What a TWAS Associates
A transcriptome-wide association study (TWAS) tests whether a trait is associated with a gene-expression component predicted from genetic variants. The prediction is made using a model that links genotypes to expression, typically estimated in a reference sample where both are available. In the trait study, an individual's predicted expression is a genetic score calculated from their genotypes and the model's fixed weights.
This target is important to state precisely. A standard TWAS does not simply test whether expression measured in the trait-study participants is associated with their trait. It tests association with genetically predicted expression. That score is intended to represent the part of expression predictable from the included genetic variants under the prediction model. It is not a complete measurement of a gene's expression, and it need not capture all genetic influences on expression.
For a simple quantitative-trait analysis, let \(Y_i\) denote the trait value for person \(i\), and let \(\widehat{E}_i\) denote that person's predicted expression score for a gene. A basic association model is
where \(\alpha\) is an intercept, \(\mathbf{C}_i\) contains included covariates, and \(\varepsilon_i\) is the residual. The null hypothesis for this gene is \(H_0:\beta_{\mathrm{pred}}=0\); the alternative is that the trait has a nonzero association with the predicted-expression score, under the chosen model. For a binary trait, an appropriate generalized linear model can replace this linear model; the target remains the predicted-expression score.
The coefficient \(\beta_{\mathrm{pred}}\) describes change in the modeled trait per unit of predicted expression, conditional on the included covariates. Its units depend on how the score and trait are scaled. The coefficient is not automatically the effect of changing a person's actual expression, nor is it by itself evidence that expression mediates a genetic effect on the trait.
Why the Prediction Score Is a Different Variable
The expression-prediction model is trained using data that connect genetic variants to measured expression. Once that model is available, genotypes in a trait sample are combined with its weights to produce a score. For illustration, if the model assigns weights \(w_1,\ldots,w_m\) to genotype dosages \(G_{i1},\ldots,G_{im}\), then a predicted-expression score can be written as
This formula describes a weighted genetic score; it does not say that every contributor to measured expression is genetic. Measured expression may also reflect environmental exposures, cell composition, regulation not represented by the score, measurement error, and other influences. Consequently, \(\widehat{E}_i\) and expression measured from that same person are distinct variables and may differ substantially.
The score also depends on the fitted model and its allele conventions. Genotype dosages must be coded for the same alleles that the weights refer to. A mismatch can change a score's sign or value and undermine its interpretation. A score is therefore not an intrinsic, model-free measurement of expression: it is the output of a specified prediction model applied to specified genotype data.
Worked Example: Calculating a Predicted-Expression Score
Suppose a hypothetical, already fitted prediction model for one gene uses two genotype dosages. Its weights are \(w_1=0.30\) and \(w_2=-0.20\) score units per dosage. For one hypothetical person, the dosages of the alleles used to define those weights are \(G_{i1}=1\) and \(G_{i2}=2\). Calculate the score.
Using the model's weights and the correspondingly coded dosages,
The person's predicted-expression score is \(-0.10\) score units. It is a value from the genetic prediction model, not a direct expression measurement such as an RNA abundance value. If the second dosage were mistakenly coded for the opposite allele, it would be \(0\) rather than \(2\), producing \((0.30)(1)+(-0.20)(0)=0.30\). That change from \(-0.10\) to \(0.30\) illustrates why dosage and weight allele definitions must agree.
Interpreting an Association and Its Assumptions
The association test is meaningful only relative to the prediction model, the trait model, and the data used. The expression weights should come from an analysis that estimates genetic prediction of expression; the trait sample must have usable genotypes for the variants the prediction requires; and variant alleles and dosage coding must be aligned. The association model also relies on its own assumptions. For example, a linear model for a quantitative trait assumes the conditional mean is adequately represented by the specified linear terms, and inference requires an appropriate treatment of residual variation and dependence.
As in other genetic association analyses, population structure and other sources of confounding can distort associations if they are not handled adequately. Covariate adjustment should be appropriate to the study design, and the analysis should not treat a numerical score as immune to problems in the data that created it. In addition, the prediction model must be informative for the relevant gene and sufficiently applicable to the target sample. If the genetic component of expression is poorly predicted, the score may be a weak proxy for that component. The earlier material on diagnosing weak or missing eQTL evidence is relevant context: limited evidence for genetic prediction can limit what a downstream TWAS score represents.
A useful justification for the association target follows directly from the score's definition. If the weights are treated as fixed, then the covariance between the score and the trait is
This follows from the linearity of covariance and fixed weights. Thus, the score-trait association summarizes trait associations with a weighted combination of genotypes. It does not, by that algebra alone, establish that changing expression would change the trait. The genetic variants may be associated with the trait through other biological routes, and a prediction score can reflect correlated genetic signals.
Worked Example: A Simple Trait Association With Predicted Expression
Consider four hypothetical participants with predicted-expression scores \(x=(0,1,2,3)\) score units and a quantitative trait \(y=(1,3,5,7)\) trait units. For this illustration, use an unadjusted linear regression with an intercept. The mean score is \(\bar{x}=1.5\), and the mean trait value is \(\bar{y}=4\). The slope is
The fitted slope is \(2\) trait units per score unit. The intercept is \(\bar{y}-\widehat{\beta}_{\mathrm{pred}}\bar{x}=4-(2)(1.5)=1\), so the fitted values for the four scores are \(1,3,5,7\). In this small hypothetical data set, the trait is perfectly linearly associated with the score. In a real analysis, uncertainty, appropriate covariates, and the model's inferential assumptions would also need to be addressed; four observations do not provide a reliable basis for broad biological claims.
