An annotation-overlap-flagged rare-disease gene-prioritisation benchmark and PMC index recipe — methods code and build artefacts

Reproducible foundation for literature-based causal gene prioritisation in rare Mendelian disease. This item is the build recipe and code: a deterministic, version-pinned pipeline for a hybrid dense-plus-sparse retrieval index over the genetics-relevant PMC Open Access corpus (52,777,395 chunks), and for the construction of a stratified 1,047-case evaluation cohort from the GA4GH Phenopacket Store v0.1.26.The cohort is released in two case-paired difficulty variants, each as its own Figshare Dataset: a standard variant with 49 uniformly-random HGNC protein-coding distractors per case, and a hard variant with 49 phenotype-similar distractors selected by HPO Resnik best-match-average similarity. They are produced by scripts/cases/18_build_candidate_lists.py and scripts/cases/18b_build_hard_candidates.py respectively, and share the same cases and the same per-case BLAKE2b seed, so the two variants are deterministic and case-paired. This item is the shared build recipe for both variants and for the index; it is meant to be cited by reference rather than duplicated.Snapshot: git tag paper-methods-v1.4. Earlier tags must not be used — they predate a correction to the reported fusion constant and label Qdrant's built-in reciprocal rank fusion k=60, where the pinned v1.14.1 in fact applies its engine default of k=2 over zero-based ranks.Contents. A single zip, paper-methods-v1.4_code_10fb864.zip, containing the corpus and cohort pipeline code, tests, environment and build configuration, MANIFEST.tsv, REPRODUCE.md and the methods documentation. It also carries the figure renderer (scripts/manuscript/render_p1_figures.py) and the released validation records: - release/index_fingerprint/ — the chunk-set fingerprint, its per-PMCID chunk-count manifest (chunk_counts_by_pmcid.tsv, 2,249,438 rows) and retrieval_substrate_validation.json. - release/cohort/retained_pmcids.txt — the exact indexed article set (2,254,388 PMC identifiers), so the genetics-relevance filter can be audited and a rebuild checked without rebuilding the index. - release/cohort/clustering_stats.json and difficulty_tie_split.json. - release/paper-methods/artifacts_manifest.tsv — every resource with its action (upload / reference / recipe-only) and licence.Per-file checksums are inside the bundle; a zip-level digest is recorded in the repository. Related datasets, built by this code: Standard cohort (random distractors): doi.org/10.6084/m9.figshare.32814449 Hard cohort (phenotype-similar distractors): doi.org/10.6084/m9.figshare.32816468Not included, by design: the 323 GB Qdrant index and the verbatim PMC chunk text, which are recipe-only because per-article licences differ — rebuild from the recipe and verify against the chunk-set fingerprint 70759656…aa39ea in release/index_fingerprint/chunk_id_fingerprint.txt; ontologies, phenopackets and models, which are referenced upstream by pinned version rather than redistributed; and the manuscript itself, which is distributed by the journal and the preprint server, not by this DOI. Licence: AGPL-3.0-or-later for the code. The two cohort datasets above are CC BY 4.0. How to citeAngulo, J., Espinos-Morato, H., & Yeste, V. (2026). An annotation-overlap-flagged rare-disease gene-prioritisation benchmark and PMC index recipe — methods code and build artefacts [Software]. Figshare. https://doi.org/10.6084/m9.figshare.32814491