Expression Prediction Weights for TWAS

Why Expression-Prediction Weights Matter

A TWAS expression predictor is a model that uses nearby genetic variants to estimate an individual's genetically predicted expression for a gene in a particular tissue. The model is trained using people for whom both genotype and expression measurements are available. Its fitted SNP coefficients, or prediction weights, are then applied to genotype data in a separate association analysis.

These weights are not interchangeable labels for a gene. They depend on the training sample, tissue, expression processing, variant set, genotype coding, and prediction method. A gene can be tested with a particular model only when the required genotype information is available and the model predicts expression adequately under the chosen validation criteria. These conditions concern whether a usable predictor exists; they do not by themselves establish that the gene is associated with a trait.

How a Predictor Is Trained

The training data contain expression measurements and genotypes for a set of individuals. For a gene in a specified tissue, let \(Y_i\) denote the processed expression measurement for individual \(i\), and let \(G_{ij}\) denote the genotype dosage for variant \(j\). A dosage commonly counts copies of a designated allele and takes values 0, 1, or 2. A cis set is a defined collection of variants in or near the gene; its genomic boundaries and quality-control rules should be specified rather than assumed to be universal.

Expression is usually processed before fitting. This may include normalization and adjustment for technical or biological covariates, such as batch or measured cell composition, when appropriate to the training design. Genotypes also require quality control and a fixed allele coding. The fitted model therefore predicts the processed expression phenotype used in training, not an abstract, universally scaled quantity. If expression is standardized in the training data, predicted values are on that standardized scale; if it is not, predictions retain the model's expression units.

A simple model estimates the weights by minimizing squared prediction errors. With many correlated cis variants and a modest expression sample, an unpenalized regression can be unstable. Common approaches therefore use regularization, such as ridge regression, the lasso, elastic net, or Bayesian shrinkage methods. These approaches differ in how they shrink or select weights, but all produce a predictor tied to the chosen variants and training data. The tuning choices, including any penalty strength, should be made using training data alone, typically with cross-validation.

A fitted weight is a predictive coefficient, not automatically a causal effect of the variant on expression. Correlated variants can share or exchange predictive signal, particularly when the model uses shrinkage or selection. A nonzero weight does not, by itself, prove that its variant directly regulates the gene.

What the Weights Mean—and Do Not Mean

For an ordinary linear regression with a single centered genotype predictor \(G\), a centered outcome \(Y\), and no other predictors, the least-squares coefficient has a direct algebraic interpretation. It is the coefficient that minimizes the sum of squared residuals. Differentiating that sum gives

The denominator is positive when the genotype varies in the training sample. Thus the coefficient is determined by the joint pattern of genotype and expression in that sample. With multiple variants, the corresponding regression accounts for their joint relationships; with regularization, the fitted coefficients also reflect the penalty. In either case, the coefficient's units depend on the genotype and expression scales. If \(G_j\) is an unstandardized allele dosage and \(Y\) is expression in a specified unit, \(w_j\) describes the fitted change in predicted \(Y\) per one-dosage increase, conditional on the model's other inputs. If either variable is standardized, the interpretation changes accordingly.

For a particular individual, the predictor adds the contributions from all included variants. This gives a model-based expression score, not a direct measurement of that individual's expression. In applications, allele orientation and genotype coding must match the specification used to create the weights. A weight for the count of one allele cannot be applied as though it were a weight for the alternate allele without an appropriate recoding.

Validation: Measuring Prediction Beyond the Training Sample

A model's fit in its training data is not sufficient evidence that it predicts expression well in new individuals. Flexible models can fit training-specific noise, and tuning a model on the same observations used to report its performance can make that performance optimistic. Prediction should therefore be evaluated using observations not used to fit or tune the model, for example a held-out validation sample or an appropriately designed cross-validation procedure.

One common summary is the coefficient of determination in the validation sample, calculated against a specified baseline. A simple version uses the validation-sample mean as the baseline and is \(R^2_{\mathrm{val}}=1-\sum_i(Y_i-\widehat{Y}_i)^2/\sum_i(Y_i-\bar{Y}_{\mathrm{val}})^2\). Here \(\bar{Y}_{\mathrm{val}}\) is the validation mean. This value can be negative when predictions perform worse than the baseline. Correlation between observed and predicted expression is also reported in some settings, but correlation alone does not assess calibration: predictions can track variation while being systematically too large or too small.

Validation performance is specific to the gene, tissue, sample, and processing choices. It is also affected by the similarity between training and target populations, including allele frequencies and linkage disequilibrium patterns. A threshold for accepting a predictor is a study or analysis choice, not a universal law. The rule should be set and reported clearly; different thresholds can produce different sets of testable gene–tissue pairs.

Which Gene–Tissue Pairs Can Be Tested?

A gene–tissue pair is eligible for a TWAS analysis using a particular prediction resource only if the corresponding expression model is available and can be applied to the analysis genotypes under the model's variant and allele conventions. In practice, eligibility depends on several linked checks:

• Model availability: weights exist for the specified gene and tissue, and the training and processing details are sufficiently documented.
• Prediction evidence: the model meets the prespecified validation criterion, if the analysis requires one.
• Variant correspondence: required SNPs can be matched to the analysis data, with compatible identifiers, genome build, and alleles.
• Usable genotype information: allele coding is aligned, and missing variants are handled only by a documented, appropriate procedure.

The exact variant set matters. A predictor trained on one cis window is not automatically the same predictor as a model for the same gene trained on a wider window or a different set of quality-controlled variants. Nor is it generally valid to discard unavailable variants and retain the remaining weights unchanged: the model's prediction rule has then changed, and its validation performance no longer necessarily applies. If an analysis uses a reduced SNP set, that choice needs a defensible model-specific procedure and appropriate performance assessment.

Training sample size alone does not determine whether a model is useful. The number of cis variants, their correlation structure, expression variability, tissue relevance, and validation design also matter. Conversely, a high in-sample fit is not a substitute for out-of-sample validation. The goal is not simply to produce a list of nonzero weights, but to establish a defined, transferable prediction rule that can be applied consistently.

Common Mistakes and a Practical Workflow

Several errors can make a correctly calculated score scientifically misleading. The most important are treating a weight as a causal effect, assuming all prediction scores are in standard-deviation units, and applying weights to allele dosages coded in the opposite direction. A fourth error is reporting training fit as though it were validation performance. These problems can be prevented by preserving the full model specification: outcome processing and scale, SNP list, allele coding, centering or scaling, fitted weights, intercept, and validation procedure.

The central distinction is between training a prediction rule and testing an association using that rule. Training produces weights and a defined expression scale; validation assesses prediction in held-out data; variant matching determines whether the rule can be applied. A usable predictor is a necessary input for a TWAS gene test, but it is not itself evidence that the predicted expression is associated with a trait.