Three Different Questions at a Locus
A gene can have a strong transcriptome-wide association (TWAS) result without being the gene whose expression mediates a nearby trait association. Conversely, variant-level colocalization can support a shared genetic signal even when a gene’s TWAS test is not statistically compelling. These results answer related but distinct questions, so integrating them requires more than counting significant outputs.
For a candidate gene in a specified tissue, TWAS tests whether genetically predicted expression is associated with the trait under a prediction model and its assumptions. Colocalization compares genetic association patterns for expression and the trait, asking whether their signals are consistent with a shared causal variant rather than distinct variants in linkage disequilibrium (LD). Neither result alone proves that changing expression of the gene causes the trait.
In the standard colocalization framework, the hypotheses are \(H_0\): neither trait has a causal association in the region; \(H_1\): only the trait has an association; \(H_2\): only expression has an association; \(H_3\): both have associations, but their causal variants differ; and \(H_4\): both have associations and share a causal variant. Posterior probabilities for these hypotheses depend on the data, model, and prior assumptions. In particular, a large \(P(H_4\mid\text{data})\) is evidence for a shared signal relative to alternatives; it is not the probability that expression mediates the trait.
A Practical Integration Framework
For each candidate gene and tissue, retain the two analyses as distinct evidence fields. The TWAS result reports the gene-level association statistic and its significance under the relevant testing procedure. The colocalization result reports posterior support for the competing signal configurations, especially the contrast between \(H_4\) and \(H_3\). A high \(H_4\) posterior is most informative when the analysis includes adequate variant coverage, appropriate LD information, and a model suited to the number of causal signals.
A useful first comparison is the posterior odds of a shared rather than distinct signal:
This ratio is a comparison of two colocalization hypotheses, not a causal effect estimate. The posterior probabilities also depend on the priors and on how well the hypotheses represent the locus. When several independent causal signals may occur, a single-signal analysis can misrepresent the comparison. Conditional or fine-mapping-based colocalization, which is covered in the earlier material on conditional and fine-mapping-based colocalization, is then more appropriate.
Match the gene, tissue, locus boundaries, variant set, genome build, and relevant LD reference. Check that the expression and trait data represent comparable genetic signals.
Report the TWAS statistic or p-value and whether it meets the prespecified threshold after the relevant correction for testing genes and tissues. Do not interpret an unadjusted p-value as genome-wide or study-wide significance.
Report the posterior probabilities for the hypotheses, particularly \(H_3\) and \(H_4\), and state whether the method accounts for multiple signals when needed.
State whether there is evidence of a TWAS association, whether there is evidence for a shared variant-level signal, and whether the results do or do not establish a causal expression mechanism.
This workflow does not create a new combined p-value or a universal gene score. TWAS and colocalization use overlapping genetic association information, and their outputs are generally not independent. Multiplying their p-values, or treating them as independent votes, can therefore exaggerate the apparent strength of evidence. A transparent report gives the two results side by side and explains their joint interpretation.
Worked Interpretations
Worked Example: Strong TWAS and Strong Shared-Signal Evidence
Suppose a hypothetical gene, GENE-A, has a TWAS p-value of \(2.0\times10^{-8}\) in a tissue. The prespecified multiple-testing threshold for the analysis is \(5.0\times10^{-6}\). A colocalization analysis for the gene’s expression and the trait reports \(P(H_0\mid D)=0.01\), \(P(H_1\mid D)=0.02\), \(P(H_2\mid D)=0.03\), \(P(H_3\mid D)=0.10\), and \(P(H_4\mid D)=0.84\), where \(D\) denotes the analyzed data.
The TWAS result passes the stated threshold because \(2.0\times10^{-8}<5.0\times10^{-6}\). The colocalization posterior probabilities sum to \(0.01+0.02+0.03+0.10+0.84=1.00\). The posterior odds for a shared rather than distinct signal are
Thus, under the stated colocalization model and priors, the data favor a shared signal over two distinct signals, and the gene also has a statistically significant TWAS association at the given threshold. Together these results prioritize GENE-A for follow-up. They do not show that expression mediates the trait association: the shared variant could affect the trait through another gene or pathway, or could influence both expression and trait through separate biological routes.
Worked Example: Significant TWAS but Evidence for Distinct Signals
For hypothetical GENE-B, suppose the TWAS p-value is \(4.0\times10^{-7}\), with the same prespecified threshold \(5.0\times10^{-6}\). Colocalization gives \(P(H_0\mid D)=0.02\), \(P(H_1\mid D)=0.02\), \(P(H_2\mid D)=0.02\), \(P(H_3\mid D)=0.78\), and \(P(H_4\mid D)=0.16\). These probabilities sum to \(1.00\).
The TWAS result is significant under the specified threshold because \(4.0\times10^{-7}<5.0\times10^{-6}\). However, the posterior odds of a shared versus distinct signal are
The colocalization analysis favors \(H_3\) over \(H_4\): in this model, distinct causal variants are better supported than a shared one. The TWAS association should not be re-described as evidence that the trait and expression signals share a causal variant. Correlated variants in LD can produce a TWAS association even when the strongest expression and trait signals are distinct. The appropriate conclusion is that GENE-B has a significant TWAS association, but these colocalization results do not support a shared causal variant at this locus.
