A Stratified Rare-Disease Benchmark Cohort for Literature-Based Causal Gene Prioritisation (n=1,047)

A deterministic, reproducible benchmark of 1,047 rare-disease cases for evaluating literature-based causal gene prioritisation. Each case pairs a patient phenotype profile (Human Phenotype Ontology terms) with a 50-gene candidate list — one causal gene and 49 distractors drawn at random from the HGNC protein-coding set — and the causal-gene label, sampled across four operational MONDO-derived disease strata (developmental, immunological, metabolic, neurological). Derived from the GA4GH Phenopacket Store v0.1.26 by a seeded, version-pinned pipeline, so the cohort regenerates exactly from public inputs.Two case-level metadata layers support fairer evaluation against knowledge-base tools: a per-case annotation-overlap flag recording whether a case's source publication is cited in the HPO disease-annotation file phenotype.hpoa for the case's own disease, which defines an overlap-absent subset of 282 cases; and publication-recency strata. The overlap classification is reproducible from public inputs, but it does not guarantee that those cases are free from leakage.The build manifest carries two digests. sha256 covers the full canonical file, including the index-derived pmc_article_count descriptor, and reproduces only against the same retrieval index. sha256_core covers the same records with that descriptor removed and is the digest to verify against if you rebuilt from the pinned files alone. The descriptor enters no metric and no inclusion decision.A case-paired hard variant, with phenotype-similar distractors selected by HPO Resnik best-match-average similarity, is deposited separately at DOI 10.6084/m9.figshare.32816468. The build recipe and validation records are at DOI 10.6084/m9.figshare.32814491.Licence: this dataset is CC BY 4.0. The construction and validation code is deposited separately under AGPL-3.0 (DOI 10.6084/m9.figshare.32814491).