From GWAS Signal to Causal Variant: Validating Non-Coding Risk Loci

A GWAS peak is a statistical signal averaged across thousands of individuals and hundreds of thousands of correlated variants. Turning it into a causal variant requires several kinds of evidence that are often presented as a single ladder, when in fact each rung tests a different thing and fails in its own way.

The failures are predictable. Credible sets are treated as if they were shortlists of causal variants rather than probabilistic summaries. Colocalization is read as proof of shared causality rather than as evidence about a shared signal. MPRA hits in a plasmid context are reported as regulatory effects. Each step is useful; each is systematically overinterpreted.

This article sets out the decision logic for the middle of the pipeline — from locus to candidate variant, and from candidate to evidence. The regulatory biology involved is in our pillar article; the enhancer-level evidence tiers are in the enhancer–gene pair validation roadmap, which this piece assumes.

Fine-mapping produces a set of variants with posterior probabilities summing to a stated confidence. Three consequences follow:

• • A 95% credible set of 40 variants is 40 candidates, not one lead. Reporting the top variant as "the" causal variant is a claim the statistics do not support.
• • Credible sets are model-dependent. Prior settings, LD reference panel choice, and sample ancestry all shift membership substantially. Sets computed with a mismatched LD panel are unreliable, and this is a common and avoidable error.
• • Power determines resolution. With small effect sizes at a locus, fine-mapping may be unable to resolve below dozens of variants no matter how the model is specified.

Practical implication: treat the credible set as the experiment's input population. Any functional design should state how many variants it tests and how it sampled them, and should prefer exhaustive testing within small sets over cherry-picking the lead.

Colocalization tests whether two association signals — a trait GWAS and an expression QTL, for example — are consistent with sharing a single causal variant. It is powerful and frequently misread.

Two further cautions. First, colocalization has low power when both signals are weak; a negative result is often uninformative rather than evidence against. Second, the eQTL must be measured in a relevant tissue; colocalization with a blood eQTL for a metabolic trait may reflect a shared regulatory architecture rather than the disease-relevant mechanism.

Editing is expensive, so the ordering of the funnel matters:

1. Intersect the credible set with functional annotations in the relevant cell type: accessible chromatin, active enhancer marks, factor occupancy, conserved motifs.
2. Check allele-specific effects in existing data where possible — allelic imbalance in expression or accessibility (ASE, AS-ATAC) is direct evidence that the variant has regulatory consequence in a physiological context.
3. Run a reporter-based assay (MPRA or a smaller targeted version) on the credible set to rank alleles. Report it as prioritization, not function, because plasmid context differs from native chromatin.
4. Select for editing: variants with consistent annotation, allele-specific support, and a plausible target gene. Typically two to four go forward.

If steps 1–3 converge on a single variant, the editing program is well justified. If they disagree, that disagreement is informative and should be reported rather than resolved by choosing the most convenient candidate.

The standard design is a pair — or better, a panel — of lines differing only at the variant:

• • Risk allele and protective allele introduced into the same parental background, ideally in both directions so that the direction of the edit does not confound the result.
• • Base editing where the substitution is a clean C•G→T•A or A•T→G•C transition; prime editing for other substitutions and small indels. Both avoid the double-strand break that would otherwise introduce indels at the site.
• • Heterozygous design where the endogenous locus is heterozygous in a relevant cell type, enabling allele-specific expression measurement from the same cells — often the most sensitive readout available.
• • Multiple clones per genotype, since clonal variation in expression can exceed the variant effect.

What to measure: the nominated target gene's expression as the primary endpoint, neighbouring genes as specificity controls, and — if the claim is mechanistic — the chromatin or factor-occupancy change predicted by the annotation.

A defensible causal claim typically includes:

The comparison of functional effect size to the population effect size is a useful sanity check. A 40% expression change from a single variant is possible in a cell model, but it should be reconciled with the modest per-allele effect seen in the GWAS rather than presented as if the two measured the same thing.

• • Calling the lead variant causal because it was the lead.
• • Reporting MPRA alone as functional validation.
• • Omitting the LD panel used for fine-mapping.
• • Using a convenience cell line with no relationship to the trait.
• • Reporting a single clone per genotype.
• • Not stating which allele was introduced, making the direction unverifiable.

Each of these is a line a reviewer can strike, and each is cheap to avoid.

EDITGENE builds isogenic variant panels by base and prime editing: risk and protective alleles in both directions, multiple independent clones per genotype, heterozygous backgrounds for allele-specific readouts, and matched reporter lines when a screening format is needed — all with sequenced edits and copy-number confirmation.

Send us the credible set and the trait-relevant cell type; we will help scope which variants warrant editing. Contact EDITGENE to request a project consultation or quote.

Fine-mapping, colocalization, and reporter assays are prioritization tools whose performance depends on sample size, ancestry, LD reference, and tissue relevance. Statistical support for a variant is not functional evidence.

Functional effect sizes measured in cell models need not match population effect sizes, and conclusions should state the cell context in which the effect was observed.

• • Nasser J, et al. Genome-wide enhancer maps link risk variants to disease genes. Nature. 2021;593:238–243.
• • Ulirsch JC, et al. Interrogation of human hematopoiesis at single-cell and single-variant resolution. Nature Genetics. 2019;51:683–693.
• • Tewhey R, et al. Direct identification of hundreds of expression-modulating variants using a multiplexed reporter assay. Cell. 2016;165:1519–1529.
• • Giambartolomei C, et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genetics. 2014;10:e1004383.
• • Wang QS, et al. Genetic and epigenetic effects of non-coding variants. Nature Reviews Genetics. 2023;24:1–18.