Worked Example: Shared Signal Evidence Without a Significant TWAS Result
Suppose a hypothetical locus has multiple association signals, so the investigators use a conditional colocalization analysis that compares the relevant expression and trait signals. For one signal pair involving GENE-C, the TWAS p-value is \(0.003\), above the prespecified threshold \(5.0\times10^{-6}\). The conditional colocalization posterior probabilities for that pair are \(P(H_0\mid D)=0.01\), \(P(H_1\mid D)=0.02\), \(P(H_2\mid D)=0.03\), \(P(H_3\mid D)=0.03\), and \(P(H_4\mid D)=0.91\).
The TWAS result does not meet the threshold because \(0.003>5.0\times10^{-6}\). For the conditional signal pair, the posterior odds for a shared rather than distinct signal are
Under the conditional model, the data strongly favor a shared signal over distinct signals for this pair. The appropriate integrated statement is therefore not “GENE-C has a significant TWAS association.” Rather, the conditional colocalization supports shared genetic signal evidence, while the specified TWAS test is not significant. These results do not conflict logically: the two procedures answer different questions and can have different sensitivity. Nor does the colocalization result alone establish that altered GENE-C expression is the causal mechanism.
Assumptions, Diagnostics, and Limits
Interpretation depends on the validity of each analysis. TWAS relies on a genetically predicted expression measure that is informative for the gene and tissue being studied, and on valid association testing and calibration. Colocalization relies on appropriately harmonized association data and LD context, adequate variant coverage, and a model whose assumptions about causal signals fit the region. The posterior probabilities are conditional on the specified data and priors; they are not assumption-free summaries.
Before combining results, check that alleles, effect directions, variant identifiers, genome build, and genomic region have been aligned. Review whether the eQTL evidence is sufficiently informative: weak or absent expression association can make colocalization inconclusive even if a TWAS result is present. Check the colocalization method’s treatment of multiple causal signals, because a locus with several signals may violate a single-signal model. Also distinguish a low \(P(H_4\mid D)\) caused by substantial support for \(H_3\) from a low \(P(H_4\mid D)\) when the data mostly support \(H_0\), \(H_1\), or \(H_2\). Those patterns do not mean the same thing.
The main alternatives are not interchangeable. A TWAS-only prioritization ranks genes by association with genetically predicted expression but does not directly establish a shared variant. Colocalization-only prioritization assesses shared genetic signal for expression and trait but does not test the same gene-level predicted-expression association as TWAS. Integrating both offers complementary information, provided the results are reported separately rather than converted into an unjustified single score.
Even concordant evidence has limits. A shared variant may influence multiple genes, tissues, or molecular processes. The variant may affect both expression and the trait without expression being the mediator. Tissue labels and expression prediction weights may not capture the relevant cell state or biology. Strong integrated evidence is therefore a basis for prioritizing follow-up, not a substitute for functional experiments or other evidence that can evaluate a specific causal mechanism.
Common Mistakes
- Equating a small TWAS p-value with colocalization. A TWAS association is not itself evidence that the expression and trait signals share a causal variant. Examine the colocalization hypotheses separately.
- Calling \(P(H_4\mid D)\) a mediation probability. It is a posterior probability for the shared-variant hypothesis under a specified model and prior, not the probability that expression causes the trait.
- Combining p-values as if independent. The results can use overlapping genetic data and are not independent votes. Report the evidence types side by side rather than multiplying their p-values.
- Ignoring the competing \(H_3\) hypothesis. A shared-signal claim needs to be interpreted against support for distinct signals, not from \(P(H_4\mid D)\) alone.
- Using a single-signal interpretation automatically. If multiple causal signals are plausible, use a suitable conditional or fine-mapping-based analysis rather than assuming one signal represents the whole locus.
Check Your Understanding
For each question, identify the evidence type and keep the conclusion within what the results support.
- A gene passes the corrected TWAS threshold, but colocalization gives \(P(H_3\mid D)=0.70\) and \(P(H_4\mid D)=0.20\). What can be concluded?
- For another gene, \(P(H_4\mid D)=0.60\) and \(P(H_3\mid D)=0.10\). Calculate the posterior odds of a shared rather than distinct signal and state what those odds mean.
- Why is it generally inappropriate to multiply a TWAS p-value by a colocalization p-value to create a combined significance measure?
- A locus may contain several causal signals, but the analysis assumes one. What is the main interpretive concern and what analysis class may be more suitable?
Solutions. (1) The gene has a significant TWAS association under the stated correction, but the colocalization results favor distinct over shared signals because \(0.70>0.20\). The TWAS result does not establish colocalization. (2) The odds are \(0.60/0.10=6\). Under the specified model and priors, the posterior probability assigned to \(H_4\) is six times that assigned to \(H_3\); this compares shared with distinct signals and is not a mediation probability. (3) The analyses can use overlapping genetic association information, so their results are not generally independent. Multiplication can misstate the combined evidence. (4) A single-signal model may fail to represent the locus and may distort the comparison of shared and distinct signals. Conditional or fine-mapping-based colocalization is more suitable when multiple signals are plausible.