Colocalizing Across Tissues And Cell Types

Why Compare Colocalization Across Tissues?

An eQTL signal can be detectable in one tissue and not another, while a GWAS association may be relevant in several tissues. Comparing colocalization across tissues asks whether the GWAS signal and the eQTL signal for a particular gene are supported as sharing a causal variant in one context, several contexts, or neither. This comparison can help prioritize where a genetic association may be connected to gene regulation.

The central interpretive challenge is that different posterior probabilities across tissues do not automatically indicate different biology. Tissues can differ in sample size, expression variability, cell composition, genotype coverage, and the strength of the eQTL. A signal that is well supported in a large tissue sample may remain uncertain in a smaller sample even if the underlying regulatory effect is similar.

A broadly shared signal is supported as colocalized with the same GWAS signal in more than one tissue or cell type. A tissue-specific sharing pattern means that sharing is well supported in one context and that another context provides evidence against sharing, or remains sufficiently informative to distinguish the alternatives. A high PP.H4 in one tissue and a low PP.H4 in a second tissue is not, by itself, enough to establish tissue specificity: the second tissue may simply be underpowered.

What Must Be Comparable?

Compare like with like. For each tissue, the analysis should concern the same GWAS signal, the same gene, and a consistent genomic region. Variant identifiers, genome build, alleles, and effect directions must be harmonized, and the variants available for analysis should be checked. The posterior probabilities also depend on prior choices; use comparable priors across tissues or explain clearly when they differ.

The single-signal interpretation of the five hypotheses assumes that the association patterns being compared can reasonably be represented by one relevant signal in each trait. When a region contains multiple causal signals, apply the conditional or fine-mapping-based approach to match the relevant signals before comparing tissues. A gene may have more than one eQTL signal, and a GWAS region may have more than one association signal; treating each region as if it had only one can confuse distinct signals with a tissue difference.

Finally, PP.H4 is a posterior probability conditional on the data, model, and priors used for that tissue. It is not a measure of the size of the regulatory effect, nor does colocalization alone establish that expression mediates the GWAS effect. Evidence for a shared variant supports a shared genetic association, not a causal mechanism from expression to disease.

Why Sample Size Changes the Evidence

For a simple eQTL regression, suppose expression is modeled as a linear function of genotype, with independent individuals, a correctly specified mean model, and residual variance \(\sigma^2\). If genotype is centered and has variance \(V_G\), the approximate standard error of the estimated per-allele expression effect is \(\sigma/\sqrt{nV_G}\). If genotype is standardized so \(V_G=1\), this becomes \(\sigma/\sqrt n\). Thus, all else equal, a larger tissue sample produces a smaller standard error and a larger absolute association statistic for the same effect.

This approximation relies on comparable genotype variation and residual variance as well as comparable measurement and sampling. In real tissue datasets, those quantities need not match: a larger sample does not guarantee more precise estimates if expression is noisier or the analyzed cell population is more heterogeneous. Therefore, sample size is an important diagnostic, not a complete explanation of differences in evidence.

The equation explains why comparing p-values or PP.H4 values as though they were direct measures of biological effect is unsafe. A larger \(n\) can strengthen eQTL evidence, which can in turn change the colocalization posterior. But PP.H4 also depends on the GWAS evidence, the pattern of association across variants and linkage disequilibrium, the priors, and the competing hypotheses. There is no general conversion from a difference in PP.H4 to a difference in regulatory effect size.

Reading Tissue Patterns

Interpret tissue-level posterior results as a pattern, not as a league table of tissues. The five posterior probabilities for a tissue sum to one; PP.H4 should be read alongside PP.H3 and the probabilities for the other hypotheses. A high PP.H4 is informative about a shared variant under the model. A low PP.H4 can instead reflect support for distinct variants, little association evidence, or uncertainty among hypotheses. Those situations have different meanings.

A stronger tissue-specific pattern requires more than a difference in PP.H4. Consider a tissue in which PP.H4 is high and another in which PP.H3 is high, with strong evidence that both traits are associated in the second tissue. The latter pattern supports distinct causal variants there, rather than merely a failure to detect an eQTL. Even then, check signal matching, variant coverage, and the assumptions of the analysis before describing the result as tissue-specific.

A Short Justification for the Sample-Size Effect

For centered genotype values \(G_i\), consider the regression estimator \(\hat{\beta}=\sum_i G_iY_i/\sum_i G_i^2\), where \(Y_i=\beta G_i+\epsilon_i\), and the residuals are independent with mean zero and variance \(\sigma^2\). Conditional on the genotypes, the estimator's variance is

If the genotype variance is approximately \(V_G\), then \(\sum_iG_i^2\) is approximately \(nV_G\). Taking the square root of the variance gives \(SE(\hat{\beta})\approx\sigma/\sqrt{nV_G}\). Consequently, under these assumptions, increasing \(n\) reduces the standard error at a rate proportional to \(1/\sqrt n\). This justifies the sample-size comparison above; it does not establish that tissues with different sample sizes have identical residual variance, genotype variance, or measurement quality.

Diagnostics, Alternatives, and Limitations

Before interpreting a cross-tissue pattern, check that the gene is measured reliably in each tissue, that the relevant variants are tested, and that allele coding, genomic coordinates, and regional boundaries agree. Inspect the full set of posterior probabilities rather than PP.H4 alone. Review uncertainty in fine-mapping and whether the candidate shared variants are consistent across tissues. If a tissue has little eQTL evidence, describe its colocalization result as weak or uncertain rather than as evidence against sharing.

A useful alternative to a simple high-versus-low PP.H4 comparison is to report, for each tissue, the full posterior distribution and the eQTL effect estimate with its standard error and sample size. This gives readers evidence about both association precision and colocalization uncertainty. A pooled or meta-analyzed eQTL can increase power, but may conceal tissue differences and does not itself identify the context in which sharing occurs. A tissue-by-tissue analysis preserves that context, while requiring careful interpretation of power and comparability.

Do not multiply tissue-level posterior probabilities to claim a joint probability of sharing across tissues unless a joint model justifies that calculation. Estimates from tissues may be statistically dependent, and even independent tissue samples do not by themselves make separately calculated posterior events independent. The joint probability requires an appropriate model and assumptions. Also, a difference between posterior values is not a formal test of different effects between tissues; such a test must account for the uncertainty and dependence of the association estimates.

These analyses have further limits. Tissue labels can represent heterogeneous cell populations, and a tissue-level eQTL may reflect cell-type proportions rather than regulation within a specific cell type. In addition, colocalization does not establish direction of effect, mediation, or a molecular mechanism. The strongest defensible conclusion is tied to the tested gene, signal, tissues, data, and model.