A map rebuilt from words
Semantic embeddings made from expert-written Reactome titles and descriptions reproduced much of the hierarchy's global shape, according to a computational analysis. The study asked whether those text-derived representations could reconstruct Reactome's biological hierarchy. The paper describes global topology as strongly conserved, while the edge-level comparison still registered local structural divergence.
The workflow used Reactome Database Version V95, released in December 2025. It turned titles and descriptions into semantic embeddings, numerical representations intended to capture meaning, with SPECTER2, a sentence transformer pretrained on scientific publication titles and abstracts with contrastive learning on a citation graph. The embeddings were then clustered and used to reconstruct a graph for comparison with the reference Reactome hierarchy.
The reference hierarchy contained 18,732 nodes and 19,613 edges. The nodes comprised 2,848 Pathways and 15,884 Reactions, including all subclasses.
To reflect the branching structure, the authors modified AGNES, a hierarchical clustering method, so it could merge several nodes at once. That produced a multiway hierarchy rather than the standard binary form. The reconstructed clusters were then connected into the Semantic Hierarchy.
The match was strongest in the network's shape
The Semantic Hierarchy matched the reference's node count at 18,732 but contained 18,703 edges, fewer than the 19,613 in Reactome. The difference in individual links was summarized by a Frobenius distance of 216.88, a measure of edge-level structural change. In plain language, the reconstruction preserved the size of the map while drawing some of its connections differently.
At the level of overall graph shape, the paper reports strong agreement. Density was 0.00011 in both graphs, and average degree, the average number of links per node, was 2.00 in the Semantic Hierarchy and 2.09 in the reference. The study's Laplacian Spectral Distance, or LSD, compares global patterns in how a graph is connected. It was 0.085, which the paper describes as high global topological conservation.
To check whether that resemblance could be explained simply by the way links are distributed across nodes, the analysis generated 500 globally rewired versions of the Semantic Hierarchy. These null models kept the degree distribution, or each node's number of connections, while scrambling the broader arrangement. The spectral distance was 0.085 for the reconstructed graph and 0.224 for the fully rewired version. The bootstrap test reported a p-value below 0.005 for the observed similarity in the rewired comparison. No confidence interval was reported for that test.
What the result does not settle
That pattern supports a limited conclusion. Semantic information in the written metadata was sufficient to recover the hierarchy's broad organization in this test, but the graph metrics do not verify that each reconstructed biological relationship is correct. A low global spectral distance can sit alongside local differences, so the result is best read as structural agreement rather than a complete biological reconstruction.
The paper presents the workflow as a preliminary feasibility contribution toward scalable, text-driven pathway reconstruction and curation. Its current evidence is a comparison between the reference hierarchy and a graph reconstructed from text, using global and local graph measures. That leaves open whether biological experts would confirm individual reconstructed relationships or whether the same approach would work in other databases and species.
The document is a preprint, supplied as arXiv version 1 dated 28 August 2026. The authors report that the software code is available at https://baltig.cnr.it/sapiens/semhier. A funding source is not identified in the supplied acknowledgments, and the authors declare no conflict of interest.
Paper data and sources
Original title: Expert Knowledge & Machine Understanding: Bridging Reactome's Ontology with LLM Semantic Embeddings
Authors: Susanna Bravi, Riccardo De Luca, Rosa Sicilia et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text