Worked examples

Start from a variant — APOE-ε4

Open the report for rs429358, the coding variant that defines the APOE-ε4 allele.

1. Read the score as a rank, not a number. The Δ on its own is a log-likelihood difference and means nothing in isolation. The report converts it into a percentile against all 6.48 M scored variants, and again within the variant's own functional class — a coding variant should be compared with coding variants, not with intergenic ones.
2. Check the sign separately from the magnitude. Strength comes from |Δ|. The sign says which allele the model finds less surprising, and nothing about the direction of any phenotypic effect.
3. Look at the six configurations side by side. Within one model size the three strand strategies agree closely; 7B and 40B do not. If your conclusion flips between them, say so.
4. Read the GWAS evidence panel. This variant carries genome-wide significant associations in all three biobanks — lipid traits in UKB and MVP, dementia and early-onset Alzheimer's in FinnGen. That is the answer to "is this a known association locus", straight from the stored summary statistics.
5. Follow one trait into the locus view. The "locus view" link on any trait row opens regional association tracks for that window with the trait already selected.

What this does not tell you: which gene the variant acts through, in which tissue, or whether it is causal rather than in LD with something causal. Inspect available records by source, with their gene, tissue, assay and allele-matching details; external resource links remain available. Gene or signal model statistics are shown on the corresponding analysis page, not in the Variant report.

Ask about a class of regions — human accelerated regions

Open the region-set module and pick Human Accelerated Regions (HAR, n=2,734), or upload your own BED file.

1. Conservation. Inspect each panel's score transformation, conservation measure, sample size and interval definition before comparing model configurations. The genome-wide values are pre-computed, so you always have a reference line to read the region set against.
2. Allele age and selection. GEVA allele age and SDS, split into all / repeat-masked / non-repeat sites. Intervals come from stratified site subsampling within chromosome × allele-frequency strata. Read the subset and pruning rules alongside the sample size; an interval is not a guarantee that all LD dependence has been removed.
3. Frequency and functional stratification. Whether the relationship survives when you hold allele frequency roughly constant, and how |Δ| distributes across region classes and coding consequences inside the set.

Watch the n. A set of a few hundred regions will give wide intervals. The per-stratum counts are printed on every panel — read them before reading the effect.

Start from a trait — UKB hypertensive diseases

Follow the manuscript's GCST90473520 example (UK Biobank, ICD10 I10–I15): open this study in the Guided Workflow or view its existing Trait Analysis results.

1. Check the original phenotype and source first. Search results can have different cohort, ancestry and phenotype definitions. Read each record's coefficient, uncertainty, z-score and within-biobank rank separately. Similar names or high ranks do not establish replication.
2. Check the trait has heritability to partition. The h² column is the trait's total observed-scale SNP heritability from the same run. A large z on a trait with h² near zero is partitioning almost nothing. The intercept is another diagnostic to interpret with the study design.
3. Switch the annotation. The signed Δ / |Δ| toggle loads the previously computed results for the other annotation; it does not rerun S-LDSC. In the earlier stored sensitivity analysis of 4,521 traits, the two agree closely (Spearman ρ = 0.99). The separate 14 September frozen sensitivity figures cover 4,767 paired phenotypes; their full sample and model definitions accompany the figures.
4. Inspect the FGF5 locus. Open chr4:80,205,413–80,333,456 with GCST90473520 already selected. The manuscript highlights rs12509595, at GRCh38 chr4:80,261,400, with 40B avgRC |Δ| ≈ 66.03. Inspect the association and score together; their overlap alone does not establish causality or formal statistical colocalization.

Read the model description on that page. This release uses the Evo2 score as the single continuous annotation, with no baseline-LD annotations. A positive coefficient for signed Δ refers to higher signed scores; for |Δ| it refers to greater score magnitude. Neither result establishes added information beyond conservation.

The second manuscript example is FinnGen F5_DEPRESSION_DYSTHYMIA. Open a 200-kb window around rs2888295 at SNRK–ANO10 with this study and 40B avgRC already selected. The window is a website navigation example; the stored study retains its original depression-or-dysthymia phenotype definition.