GWAS and Colocalization: eQTL, pQTL, sQTL, mQTL

QTL mapping · Colocalization (coloc, eCAVIAR) · SMR · Pleiotropy · Mendelian randomisation

Published 2026 · Last updated 2026-09-27 · Hailu Zhou

A GWAS hit is a coordinate, not a mechanism. The credible set rarely contains a coding variant, and the nearest gene is frequently the wrong culprit. Colocalization is the step that turns a statistical signal into a testable molecular hypothesis: does the disease association and the molecular QTL association share the same causal variant? If you skip this step and let proximity or LD guide you, you will spend months chasing the wrong transcript.

What QTL mapping actually estimates

A quantitative trait locus is a genomic region where genetic variation associates with a molecular phenotype. The molecular phenotype defines the assay: eQTL for gene expression, pQTL for protein abundance, sQTL for splicing ratios, mQTL for DNA methylation. The mapping machinery is ordinary regression of the molecular trait on genotype, most often a linear mixed model with a kinship or genotype-derived random effect to control population structure and relatedness. Per-gene permutation gives an empirical threshold, and one reports a variant-level association statistic for every SNP in a cis window (usually 1 Mb) around the gene or feature.

The key output for colocalization is not the top SNP but the full summary statistics vector: effect sizes and standard errors across the locus. That vector is what lets us ask whether the two traits share a causal variable. Two practical cautions matter. First, eQTL effect sizes are context-specific — a liver eQTL does not transfer to brain — so the tissue must match the disease. Second, low-powered QTL studies yield noisy vectors and produce apparent colocalization with nothing at all; check the number of significant cis SNPs before believing any result.

Colocalization: the shared-causal-variant question

Two traits colocalize when a single causal variant explains both association signals. The Bayesian framework of Giambartolomei and colleagues enumerates five hypotheses: no association, association with trait 1 only, trait 2 only, both driven by distinct causal variants, and both driven by one shared variant1. Priors on p1, p2, and p12 shape posterior probability; the recommended colocalization support is a posterior probability of H4 above roughly 0.8, paired with low H3 to rule out linkage. This matters because H3 and H4 are often confused: two traits can be strongly associated in the same region yet be non-colocalizing if they are driven by different variants in LD.

eCAVIAR attacks the same question from the fine-mapping side, using an approximate causal-variant posterior and reporting a colocalization posterior probability defined over shared credible-set variants2. Its advantage is that it propagates fine-mapping uncertainty rather than conditioning on a single lead SNP. In practice, coloc with a sensitivity analysis over priors and eCAVIAR on the same locus are complementary reads, and agreement between them is the strongest evidence short of functional work.

SMR: fast, but with a known blind spot

SMR tests pleiotropy between a QTL and a complex trait through a single instrumental variable, using the top associated SNP per gene as the instrument3. It is cheap, runs genome-wide in minutes, and scales trivially to thousands of traits. The price is its one-SNP design: SMR reports a shared association, and it cannot distinguish a shared causal variant from two linked variants in LD. The standard remedy is HEIDI, which tests heterogeneity of the effect estimates and flags loci where linkage alone explains the signal. Run SMR for discovery, then confirm any candidate with coloc before writing it up.

Pleiotropy testing and what it does not solve

Pleiotropy is a variant or locus influencing multiple traits; it is the rule, not the exception. Tests for it operate at two levels. At the variant level, one can ask whether the same signal associates with many molecular phenotypes, which is informative but confounded by LD. At the causal level, methods like HyPrColoc examine a set of traits jointly and ask for a single shared causal variant across all of them, returning clusters of traits explained by one signal4. This is the honest framing: pairwise coloc answers "do A and B share a variant", while multi-trait approaches answer "do A, B, C, and D share one". Neither proves direction, and neither replaces experimental perturbation.

Colocalization as the bridge to Mendelian randomisation

Mendelian randomisation estimates a causal effect of an exposure on an outcome using genetic instruments, under assumptions of relevance, independence, and exclusion restriction. Colocalization is what protects the third assumption. If the QTL instrument and the disease signal do not share a causal variant, the instrument is almost certainly acting through a different pathway — horizontal pleiotropy — and the MR estimate is uninterpretable no matter how small its p-value. The workflow I recommend is: derive instruments from a tissue-matched QTL study, run MR, and then require colocalization between the exposure QTL and the outcome for every instrument used. Report both. An MR result without coloc is a hypothesis; with coloc it is a mechanistically anchored estimate.

A pragmatic pipeline

None of this is a substitute for perturbation experiments. It is a filtering discipline that keeps you from expending wet-lab effort on associations that were never going to replicate mechanistically.

References

  1. Giambartolomei C, Vukcevic D, Schadt EE, et al. Bayesian test for colocalisation between pairs of genetic association studies using summary statistics. PLoS Genet. 2014;10(5):e1004383.
  2. Hormozdiari F, van de Bunt M, Segrè AV, et al. Colocalization of GWAS and eQTL signals detects target genes. Am J Hum Genet. 2016;99(6):1245-1260.
  3. Zhu Z, Zhang F, Hu H, et al. Integration of summary data from GWAS and eQTL studies predicts complex trait gene targets. Nat Genet. 2016;48(5):481-487.
  4. Foley CN, Staley JR, Breen PG, et al. A fast and efficient colocalization algorithm for identifying shared genetic risk factors across multiple traits. Nat Commun. 2021;12:764.
  5. Wallace C. A more accurate method for colocalisation analysis allowing for multiple causal variants. PLoS Genet. 2021;17(9):e1009440.
  6. Hormozdiari F, Kostem E, Kang EY, et al. Identifying causal variants at loci with multiple signals of association. Genetics. 2014;198(2):497-508.
  7. Davey Smith G, Hemani G. Mendelian randomization: genetic anchors for causal inference in epidemiological studies. Hum Mol Genet. 2014;23(R1):R89-R98.
  8. GTEx Consortium. The GTEx Consortium atlas of genetic regulatory effects across human tissues. Science. 2020;369(6509):1318-1330.