An automated audit of dataset identifiers, release pointers and AI-coded reuse in TCGA-naming papers, 2024–2025
Isaiah Neal
Abstract
Background. Machine-readable links between papers and datasets inform credit and assessments of data reuse. We measure formal dataset identifiers and release pointers in 2024–2025 open-access PubMed-source papers naming "The Cancer Genome Atlas", and separately estimate presence of TCGA's program-level dbGaP accession, phs000178, among AI-coded 2024 direct reusers. Identifier target roles, paper membership and AI reuse judgements are recorded separately. This is an unvalidated computational audit.
Methods. The Europe PMC frame contained 10,169 open-access papers with full text and source filter SRC:MED. We compared the default "phs000178" query with scanner-observed accession presence in retrieved JATS: accession contexts in enriched strata were read, and the remaining 9,913 records were scanned with every candidate hit adjudicated. A random sample of 100 detector-negative stratum-C papers estimated direct TCGA reuse. A separate Codex pass reuse-coded all 73 accession-positive papers. A full-frame JATS inventory added selected dataset DOI families and release pointers. We separately AI-adjudicated every historically recorded expanded positive paper, classified deposit targets as redistributed inputs, study outputs, mixed or unspecified, and reuse-coded the positive set from located source passages. We also measured identifier presence in CPTAC- and TCIA-naming papers under different definitions. An author AI-assisted full-text check (n=18, existing labels visible) prompted a separate corpus-wide reuse screen; identifier absence was never a valid non-reuse label. Frozen outputs and code form a local reproducibility package.
Results. The new screen flags probable reuse in 7,973 and 9,874 papers, including 7,786 and 9,634 historical identifier-negatives. These are screening results requiring semantic review. The AI-adjudicated formal-identifier union comprised 101/10,169 naming papers in 2024 (0.99%) and 131/12,356 in 2025 (1.06%). The raw expanded sets contained 197 and 253 papers; AI review of all historical positive papers retained 189 (1.86%) and 235 (1.90%), with 1 unresolved 2024 paper and 1 unresolved 2025 paper. The formal unions included 98 and 126 AI-coded direct reusers, respectively. Against scanner-observed accession presence, the default phs000178 query had precision 100% (95% Wilson interval 94.8–100) and recall 95.9% (88.6–98.6). Of 73 accession-positive 2024 papers, 72 were coded as direct reusers. The stratum-C sample had 78/100 AI-coded reusers, yielding an exploratory accession share of 0.83–1.04% of estimated reuse under the stated sampling and stratum-B envelope. Positive-set reuse intersections were measured for both years; a frame-wide reuse denominator was estimated only for 2024. TCIA and CPTAC provide descriptive naming-paper comparisons under different identifier definitions.
Conclusions. The AI-adjudicated historical inventory records formal identifiers from the specified families in about 1% of these naming papers. This is identifier presence in available JATS, not a complete measure of formal citation or identifier capture among all TCGA reusers. The author check had labels visible and supplies no independent validation. Historical identifier extraction depends on proximity to a TCGA mention; the new full-frame reuse screen has no identifier requirement. Automated cross-checks and hypothetical sensitivity analyses examine these dependencies without estimating human-validated accuracy.