A Phenotype-Similar Distractor Variant of the Stratified Rare-Disease Cohort for Literature-Based Causal Gene Prioritisation (n=1,047)

A hard variant of the stratified rare-disease benchmark cohort. It contains the same 1,047 cases as the standard cohort (DOI 10.6084/m9.figshare.32814449) — identical case identifiers, phenotype profiles, causal-gene labels and metadata layers — and differs only in the 49 distractor genes per case.Distractors are the genes phenotypically most similar to each case, scored by best-match-average of Resnik term similarities over the HPO gene-to-phenotype annotations, taking the top 49 after excluding the causal gene and any gene annotated to the case's own disease(s). That exclusion means no distractor is a curated alternative cause for the recorded diagnosis. Distractors are treated as negatives relative to the single recorded causal gene; for cases with digenic or oligogenic contributions this labelling may be incomplete.The variant raises difficulty sharply on a common scale: a distractor matches or exceeds the causal gene's phenotypic similarity in 42.8 % of cases, against 7.4 % with random distractors. Because the two variants share cases, seed and ground truth, the difference is attributable to distractor selection alone, giving a difficulty axis orthogonal to the leakage axis.The lists are not neutral across tool classes. Distractors are selected using the same annotation resource from which knowledge-base tools derive their gene-phenotype associations, so results should be read as performance against an HPO-similarity-defined adversary. Standard-versus-hard comparisons are informative within a tool; comparisons between tool classes are more safely made on the standard variant.The build manifest carries a full-file digest and an index-independent core digest; verify against the core digest if you rebuilt from the pinned files alone. 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).