PrediXcan And FUSION Modeling Choices

Two Decisions: How to Predict Expression and How to Test It

A transcriptome-wide association study (TWAS) uses genetic predictors of gene expression to test whether genetically predicted expression is associated with a trait. This involves two related but distinct decisions: how to estimate expression-prediction weights from an expression reference panel, and how to combine those weights with GWAS data to test an association. PrediXcan-style and FUSION-style analyses are useful ways to organize these choices, but they are not two wholly separate statistical ideas.

In a typical training panel, individuals have both measured gene expression and genotypes. A prediction model estimates which genetic variants predict expression of a gene in a specified tissue. The estimated weights are then used with trait data: either individual-level genotypes and phenotypes, or GWAS summary statistics and linkage disequilibrium (LD) information. The choice of prediction model affects which variants receive weights; the available GWAS inputs affect how the association test is computed.

This is the central distinction: the expression reference panel estimates prediction weights, while the GWAS data support the association test. The weight-estimation procedure is not itself evidence that the gene is associated with the trait. Earlier material on building expression-prediction weights and deriving a TWAS test statistic supplies the basic framework; here the focus is how common modeling and input choices differ.

Training Weights: Prediction Is a Model-Selection Problem

Suppose expression for one gene has been measured in \(n\) reference individuals, and genotypes for \(m\) cis variants are available. A basic linear prediction model is

Here \(E_i\) is the expression measurement for individual \(i\), \(X_{ij}\) is the coded genotype at variant \(j\), and \(\epsilon_i\) is the unexplained component. In practice, expression and genotypes are generally adjusted or residualized for relevant covariates, such as technical factors and ancestry-related variables, before or within model fitting. A fitted coefficient vector supplies prediction weights. The training observations must contain the required genotype and expression measurements, and the prediction target is expression variation that can be predicted from the included genotypes.

With many correlated variants and a modest expression sample, ordinary least squares may be unstable or not uniquely determined. Common choices include elastic net or LASSO regularization, which shrink coefficients and may set some to zero; ridge-like or BLUP approaches, which distribute signal across correlated variants; and a single-variant model, which uses only the strongest selected variant. FUSION-style weight training commonly compares a set of such models and selects or evaluates a model using cross-validation. PrediXcan-style analyses commonly use trained prediction models, often elastic-net weights, to form genetically predicted expression. The exact model library and selection procedure depend on the implementation and reference weights being used.

These methods make different assumptions about the architecture of predictable expression. A sparse model is useful when a small number of variants carry most predictive information, while a ridge-like model can be more stable when many variants have small effects. Cross-validation assesses predictive performance on held-out expression-panel individuals; it does not establish that the selected model is optimal in every population or that its gene-level trait association is causal.

Using the Weights With Trait Data

With individual-level data, one can calculate predicted expression for each GWAS participant from their genotypes and then test its association with the trait. The analysis requires participant genotypes at the model's variants, a trait measurement, and any covariates used in the association model. A linear regression is suitable for a continuous trait under the corresponding regression assumptions; other trait types require an appropriate outcome model. For a simple linear model, the gene-level coefficient describes the estimated change in trait per unit of predicted expression, conditional on the included covariates. Its units depend on the expression and trait scales.

With GWAS summary statistics, individual predicted expression values are not available. Instead, a summary-statistic method combines variant-level GWAS association estimates using the expression weights and the LD among the variants. FUSION-style testing is designed to support this route using summary statistics and an LD reference, as well as individual-level data in applicable workflows. PrediXcan-style testing is often described as regressing a trait on predicted expression in individual-level data; S-PrediXcan is a summary-statistic extension. Thus, it is inaccurate to treat the names as meaning that one can only use individual data and the other only summary data.

For a simplified summary-statistic calculation, let \(z_j\) be the GWAS \(z\)-score for variant \(j\), let \(w_j\) be its expression weight on a compatible genotype scale, and let \(R\) be the LD correlation matrix for the same variants. Under the null and the usual approximation that variant \(z\)-scores have covariance \(R\), the weighted score and its variance are

The denominator matters because correlated variants do not provide independent evidence. This calculation assumes compatible variant alleles and ordering, a suitably representative LD reference, and summary statistics whose scale and sampling properties support the approximation. If variants are missing, alleles are mismatched, or LD is poorly represented, the calculated statistic may be unreliable.

What the Result Means, and What It Depends On

A gene-level TWAS result reports evidence that genetically predicted expression, constructed from a specified weight model, is associated with the GWAS trait. It depends on the gene, tissue, variant set, training panel, prediction model, and association inputs. Different choices can yield different weights and therefore different results. A significant result does not mean that measured expression in the GWAS participants was directly tested, nor does it show that changing the gene's expression would change the trait.

Model performance is one diagnostic. Cross-validated correlation or prediction error in the expression panel indicates how well a model predicts expression for held-out members of that panel. A model with weak prediction may have little power for a trait association test, while a high prediction score does not ensure transportability to the GWAS population. Check that the gene has usable weights for the tissue of interest, that the relevant variants are present and harmonized, and that the LD reference corresponds reasonably well to the tested population.

PrediXcan-style individual-level regression directly constructs predicted expression for each person, which can make covariate adjustment and the modeled coefficient straightforward, but requires participant-level genotypes and trait data. Summary-statistic testing can be applied when only variant-level GWAS results are available, but depends on LD information and variant alignment. Within either route, changing the expression prediction model can change the weight vector and the result. The appropriate comparison is therefore not a universal winner, but a set of modeling and data requirements matched to the available evidence.

A common mistake is to say, “The gene is significant, so its expression causes the trait.” The test only establishes an association between the trait and a genetically weighted expression predictor under the specified model. Variants influencing the predictor may also relate to the trait through other pathways, and correlated variants can complicate which signal is being represented. A TWAS result is a gene-level association finding, not a causal conclusion.

The key limitation is that a weight vector is an estimated, population- and tissue-dependent predictor, not a direct measurement of expression in the GWAS sample. Results can be weak when expression is poorly predictable, when the reference and target populations differ, or when the relevant variants are absent or poorly imputed. These constraints make careful reporting of tissue, model, training source, variants, and testing inputs essential.