What a Significant TWAS Result Says
A transcriptome-wide association study (TWAS) tests whether a trait is associated with genetically predicted expression of a gene. A small p-value is evidence against the test’s null hypothesis of no association between that prediction and the trait, under the model and assumptions used. It is not, by itself, proof that the gene’s expression causes the trait.
The distinction matters because the prediction is built from genetic variants. Those variants may be associated with the trait through an effect on expression, but they may also tag a different causal variant through linkage disequilibrium (LD), or cause several genes’ predicted expression values to move together. A TWAS signal can therefore help prioritize genes and hypotheses without uniquely identifying a causal mechanism.
The test’s interpretation assumes that the prediction weights and trait-association analysis are suitably specified, that the standard errors are calibrated for the study design, and that the weights are relevant to the population being analyzed. Prediction quality, ancestry differences, sample dependence, and other modeling choices can affect the result. Earlier material on PrediXcan and FUSION modeling choices describes how choices of weights and testing procedures can differ; here the focus is what a detected association can and cannot mean.
Three Routes to a TWAS Signal
A useful way to understand the ambiguity is to consider three possible data-generating explanations. They can produce similar association statistics, even though their biological interpretations differ.
- Expression-mediated effect. A genetic variant changes expression of a gene, and that expression change contributes to the trait. This is the causal-expression explanation of interest, but a TWAS association alone does not distinguish it from the alternatives.
- Linked causal variant. A variant affects the trait through a route other than the tested gene’s expression. A variant used to predict expression is correlated with that causal variant because of LD, so predicted expression is associated with the trait even when expression has no causal effect.
- Correlated gene predictions. Two or more genes can have predicted expression values correlated through shared or correlated genetic predictors. A trait association with the shared genetic component can then generate associations for multiple genes. Several significant genes do not necessarily imply several causal expression effects.
These explanations are not mutually exclusive in a real genomic region. For example, a genuine expression effect could coexist with direct effects of nearby variants. A single marginal TWAS statistic does not decompose those contributions.
A Short Derivation: How LD Creates Association
Suppose genotypes are centered and standardized, with covariance matrix \(R\), and the trait follows the linear model \(Y=\beta^T G+\varepsilon\). Here \(\beta\) describes genetic effects on the trait, and assume \(\operatorname{Cov}(G,\varepsilon)=0\). For a gene with prediction weights \(w_g\), predicted expression is \(P_g=w_g^T G\). Its covariance with the trait is
The last equality follows because the covariance between \(G\) and \(\varepsilon\) is zero, while \(\operatorname{Cov}(G,G)=R\). Thus, the predicted expression can covary with the trait whenever the weights overlap, through \(R\), with variants that affect the trait. This covariance can be nonzero even if expression itself is absent from the trait-generating model.
In a simple linear regression of \(Y\) on \(P_g\), the population slope is \(\operatorname{Cov}(P_g,Y)/\operatorname{Var}(P_g)\), provided \(\operatorname{Var}(P_g)>0\). The TWAS test assesses whether that association is distinguishable from zero, with an uncertainty estimate appropriate to its design. It does not test whether the gene’s expression is a mediator rather than a correlated genetic proxy.
Worked Example: A Signal Under a Prediction-Score Model
Consider a hypothetical study with 800 participants. Suppose the predicted expression \(P\) and trait \(Y\) are standardized, and the data-generating model is \(Y=0.35P+U\), where \(U\) is independent of \(P\) with variance \(0.8775\). Then \(\operatorname{Var}(Y)=0.35^2+0.8775=1\). Suppose the fitted TWAS slope is \(0.350\), with \(SE=0.0331\).
The test statistic is
The two-sided normal-reference p-value is approximately \(3.9\times10^{-26}\), so it is very small. In this hypothetical data-generating model, the association arises because the trait depends on the prediction score \(P\); this does not specify an expression-mediated biological effect. In an actual TWAS, however, observing a similarly small p-value does not establish that this is the model that generated the data. The same statistic can arise from LD with a direct trait variant, as the next example shows.
Worked Example: A Linked Variant Produces a Strong TWAS Association
Suppose two standardized variants \(G_1\) and \(G_2\) have correlation \(0.90\). Variant \(G_1\) affects a standardized trait directly, while \(G_2\) is used to predict a gene’s expression. Let
with \(\varepsilon\) uncorrelated with both variants. Then \(\operatorname{Var}(Y)=0.30^2+0.91=1\). Assume the predicted expression is \(P=G_2\). The population association slope of \(Y\) on \(P\) is
There is no expression term in the trait model: the trait depends directly on \(G_1\), not on the gene’s expression. Nevertheless, the association is substantial because \(G_2\) tags \(G_1\). For a hypothetical sample of \(n=1000\), an approximate standard error for the slope is \(\sqrt{(1-0.270^2)/1000}=\sqrt{0.9271/1000}\approx0.03045\). Thus
The two-sided normal-reference p-value is about \(7.5\times10^{-19}\). This is strong evidence of association between predicted expression and the trait under the model, but it does not show that expression mediates the \(G_1\) effect. The numerical example checks the key point: LD can yield an extremely significant TWAS result without a causal expression effect.
