Reprocessability audit of published human disease bulk RNA-seq: analysis pipeline and result tables

What fraction of published human disease bulk RNA-seq can actually be uniformly reanalysed, and where is it lost? This deposit contains the complete analysis pipeline (R 4.6.1, Bioconductor 3.23) and every result table for a census-scale audit of that question, measured as a cascade from “exists in the literature” to “usably covered by a uniform reprocessing resource” over 3,631 GEO series (from 16,820 retrieved) under 1,124 pre-registered MeSH descriptors. Headline result Of the series whose raw reads are public, 71.0% are not usably reprocessed by any of recount3, ARCHS4 or DEE2. Layer Rate L0 audited population 3,631 series L1a design eligible 24.7% weighted L1b + ≥15 cases / ≥10 controls, counted on sample accessions 7.4% weighted L1c + ≥15 / ≥10, counted on distinct donors 4.1% (conditional, n = 199) L3 raw reads public (dual-route adjudicated) 94.1% L4 present in ≥1 uniform resource 46.7% (bounds 46.7–57.2%) L5 usably covered 27.3% (bounds 27.3–38.8%) Two further results are load-bearing: The unit of counting changes the answer. Among the 199 series where donor identity resolves, 4.1% pass the replication criterion when counted on distinct donors versus 15.1% when counted on sample accessions — a 3.73-fold difference, consistent with the independently measured 2.65× accession-to-donor inflation (Manski bounds 1.09–10.77×; Imbens–Manski 95% CI 1.06–11.88×). A single programmatic route is not sufficient evidence of absence. Of 237 series an Entrez route could not resolve, 24 are contradicted by the primary GEO record, most reading verbatim “Raw data are available in SRA”. How the numbers are constructed Three conventions are applied throughout and are what the deposit is really for: Three-valued outcomes. resolved / not_found / unresolved. An unresolvable lookup is never counted as an absence. Layers summarised as a single rate are reported as bounded intervals with the count of unassessable records stated; the lower bound is the confirmed count. Coverage is a proportion of sequencing runs — the fraction of a series’ own runs present in a resource, bounded in [0, 1] by construction and asserted in the test suite. Reads-absent requires dual-route agreement, and the three absence classes are never pooled: explicitly withheld (28), controlled access (50), and no evidence of deposition (96). A further 39 GEO records carry no “Raw data” status line at all — the submitter made no claim. Contents The L0 attrition from 16,820 retrieved series to the 3,631 audited; the L3/L4/L5 census with three-valued per-resource outcomes and bounded layer rates; the dual-route L3 adjudication; the three-layer L1 table; the accession-to-donor inflation analysis with partial-identification bounds and Imbens–Manski confidence intervals; an EFO/DOID→MeSH crosswalk enabling descriptor-level attribution validation (83.1%, 95% CI 75.7–89.0, which clears the pre-registered 0.80 threshold); sensitivity analyses on the case definition and on the excluded population; PRISMA-S and PRISMA 2020 reporting tables; the full regression suite; and run_all.R, a single command that regenerates every table and figure from cache without contacting any external resource. The deviation log is included in full and deliberately untidied. It runs to 59 entries, 21 of which record defects in this pipeline’s own code, together with the numerical corrections that followed — including two that undercut the study’s own argument before they were found. It is a reviewer-facing audit trail, not an appendix. Provenance Three pre-registration freezes (disease selection, attribution verification, sampling plan) were committed before the work they govern. Their commit, tree and configuration-blob hashes carry OpenTimestamps proofs anchored in Bitcoin blocks 961233 and 961237, in freeze_proofs/. This is an upper bound on when that content existed, from a source with no trusted authority — it does not confirm git’s own committer dates, which are self-asserted and rewritable. tests/test_freeze_integrity.R separately proves from git that no pre-registered parameter changed after any freeze. Licensing This record carries two licences. Source code, tests, configuration and workflow are MIT. Result tables, figures, documentation and the deviation log are CC-BY-4.0. Cryptographic proofs and the public accession list are CC0-1.0, because they are facts. Zenodo’s metadata can list the licences but cannot say which files each covers — LICENSING.md in the deposit root is the authoritative file-by-file mapping. Data and disclosure No third-party dataset is re-hosted. All primary data are publicly available from NCBI GEO, SRA, ENA and the recount3, ARCHS4 and DEE2 compendia; accessions, exact retrieval URLs, retrieval dates and fetch scripts are in data/README.md, with SHA-256 checksums of each large third-party file as retrieved. No new sequence data were generated. No FASTQ was downloaded and no aligner or quantifier was run — this study measures what public resources contain, it does not add to them. Controlled-access data (dbGaP/EGA) is recorded as a classification and never accessed. The analysis code was generated and executed by an AI coding agent (Claude Code, Anthropic; Claude Opus 5) under the author’s direct supervision, with all numerical results independently verified against the archived result tables. Full disclosure, including errors the agent made and corrected, is in docs/AI_USE_STATEMENT.md.