Identifying translational science through embeddings of controlled vocabularies

Objective: Translational science aims at “translating” basic scientific discoveries into clinical applications. The identification of translational science has practicality such as evaluating the effectiveness of investments made into large programs like the Clinical and Translational Science Awards. Despite several proposed methods that group publications—the primary unit of research output—into some categories, we still lack a quantitative way to place papers onto the full, continuous spectrum from basic research to clinical medicine. Methods: Here we learn vector-representations of controlled vocabularies assigned to MEDLINE papers to obtain a Translational Axis (TA) that points from basic science to clinical medicine. The projected position of a term on the TA, expressed by a continuous quantity, indicates the term’s “appliedness.” The position of a paper, determined by the average location over its terms, quantifies the degree of its “appliedness,” which we term as “level score.” Results: We validate our method by comparing with previous techniques, showing excellent agreement yet uncovering significant variations of scores of papers in previously defined categories. The measure allows us to characterize the standing of journals, disciplines, and the entire biomedical literature along the basic-applied spectrum. Analysis on large-scale citation network reveals two main findings. First, direct citations mainly occurred between papers with similar scores. Second, shortest paths are more likely ended up with a paper closer to the basic end of the spectrum, regardless of where the starting paper is on the spectrum. Conclusions: The proposed method provides a quantitative way to identify translational science.

💡 Research Summary

The paper presents a novel, fully quantitative method for positioning biomedical publications on a continuous spectrum from basic research to clinical application, thereby enabling the systematic identification of translational science. The authors leverage the controlled vocabulary of Medical Subject Headings (MeSH) assigned to MEDLINE articles. First, they construct yearly co‑occurrence matrices of MeSH terms (using a five‑year sliding window) and embed these matrices into a low‑dimensional vector space (5–10 dimensions) using the LINE algorithm (with GloVe as a robustness check).

From the embedded vectors they define a “Translational Axis” (TA) that connects the centroid of “basic” terms (those rooted in the cell, molecular, and animal sub‑trees of the MeSH hierarchy) with the centroid of “applied” terms (those rooted in the human sub‑trees). Each term’s projection onto the TA, measured by cosine similarity, yields a continuous “Level Score” (LS) ranging from –1 (purely basic) to +1 (purely applied). A paper’s LS is simply the average of the LS values of all MeSH terms attached to that paper.

The method is validated against several established categorizations. Clinical‑trial papers, Phase I–IV trial papers, Weber’s seven MeSH‑based categories, and Narin’s four‑level journal classification all show LS distributions that align with expectations, confirming that the continuous scores capture the same broad trends while revealing finer‑grained variation within each category. For instance, papers containing only cell/animal terms have median LS ≈ –0.19, whereas those with only human terms have median LS ≈ +0.48.

Applying LS to the entire MEDLINE corpus (≈15.7 M papers from 1980–2013) yields insightful portraits of journals, disciplines, and the literature as a whole. Classic “basic” journals such as Journal of Biological Chemistry and Cell cluster around negative LS values (≈ –0.2), while clinical powerhouses like NEJM and JAMA sit near +0.5. Multidisciplinary venues (Nature, Science) display broad LS distributions, indicating they publish across the spectrum and suggesting the existence of intermediate “levels” beyond the four traditional categories. Discipline‑level analysis shows Cell Biology, Biochemistry, and Molecular Biology near the basic end, whereas Nursing and Health Services Research lie toward the applied end, with fields such as Immunology, Physiology, and Pharmacology occupying intermediate positions.

Citation‑network analysis uncovers two key patterns. Direct citations preferentially link papers with similar LS values, indicating a homophily of “research level” in citation behavior. Moreover, when tracing shortest paths across the entire network, the terminal nodes of these paths tend to be more basic (lower LS) regardless of the starting paper’s position, reflecting a structural bias toward basic research as a hub in the knowledge flow.

In sum, the authors deliver a scalable, automated framework that quantifies the “basicness” or “appliedness” of any biomedical article using only its MeSH annotations. The Level Score provides a continuous, interpretable metric that can be used by funding agencies, program managers, and scholars to assess the translational impact of research portfolios, monitor the effectiveness of large translational initiatives (e.g., CTSA), and explore the dynamics of knowledge diffusion across the basic‑clinical continuum. The embedding‑based approach also ensures adaptability to emerging terminology and evolving research domains, offering a sustainable solution for long‑term monitoring of translational science.