RaDiCal benchmark data: cohort manifest and accuracy results (v06-01)
收藏资源简介:
This dataset contains the benchmark inputs and outputs supporting the RaDiCal preprint "Phenotype-robust rare-disease variant prioritisation with RaDiCal: 96.0% Top-1 accuracy maintained under 50% HPO term dropout on a 200-case benchmark." Three files are included: (1) COHORT.json v2.0 — the 200-case cohort manifest curated from GA4GH Phenopacket Store v0.1.26 (seed=42), including HPO terms, causative gene, disease identity, and difficulty tier for each case; (2) RESULTS_reranker_am_final.json — ideal-mode benchmark results (Baseline 4, git commit f1c7a81, Exomiser 14.0.0 + AlphaMissense 2-way blend weight 0.30), 192/200 Top-1 accuracy (96.0%, 95% Wilson CI [92.3%, 98.0%]); (3) RESULTS_degraded_50pct.json — degraded-mode results (Baseline 6, 50% HPO term dropout per-case seed hash(case_id) XOR 42, git commit f1c7a81), 192/200 Top-1 accuracy (96.0%, 95% Wilson CI [92.3%, 98.0%]; delta = 0.0 pp vs ideal; McNemar p=1.000). Produced by RaDiCal (Virtual Cognition Pty Ltd, Bass Coast, Victoria, Australia). The production codebase is available to bona fide research and clinical-validation partners under appropriate confidentiality arrangements; see the preprint Code Availability section.



