A closer look at how GWAS Harmonizer handles allele orientation, statistical contradictions and unresolved identifiers.
Five scenarios drawn from 100,000 selected non-palindromic, biallelic SNPs in GCST90018642. Repeated variants across scenarios are counted as scenario-row evaluations.
scenario-row evaluations with exact recovery of the scored coordinates, alleles and statistical fields.One source dataset. This does not establish accuracy for indels, palindromes or liftover.
Illustrative allele examples explain each operation; they are not extracted benchmark rows. Measured results come from the controlled perturbation report.
Deterministic selection from 9,834,126 eligible records, seed 20260816. Target GRCh38; strict submission mode; reference alignment and conservative identifier handling. All five recorded runs were processing-ready but not submission-ready. The original systematic bundles predated allele-aware rsID validation, and ambiguous confidence-interval blockers occurred despite no CI columns. Exact field recovery is separate from submission readiness.
Source SHA-256: 7f0e747b719a54dbfc81ba07ecbe7568962687b7ff8be06def06f79f266bc2e4
Injected p-value contradictions detected. No flags among 95,000 non-injected mixed-file rows.
Detection sensitivity: 95% Wilson CI 99.923–100%. Rows and reported p-values were retained; submission readiness was blocked.Of 5,000 blank rsIDs: 4,002 correct assignments, 998 unresolved, zero wrong assignments.
Assignment precision 100%; recall 80.04%. One merged canonical-equivalent rsID counted as correct.September 30 assays add 256 GWAS Catalog SNPs and 1,024 FinnGen SNPs outside the specified earlier subsets. Small chromosome-1 assays; no speed or general-accuracy ranking.
| Assay | Harmonizer | GWASLab 4.2.1 | tidyGWAS 1.0.0 |
|---|---|---|---|
| GCST90018642 | 256 / 256 | 256 / 256 | 256 / 256 |
| FinnGen R13 T2D | 1,024 / 1,024 | 1,024 / 1,024 | 1,024 / 1,024 |
All three tools retained all 1,280 records with equivalent alleles and beta/SE/p. GWASLab EAF differences agree with float32 rounding (maximum absolute difference <4.3 × 10⁻⁸). tidyGWAS preserved valid alternative effect orientations. Neither is treated as a scientific error.
Harmonizer original/final fields verified, including 960 transformations.
tidyGWAS also retained raw records and row IDs for reconstruction. Traceability is not unique to Harmonizer; its explicit action records accompany the browser download.Prespecified assertions passed. One valid OR/CI swap was unexpectedly quarantined.
OR 0.5, CI [⅓, ⅔] should become OR 2, CI [1.5, 3]. This local integration limitation remains unresolved. Small synthetic tests do not establish empirical indel, palindrome or liftover accuracy.Scientific implementation frozen at commit 15e8878. Four perturbation classes; independent reference-base checks; no overlap with the specified earlier truth/subset registers. The FinnGen full source had been processed historically, so this is not blinded, never-before-processed data. The GWAS Catalog extension is same-study and same-region.
Of 14 production-download scenarios, 13 had the expected disposition and fields. All five expected quarantines occurred; eight of nine expected releases occurred. No incorrect release was observed in these small scenarios.
Download September 30 report ↗Four public subset runs exercise configured behavior and audit accounting. Retention is an observational outcome, not ground-truth accuracy.
Each bar uses its own input count as the denominator. Open a dataset for its exact counts, input checksum and recorded limitations.
Completed full-file runs extend format coverage. They are not additional independent studies or controlled accuracy tests.
Build-corrected supplemental runs completed for GWASLab 4.2.1 and tidyGWAS 1.0.0 on PLINK2 and Pan-UKBB inputs. MungeSumstats 1.20.0 failed on both GRCh37 inputs because the required SNPlocs reference was missing.
| Input | GWASLab | tidyGWAS |
|---|---|---|
| Pan-UKBB | 98.01% | 98.03% |
| PLINK2 | 99.95% | 99.99% |
Download comparator feasibility report ↗
No locally verifiable Pan-UKBB Harmonizer bundle is included. These runs do not support a speed, usability or scientific-accuracy ranking.
Bell’s Palsy summary statistics from All of Us EUR, AFR and AMR; FinnGen; EstBB; UK Biobank; and JPBB were prepared for METAL. This application is separate from the benchmark and is not counted toward its accuracy endpoints.
Download the reports behind this page. Large inputs and row-level evidence packages are not hosted here. Release archival and DOI remain pending.
Source reports include references to local archived packages that are not hosted here. Their historical limitations should be read alongside each result. The September 30 extension is shown separately with its measured limitations.