TWAS Across Tissues: Sharing And Heterogeneity

Why Compare TWAS Results Across Tissues?

A gene may have TWAS associations in several tissues, in one tissue, or in tissues with estimates that differ in direction or magnitude. These patterns can help describe where genetically predicted expression is associated with a trait. But a list of tissue-specific p-values is not itself a comparison: a small p-value in one tissue and a larger one in another do not establish that the underlying effects differ.

The central task is to compare estimates and their uncertainty. Tissue-specific TWAS estimates may be correlated because tissues share participants, genetic variants, or expression-prediction information. Their precision may also vary because prediction models differ in accuracy and sample size. A sound comparison therefore requires a common effect scale and, when possible, the covariance among estimates. As emphasized in the earlier material on interpreting TWAS signals without claiming causality, a TWAS association is evidence about genetically predicted expression, not proof that changing expression in a particular tissue causes the trait.

The estimates must refer to comparable quantities. If predicted expression is measured on different scales across tissues, a coefficient per unit of predicted expression need not be directly comparable. One option is to express each effect per one standard deviation of genetically predicted expression, provided that this is the intended scale and the estimates are transformed consistently. Record the scale, outcome units, effect direction, and allele or model conventions before interpreting cross-tissue differences. A comparison of differently scaled coefficients can create apparent heterogeneity or conceal real differences.

Association, Sharing, and Heterogeneity Are Different Questions

Three questions are often conflated. First, is there evidence of an association in at least one tissue? Second, is there evidence that effects are compatible with a common effect across tissues? Third, do the data show more cross-tissue variation than expected from sampling error under a common-effect model? These require different statistics.

A joint Wald test of no association in any tissue uses the vector of estimates and its covariance matrix. Under the null hypothesis \(H_0:\boldsymbol{\beta}=\boldsymbol{0}\), and assuming an approximately multivariate normal estimate with a valid nonsingular \(\boldsymbol{\Sigma}\), the statistic

has approximately a chi-square distribution with \(T\) degrees of freedom. It tests whether the vector is zero; it does not test whether tissue effects are equal. A large \(W\) can arise from a shared effect or from effects confined to a subset of tissues.

For a common-effect summary, assume the tissue estimates satisfy

where \(\theta\) is the common effect on the chosen scale, \(\boldsymbol{1}\) is a vector of \(T\) ones, and \(\boldsymbol{\Sigma}\) is known or treated as adequately estimated. The generalized least-squares (GLS) estimator and its standard error are

This uses more precise information while accounting for correlation; it is not generally the simple average of tissue estimates. A common-effect estimate is useful as a summary only if a common effect is a reasonable working model.

Why the GLS and Heterogeneity Formulas Work

The GLS estimate is the value of \(\theta\) that minimizes the covariance-weighted squared residuals \(L(\theta)=(\hat{\boldsymbol{\beta}}-\theta\boldsymbol{1})^\top\boldsymbol{\Sigma}^{-1}(\hat{\boldsymbol{\beta}}-\theta\boldsymbol{1})\). Differentiating with respect to \(\theta\) gives

Setting this derivative to zero yields \(\boldsymbol{1}^\top\boldsymbol{\Sigma}^{-1}\hat{\boldsymbol{\beta}} =\theta\boldsymbol{1}^\top\boldsymbol{\Sigma}^{-1}\boldsymbol{1}\), which gives the stated GLS estimator. Since it is a linear combination of a multivariate normal vector, its variance is

For \(Q\), subtracting the fitted common effect leaves residual variation in the \(T-1\) directions orthogonal to the fitted common-effect direction, using the covariance-weighted metric. Under the stated normal model, this residual sum of squares has a chi-square distribution with \(T-1\) degrees of freedom. If covariance is misspecified, estimates are not approximately normal, or too few independent observations support covariance estimation, this reference distribution may be inaccurate.

Worked Comparisons Across Tissues

Interpreting Patterns and Checking Assumptions

A defensible interpretation separates the estimated pattern from the strength of evidence. Consistently directed estimates with compatible magnitudes can support a description of a shared association across the tissues studied. A significant heterogeneity statistic indicates that a common-effect model fits poorly, but it does not identify which tissue differs or explain why. Conversely, failure to reject homogeneity can reflect imprecision, especially where expression prediction is weak.

Unequal prediction accuracy affects precision and power. A tissue with a weaker expression-prediction model may have a larger TWAS standard error, so its estimate can be less informative even when its underlying association is similar. Do not interpret a less significant result as evidence of a smaller biological effect. Nor should a highly precise tissue automatically be treated as the only relevant tissue: precision and tissue-specific effect are distinct.

Before reporting cross-tissue results, check that:

• Effects use comparable expression and outcome scales, directions, and gene definitions.
• Standard errors and cross-tissue covariances correspond to the actual study design, including sample overlap where relevant.
• The covariance matrix is positive definite and not so poorly estimated that its inverse is unstable. Near-duplicate tissue estimates can make multivariate tests numerically fragile.
• Each tissue has enough prediction accuracy for its estimate to be informative; very weak prediction can yield unstable or uninformative estimates.
• The common-effect model is treated as an assumption to assess, not as an automatic description of biology.

When covariance is unavailable, an independence-based comparison may be shown as an explicitly limited approximation, but it can misstate uncertainty. Alternatives include reporting the tissue estimates and standard errors without a formal pooled test, using a covariance-aware joint test when the scientific question is any association, or fitting a model that allows between-tissue variation when that is scientifically justified and the data support estimating it. These approaches answer different questions; none turns a TWAS association into evidence of causation. Results across many genes and tissues also raise a separate multiplicity problem, which must be handled in the relevant analysis rather than by interpreting isolated p-values.

Common Mistakes and Limitations

A heterogeneity test is not a tissue-localization procedure. A large \(Q\) says that the common-effect model is inconsistent with the observed estimates and assumed covariance; it does not establish a causal tissue, rule out model misspecification, or identify the source of the discrepancy. A small \(Q\) does not establish exact equality. In addition, a common TWAS association across tissues can arise because the prediction models use overlapping genetic information or because estimates are correlated, rather than because expression has an independent causal role in every tissue.

The key reporting practice is therefore to show the tissue-specific estimates and standard errors, state the scale and covariance assumptions, and describe any common-effect or heterogeneity test according to the question it actually answers. Treat the result as a pattern of association across the measured tissues, with uncertainty, rather than a definitive map of tissue-specific mechanism.