The research behind the atlas

Coverage & study windows

The atlas covers small-rodent acute-injury therapy studies published between 1980 and 2020. Its 5,033 included publications combine 130 from 1980–1989, 1,257 from 1990–2010 and 3,646 from 2011–2020.

Individual study detail is partial for 1990–2010. A publication can be associated with more than one therapy; the atlas retains all 4,098 study associations. Historical clinical-trial and regulatory notes can extend beyond the research window. This dataset was generated on 25 August 2026.

Definitions that matter

• Resolved drugs: records with a ChEMBL identifier (635). The atlas uses this identifier-based definition throughout.
• Trauma drugs: resolved drugs flagged as studied in physical-trauma models (333). The broader corpus also covers chemical/toxin injury, ischaemia and radiation.
• Any human trial: derived from the any-trial flag, for any indication. Trial registration is not proof of benefit.
• Stringent acute-injury trial: aggregate counts of therapies with injury-indication trials, reported separately from record-level injury-link flags. Both answer useful but different questions.
• Approval: strict trauma approval, caveated trauma use and non-drug/device classifications remain separate. ChEMBL maximum phase is a development indicator, not regulatory approval.
• Causality: listing a therapy in the corpus does not demonstrate that these studies produced its approval. The atlas does not infer a number of new therapies caused by the research.
• Safety: editorial annotations organise reported narratives. Historical removals, ambiguous findings and narratives without positive signals are separated.

Counts & denominators

Publication counts, therapy records and trial definitions are related but not interchangeable. The table below shows how common measures relate.

10 consequential classifications do not yet list attached primary citations. 3 publications have conflicting title or year metadata across therapy associations; both extractions remain available for inspection.

Eligibility criteria

Studies were eligible if they (1) used a small animal model (including rat, mouse, guinea pig, hamster, rabbit etc.) of acute injury; (2) administered a therapeutic intervention intended to treat the injury; (3) reported an injury-related outcome versus a comparator; (4) were published 1980-2020; and (5) were original research in English. Studies were excluded if they were fundamental biology without an intervention, reviews or editorials, lacked a comparator, or (for the 1990-2010 review) tested therapies already in routine clinical use at publication. Acute injury was defined as tissue damage over hours to days rather than chronic or degenerative disease. For the 1980–1989 and 2011–2020 tranche, a scope ruling placed physical trauma, chemical or toxin-induced injury and bacterial endotoxin (LPS) in scope, and placed live infection or sepsis without a toxin trigger, and ex vivo or isolated-organ preparations, out of scope.

Information sources and search

MEDLINE via PubMed was searched with Boolean combinations across acute-injury, therapy, and rodent-model term sets, with PubMed Automatic Term Mapping active. A single database is a limitation; reference-list screening and an Embase sensitivity check for exemplar therapies identified no additional eligible studies. The complete search strings are in the accompanying supplementary files.

Study selection

For 1990-2010, five reviewers screened records in two phases (title/abstract, then full text), each record assessed by at least two reviewers with group-consensus voting; an independent large-language-model classification of a stratified 200-study sample (100 included, 100 excluded; seed 42) provided cross-validation, with human-AI agreement quantified by Cohen's kappa and discordant records re-reviewed. For 1980-1989 and 2011-2020, screening was automated: each record was classified by three independent blind passes of a large language model against the same criteria, decisions were auto-settled only on unanimity, non-unanimous records were escalated, and an objective PubMed-metadata cross-check (article type, species, retraction) was applied. Deduplicated study-level decisions were recorded throughout.

Therapy grouping and clinical mapping

Studies of the same agent were grouped by active ingredient or biological entity (not formulation, dose or route), resolved where possible to a ChEMBL identifier, which also supplied synonyms, mechanism and target. Human translation was established from ClinicalTrials.gov registrations, regulatory records and the peer-reviewed literature, not from preclinical vote-counting. Combination therapies were classified by the novel component. Devices and physical interventions were classified separately and excluded from drug denominators.

Risk of bias and synthesis

Risk of bias in included animal studies was framed against the SYRCLE domains (selection, performance, detection and attrition); consistent with the wider preclinical literature, reporting of randomisation, allocation concealment and blinding is sparse. Formal per-study SYRCLE scoring is provided in the accompanying publication rather than in this interactive tool. Translation proportions are presented with 95% Wilson score confidence intervals and under multiple denominators so readers can see how conclusions depend on definitional choices.

Limitations

• The review used a single bibliographic database. Publication bias in the preclinical literature means observed translation rates should be read as an upper bound, not a proven ceiling.
• Species, model, timing, route, dose and outcome fields are machine-extracted or metadata-derived. Missing values remain missing, and distinct therapy-level extractions are preserved.
• Human screening covers 1990–2010; the extended windows used automated screening with validation steps described in the methodology above.
• Some classifications and safety descriptions lack direct primary citations in the record. Names and prose references are displayed as recorded; URLs are clickable only when an actual URL is present.
• Clinical failure, a safety warning and absence of an identified trial are distinct findings. None alone establishes that repeating a mechanism is futile or that an untested agent is a promising candidate.
• This dataset is not live-refreshed from PubMed, ChEMBL, ClinicalTrials.gov, FDA or EMA. Regulatory status and warnings may have changed since August 2026.

Open to correction

Send anomalies, missing studies or proposed corrections to [email protected], including the record identifier, dataset date and supporting sources. Each dossier provides a prefilled correction link.