The coefficient describes an association with the score as defined. It does not mean that increasing measured expression by one unit would cause the trait to increase by two units. Even if the score is designed to predict expression, the trait analysis in this example did not use measured expression as its predictor.
Worked Example: Predicted and Measured Expression Can Differ
Suppose, in a hypothetical sample, the predicted-expression scores are \(x=(0,1,2,3)\), measured expression values are \(e=(2,3,2,5)\) expression units, and trait values are \(y=(1,3,5,7)\) trait units. The score-trait regression slope is \(2\), as calculated above. Now calculate the slope from a separate regression of the trait on measured expression to see why the targets should not be conflated.
Here \(\bar{e}=3\) expression units and \(\bar{y}=4\) trait units. The slope using measured expression is
The measured-expression slope is approximately \(1.333\) trait units per expression unit, whereas the predicted-expression slope is \(2\) trait units per score unit. These numerical slopes are not directly comparable because their predictor units differ, and this tiny example does not support a significance claim. Its purpose is to show that the two regressions use different predictors and need not yield the same association. A TWAS result should therefore be described as an association with predicted expression, not silently substituted for an association with measured expression.
What a Significant TWAS Result Does and Does Not Say
A small p-value for a gene's TWAS test is evidence against the specified null of no association between the trait and that gene's predicted-expression score, subject to the model and assumptions. It is not the probability that the null is true. Nor does it establish that the gene is the causal gene at a locus, that the gene's expression mediates a genetic effect, or that the trait would change under an intervention on expression.
A TWAS signal can arise because the predicted score captures genetic variation associated with the trait, but that association may have different biological explanations. For example, a variant may influence the trait through another molecular process, or correlated variants may contribute to both expression prediction and trait association. A genetic score association is a useful functional clue, not a stand-alone causal conclusion. Evidence such as colocalization addresses a different question about shared genetic signals; a TWAS association alone does not answer that question.
The main alternatives answer related but distinct questions. Direct expression-trait analysis tests association with expression measured in the participants and is subject to the limitations of that measurement and study design. An eQTL analysis tests genotype-expression association. TWAS instead tests trait association with a model-based expression score. These approaches can agree or disagree because their predictors, samples, and estimands differ.
| Analysis | Primary association target | Interpretive focus |
|---|---|---|
| Direct expression-trait analysis | Measured expression and trait | Whether observed expression covaries with the trait in the analyzed sample |
| eQTL analysis | Genotype and measured expression | Whether genetic variation is associated with expression |
| TWAS | Predicted expression and trait | Whether a genetically derived expression score is associated with the trait |
These methods are not interchangeable. In particular, a measured-expression association is not a substitute definition for TWAS, and a TWAS association does not by itself demonstrate an expression-mediated effect. Each result should be reported with its actual target and units, and interpreted in light of the prediction model and trait-analysis assumptions.
Diagnostics and Common Pitfalls
Before interpreting a TWAS result, check that the genotype variants required for the score were available and aligned to the weight alleles, and that the score was calculated as intended. Inspect whether the model is informative for the gene and whether its use in the target sample is reasonable. For the trait association, examine whether the chosen model and covariates are suitable, and whether unusual observations or data-quality problems could dominate the result. A score that varies little across the sample, or that is poorly predicted in the relevant setting, offers limited information even if a coefficient can be computed.
A common mistake is: “The TWAS is significant, so measured expression of this gene causes the trait.” This reasoning changes the tested predictor and turns association into causation. The correction is: the test indicates an association between the trait and the genetically predicted-expression score under the fitted model. Further evidence is needed to distinguish causal mediation from other explanations.
Another pitfall is to compare a TWAS coefficient directly with a coefficient from measured expression without checking their units and definitions. The score may be standardized or otherwise scaled, whereas measured expression may use a different scale. Report what predictor was used, how it was scaled, and the trait units; do not imply that the coefficients describe the same contrast unless that has been established.
Check Your Understanding
Answer each question, then compare your reasoning with the brief solutions below.
- In one sentence, state the association target of a TWAS and contrast it with measured expression.
- A gene has a TWAS slope of \(-0.4\) trait units per score unit. What does the slope describe, and what does it not establish?
- A researcher finds a significant TWAS result and concludes that experimentally increasing the gene's expression will lower the trait. Identify the interpretive error.
- Why must genotype dosages be coded for the same alleles used to define prediction weights?
Solutions. (1) A TWAS tests association between the trait and expression predicted from genotypes using a specified model; measured expression is a separate variable. (2) It describes the modeled trait difference per one-unit difference in predicted-expression score, conditional on included covariates; it does not establish an effect of intervening on actual expression. (3) The researcher has treated association as causation and substituted a genetic score for measured expression. (4) The weights refer to particular alleles, so coding the opposite allele can change a score's direction or value and invalidate the intended prediction.