adverse events, and design features, but they cannot validate the clinical truth of any specific safety signal or endpoint beyond the source registrations themselves. Performance also varied across the design-recommendation classes, and that variability will need to be understood before any broader operational use.
The next question is whether the same ontology-first pipeline can be extended beyond vaccines to other trial domains where naming inconsistency and fragmented data slow evidence synthesis. The framework is already built to test that idea; the remaining work is showing whether the gains hold in other therapeutic areas with different vocabularies, different design conventions, and less ontological cohesion.