Skip to content

Statistical Colocalization

Summary

Statistical colocalization is a Bayesian or frequentist analysis method in statistical genetics that determines whether two or more phenotypic traits share a common causal genetic variant at a specific genomic locus. It distinguishes true shared genetic etiology (pleiotropy) from mere physical proximity (linkage disequilibrium) between distinct causal variants. This approach is widely used to integrate genome-wide association study (GWAS) signals with molecular traits like gene expression (e.g., eQTL) and protein levels (pQTL).

Core Concepts

When multiple traits show significant genetic associations within the same genomic region, it could be due to: 1. Horizontal Pleiotropy: A single causal variant directly affects multiple traits. 2. Linkage Disequilibrium (LD): Two distinct causal variants, each regulating one trait, are in high correlation, leading to a correlated association signal across the locus.

Statistical colocalization methods model the patterns of association to calculate the posterior probability of a shared causal variant.

Key Frameworks and Algorithms

Several tools have been developed to perform colocalization: - COLOC: A pairwise Bayesian method that evaluates five mutually exclusive hypotheses ($H_0$ to $H_4$). $H_4$ represents the hypothesis that both traits share a causal variant. - MOLOC: An extension of COLOC that supports multi-trait colocalization for three or four traits simultaneously. - HyPrColoc: A computationally scalable method developed by the Burgess Group that uses hypothesis prioritisation to compute the Posterior Probability of Full Colocalization (PPFC) for dozens of traits in a single run. - eCAVIAR / ENLOC: Colocalization methods that relax the single causal variant assumption, allowing for multiple causal variants at the same locus.

Applications

Colocalization is crucial for: - Causal Gene Identification: Priorizing candidate causal genes by colocalizing disease GWAS signals with cis-eQTLs and cis-pQTLs. - Drug Target Validation: Determining if a target protein's genetic association colocalizes with disease risk. - Pathway Analysis: Mapping out molecular cascades by tracing colocalization across transcriptomic, proteomic, and metabolomic datasets.

Promise and Limitations of Probabilistic Colocalization

A systematic evaluation of probabilistic (SNP-level) colocalization methods — coloc, eCAVIAR, and ENLOC/fastENLOC — showed all three share the same underlying Bayesian target, P(g=1, d=1 | eQTL data, GWAS data), differing only in how they parameterize the prior probability of a shared causal variant.[3]

  • The enrichment prior dominates false-positive risk: of the analytical choices that shape colocalization outcomes, the specification of the prior enrichment level (the ENLOC parameter α1, quantifying how enriched molecular QTLs are among GWAS hits) was the single factor capable of severely inflating false-positive colocalization calls when set too aggressively. SNPs with only modest marginal evidence from either the eQTL or GWAS analysis are the most sensitive to this prior; SNPs with strong or weak evidence in both studies are comparatively robust to it. The recommended remedy is to estimate the enrichment prior from the observed data (as fastENLOC does via multiple imputation) rather than fixing it a priori, and to keep fine-mapping and colocalization as separate analytical steps.
  • Practical power is limited: beyond prior specification, numerous uncontrollable practical factors — reference-panel/study population mismatch, sample size, and LD structure differences — each erode power further. In an extreme cross-population mismatch (Yoruba GWAS data analyzed against Finnish eQTL data), colocalization power dropped to roughly 50% of the matched-population baseline, and only part of that loss was attributable to enrichment-parameter misestimation.
  • Scale of the problem: applying this framework to 4,091 complex-trait GWAS against multi-tissue GTEx v8 eQTL data suggested that colocalization between molecular QTLs and causal complex-trait associations is widespread in principle, but only a small fraction of true colocalization events can currently be confidently identified given the power limitations of available eQTL and GWAS sample sizes — a benchmark the authors offer for calibrating expectations in future integrative genetic studies.

Complementarity with Sequence-Based Sign Prediction

Colocalization derives a GWAS locus's direction of effect from a measured eQTL, so it inherits that study's power constraints — allele frequency, sample size, tissue coverage and population matching. Sequence-to-function models predict the direction from DNA alone and therefore fail under different conditions.

Benchmarking across 18,537 GWAS credible sets, AlphaGenome and COLOC ($H_4 > 0.95$) resolved the direction of effect for largely non-overlapping sets of loci, so using both increases total yield rather than corroborating the same calls. AlphaGenome resolved approximately 4-fold more credible sets in the lowest minor-allele-frequency quintile — the stratum where eQTL power is weakest — which the authors attribute to its reduced dependence on the population-genetics parameters governing association detection.[4]

Synthesis: the power losses documented above for colocalization [3] — enrichment-prior misspecification, LD/population mismatch, small eQTL sample size — are all properties of the observational study design, whereas the sequence-model failure modes are long variant-to-TSS distance and unmodelled cell types [4]. The two method families therefore appear to fail under disjoint conditions. No cited study has jointly tested this claim.

See Direction-of-Effect Assignment at GWAS Loci for the full comparison and a suggested combined workflow.

Molecular Trait Choice Changes the Colocalization Result

Colocalization is only as informative as the molecular trait supplied on one side of it. Splitting splicing QTLs into productive and unproductive classes (LeafCutter2) found that 225 of 2,897 GWAS loci (7.8%) across 20 complex traits colocalize with at least one unproductive-splicing QTL in at least one GTEx tissue — a signal invisible when productive and unproductive events are pooled, since only the unproductive class correlates with host-gene expression.[5] See RNA-seq-Derived Molecular Phenotypes.

Citations