Probabilistic Mechanism Graphs Resolve the GWAS→Medicine Paradox via Decision‑Grade Calibration: Complete Dataset

This deposit contains the full data and reproducibility artifacts used to develop and validate a probabilistic, mechanism‑aware GWAS→gene prioritization pipeline. It provides raw GWAS summary statistics, regulatory annotations (ABC, PCHi‑C), processed analyses, benchmark definitions, and all scripts needed to reproduce reported calibration and validation results. Contents (high level) Raw GWAS summary statistics (14 files; ~20.9 GB; total variants ~313.9M; genome‑wide significant variants ≈642,451). Regulatory annotations and enhancer→gene data (ABC, PCHi‑C; ~3.8 GB). Processed outputs and reproducibility artifacts: comprehensive_gwas_analysis.json, calibration_metrics.tsv, replication_summary.yaml, MANIFEST.json, checksums and README.md. Validation datasets: Tiered benchmarks including Mendelian/drug targets and Tier‑3 CRISPRi pairs (n = 863). Full analysis and validation code (validation bundle) with notebooks, scripts, and environment files. Key quantitative results (reproducible) Ensemble CRISPR benchmark AUPRC ≈ 0.71 (ensemble), ABC-only 0.65; PCHi‑C +0.06 over baseline. Overall Path‑probability Recall@20 ≈ 76% (benchmarked vs L2G ≈ 58% improvement; disease‑level Recall@20: Alzheimer’s 72%, IBD 70%, Breast Cancer 65%). Expected Calibration Error (ECE) ≈ 0.012 [95% CI 0.009–0.015]; Tier‑3 (CRISPR) ECE ≈ 0.016. eQTL replication: 78% (661/847) with r ≈ 0.89 effect size correlation.All metrics include methodology (bootstrap CIs, locus‑stratified bootstraps) and code in the validation bundle. Methods Fine‑mapping: SuSiE posterior inclusion probabilities. Colocalization: coloc.susie. Enhancer linking: ABC model + PCHi‑C ensemble. Path probability: joint inference across variant→cCRE→gene→tissue layers. Calibration & validation: ECE, bootstrap CIs, independent CRISPR benchmarks, and prospective STING‑seq validation. Limitations & how we mitigated them Cohort and tissue coverage: limited ABC/PCHi‑C coverage reduces recall in tissues (e.g., breast cancer) — mitigation: report disease‑specific results, include per‑locus provenance and encourage tissue‑targeted follow‑up. Label noise & gold‑standard incompleteness: addressed by tiered benchmarking, bootstrap CIs, and explicit sensitivity analyses (provided). Overclaim risk: statements are probabilistic and validated across orthogonal experimental benchmarks; experimental confirmation remains required before therapeutic decisions. Large‑file & delivery constraints: checksums, manifest, curl‑recommended transfer instructions, and stored upload progress are provided for reliable file transfer and integrity checks. How to use Start with README.md and the MANIFEST.json. For large files, use the provided bucket/ftp locations or curl with the checksum verification. Reproduce analyses by running the step-by-step pipeline scripts (requirements listed in requirements.txt / environment.yml). Example notebooks are included. Citation & contact Cite: Ashuraliyev, A. (2025). Mechanism-Aware GWAS Causal Graph Inference: Complete Dataset v5.0. Zenodo. doi: 10.5281/zenodo.17880202 For questions / data access issues: contact the corresponding author (see CITATION.cff and README.md).