A reproducibility benchmark of transcriptomic prognostic signatures in pancreatic ductal adenocarcinoma: protocol and feasibility assessment

Protocol and feasibility assessment for a systematic review and independent multicohort benchmark of published transcriptomic prognostic signatures for overall survival in pancreatic ductal adenocarcinoma (PDAC). IMPORTANT: no comparative outcome analysis has been performed. This record documents what was planned and what was found to be feasible. It contains no C-index, no hazard ratio and no pooled estimate, and none exists. Methods. PubMed and Europe PMC were searched from inception. Records were screened against prespecified eligibility criteria, with a recorded reason for every exclusion. Candidate model formulas were recovered from open-access full text. Public cohorts were assessed directly against repository interfaces rather than from abstracts. Gene symbols were resolved against each cohort's own annotation. Results. The search retrieved 1,631 records; screening excluded 1,254 and retained 377 candidates, ranked by the likelihood of a reconstructable formula. Thirteen models with literally printed coefficients were staged, carrying 71 coefficients. Five public cohorts were verified: TCGA-PAAD (175 usable survival records, 100 deaths), GSE21501 (102, 66), GSE57495 (63, 42), GSE62452 (65, 49) and GSE28735 (42, 29), totalling 447 unique patients and 286 deaths. All 112 gene symbols resolve against the GENCODE v36 annotation used by GDC matrices. Feasibility verdict. Against the prespecified stop/go criteria, the cohort (5 of 3 required), patient (447 of 300) and death (286 of 100) thresholds are met. The model threshold is NOT met: thirteen formulas are staged but none has received the two independent human confirmations the protocol requires. That is a governance state, not a data limitation, and it is the single step separating this study from its quantitative phase. The analysis lock remains closed and the production model registry is empty. Methodological finding. Recovering printed formulas mechanically failed in ways that produce plausible but wrong results rather than errors: Unicode minus signs and space-separated signs were dropped, reading seven of fourteen coefficients in one signature as positive when they are negative; adjacent terms ran together into a false gene symbol, losing a gene and turning a 14-gene signature into 13; sentence-final punctuation was absorbed into symbols; per-character markup flattened gene symbols into spaced letters. A cross-check comparing the parsed gene count against the count declared in the title catches the first two classes. Separately, the symbol AP2 is an alias of both FABP4 and GTF3A and cannot be resolved mechanically, so the model using it is blocked from scoring. Preprint. Not peer reviewed. This work was produced with substantial AI assistance (Claude, Anthropic). The author takes full responsibility for the content. The two independent human confirmations the protocol requires have not been performed and remain outstanding.