Worked Example: Correlated Predictions Give Multiple Gene Signals
Suppose two genes have standardized predicted expression values \(P_A=G_1\) and \(P_B=G_2\), where the standardized variants have correlation \(0.80\). Let the standardized trait follow \(Y=0.20G_1+\varepsilon\), where \(\operatorname{Var}(\varepsilon)=0.96\), so \(\operatorname{Var}(Y)=1\). The trait model contains no expression effect.
For gene A, the association slope is \(\operatorname{Cov}(G_1,Y)=0.20\). For gene B, it is \(\operatorname{Cov}(G_2,Y)=0.80(0.20)=0.160\). With \(n=1000\), approximate standard errors and test statistics are
Both marginal associations would be statistically strong, although the trait in this hypothetical model is driven directly by \(G_1\), not by either gene’s expression. The two gene predictions share correlated genetic information, so their signals are not independent evidence for two causal genes. The example illustrates why counting significant genes is not equivalent to counting causal expression effects.
What Can Help, and What Remains Unresolved
A significant TWAS result is useful as an association-level finding: it identifies a predicted-expression phenotype whose genetic component tracks the trait in the analyzed data. It can prioritize genes for further study, but causal interpretation requires additional evidence and assumptions. The earlier material on linkage disequilibrium and colocalization explains how shared variant evidence is evaluated. Colocalization can help distinguish a shared causal signal from two nearby signals that merely lie in LD, but even shared genetic association does not automatically prove that expression mediates the trait effect.
Other approaches answer different questions. An eQTL analysis asks whether genetic variants are associated with measured expression. Fine-mapping seeks to narrow plausible causal variants. Colocalization assesses whether trait and expression associations are compatible with sharing a causal variant. A TWAS instead tests association with a weighted genetic prediction of expression. These methods can inform one another, but none should be described as proving a causal expression mechanism without the assumptions needed for that claim.
Diagnostics can reveal reasons to be cautious. Check whether the prediction is reliable in a relevant reference population, whether the association depends heavily on a small number of variants, and whether LD or allele alignment has been handled appropriately. Compare the gene-level result with the local variant-association pattern and examine whether nearby genes have correlated predictions. If small changes in weights or modeling choices substantially change the result, the identity of the prioritized gene may be unstable. Conversely, a robust association across reasonable analyses supports the stability of the association, not necessarily its causal interpretation.
A causal claim would require a defensible argument that the genetic instruments affect the trait through the expression process under study, rather than through another pathway, and that relevant confounding and model violations are addressed. Those conditions are not tested merely by obtaining a small TWAS p-value. Expression perturbation evidence, functional data, and carefully interpreted genetic evidence may strengthen a mechanistic case, but each has its own assumptions and limitations.
Common Mistakes and Limitations
- Wrong reasoning: “The TWAS p-value is very small, so this gene causes the trait.” Correction: A small p-value supports association with predicted expression under the test; LD or correlated predictions can produce that association without mediation.
- Wrong reasoning: “Two significant genes mean two causal genes.” Correction: Their predictions may share correlated genetic components. The marginal tests do not establish distinct or independent causal effects.
- Wrong reasoning: “A TWAS coefficient is the effect of experimentally changing expression.” Correction: It is an association slope for a genetically predicted quantity, on the analysis scale. It need not equal an intervention effect.
- Limitation: A statistically significant association can still be affected by poor prediction, population differences, LD structure, or model misspecification. Statistical significance does not remove those concerns.
Check Your Understanding
Answer each question, then compare your reasoning with the solutions below.
- In the derivation, what makes \(w_g^T R\beta\) nonzero even when the trait model contains no expression term?
- A TWAS test has a two-sided p-value of \(10^{-12}\). What does that result support, and what does it not establish?
- Two genes have significant marginal TWAS results and highly correlated predicted expression. Why is it not valid to conclude that both genes independently cause the trait?
- In the linked-variant example, what is the predicted-expression association slope when the variant correlation is \(0.90\) and the direct trait effect is \(0.30\)?
- Name one diagnostic that can reveal instability or a possible LD explanation, and state why passing it still does not prove causality.
Solutions. (1) The covariance matrix \(R\) transmits association between the prediction weights and trait-effect variants through LD; therefore \(w_g^T R\beta\) can be nonzero without mediation. (2) It supports evidence of association between predicted expression and the trait under the test model. It does not establish that changing actual expression changes the trait or distinguish mediation from LD. (3) Their predictions can share genetic information, so both marginal results may reflect one underlying genetic association rather than independent expression effects. (4) The slope is \(0.30(0.90)=0.270\). (5) One can examine whether the result is sensitive to reasonable prediction-weight choices or whether nearby variants show a plausible LD-based explanation. Stability or a coherent variant pattern can strengthen the association interpretation, but neither alone rules out other causal pathways.