Feature Loadings As Omics Signatures

What a Loading Says About a Feature

A latent factor is not directly measured, so its meaning is inferred from how it relates to observed features. In a multi-omics model, a factor may have one loading for each feature in each view: those loadings summarize which measurements are associated with movement along that latent direction. A collection of loadings can therefore be read as a candidate omics signature. It is a model-based description of the data, not by itself evidence of a biological mechanism or a causal effect.

Let \(Y_{ij}^{(v)}\) denote the measurement for sample \(i\), feature \(j\), and omics view \(v\). Let \(z_{ik}\) be sample \(i\)'s coordinate on factor \(k\), and let \(\lambda_{jk}^{(v)}\) be the loading of feature \(j\) in view \(v\) on that factor. The loading tells how the model's expected response for that feature changes as the factor coordinate changes, with the other factor coordinates held fixed. The exact response scale depends on the observation model.

For a continuous feature, a common conditional-mean model is

Here \(\mathbf{z}_i\) is the vector of factor coordinates for sample \(i\), and \(\mu_j^{(v)}\) is a feature-specific intercept. The usual interpretation assumes the conditional mean is correctly specified and the residual has conditional mean zero. A positive loading means that the expected measurement increases as the factor coordinate increases, holding the other coordinates fixed; a negative loading means it decreases. The loading is measured in the feature's units per unit of factor coordinate.

For a count feature with a log link, the linear predictor may instead be

For a binary feature, a logistic observation model may use

These examples rely on the relevant observation model being appropriate for the feature and on its link function being correctly specified. For count and binary features, a loading is a change on the linear-predictor scale, not a fixed change in the observed mean or probability. Thus loading signs have a consistent direction on the link scale, but the corresponding change in observed values depends on the feature's baseline and the size of the factor-coordinate change.

Magnitude, Direction, and Units

The absolute value \(|\lambda_{jk}^{(v)}|\) describes the size of the modeled association on the model's linear-predictor scale. Its practical meaning requires the units of both the feature and the factor coordinate. In a continuous model, for example, a loading of \(0.8\) means an expected increase of \(0.8\) feature units for a one-unit increase in the factor coordinate, with other factor coordinates held fixed. It does not mean an increase of \(0.8\) standard deviations unless the variables were scaled to make that so.

Raw loading magnitudes generally should not be compared across features measured in different units or across views with different preprocessing. A loading of \(3\) may be large for a feature measured in small units and modest for one measured in large units. Even within a view, the interpretation depends on whether features were centered or scaled before fitting. A standardized comparison can be useful for continuous features: if \(s_{Y_j^{(v)}}\) is the observed standard deviation of a feature and \(s_{z_k}\) is the standard deviation of its factor coordinate, then

This quantity represents the modeled change in the feature, in standard-deviation units, associated with a one-standard-deviation change in the factor coordinate under the continuous linear-mean model. It helps compare continuous features on a common scale, but it does not make fundamentally different observation models equivalent. For count and binary features, report a link-scale contrast or transform a specified coordinate change back to the mean or probability scale.

A loading's sign gives direction only relative to the chosen coding of the factor coordinate. If the coordinate is reversed, the loadings' signs reverse as well while the model's fitted relationships can remain unchanged. Therefore, statements such as “high factor values are associated with higher expression” require a stated factor orientation. The sign indicates a model association, not whether a feature increases over time or causes another feature to change.

Signatures in Count and Binary Views

For count and binary features, a useful interpretation starts with a change in factor coordinate and then translates the loading through the link function. In a log-mean model, a change \(\Delta z_k\) changes the conditional mean by a multiplicative factor \(\exp(\lambda_{jk}^{(v)}\Delta z_k)\), with other coordinates fixed. This is a ratio of expected counts, not an additive change in counts. In a logistic model, the same expression is a ratio of conditional odds, not a difference in probabilities.

Sparsity and the Shape of a Signature

A factor signature may be dense, with many features having appreciable loadings, or sparse, with most features having zero or small loadings and a limited set carrying most of the pattern. If the fitted model constrains a loading to be exactly zero, that feature has no modeled contribution from that factor in that parameterization. A small but nonzero estimate is not the same as an exact zero. Whether it is substantively negligible depends on measurement scale, uncertainty, and the purpose of the analysis.

In practice, a signature is often summarized by listing features with the largest absolute loadings, while retaining their signs and the units or scale used. A threshold can be used to label loadings as “large enough to report,” but it is a descriptive rule, not automatic proof that selected features are important or that unselected ones have no association. Report the threshold and consider whether the listed features remain prominent under reasonable refitting or perturbations of the data.

Justifying the Conditional Interpretation

The linear-mean interpretation follows directly from the model, and it clarifies why loadings describe conditional associations rather than simple one-feature correlations. For a continuous feature, compare two factor-coordinate vectors that differ only in coordinate \(k\), by \(\Delta z_k\). Under the specified model,

Here \(\mathbf{e}_k\) is a vector with value one in coordinate \(k\) and zero elsewhere. The difference is exactly the loading times the coordinate change because all other coordinates are unchanged. The same subtraction on a log-mean or log-odds scale gives \(\Delta z_k\lambda_{jk}^{(v)}\) on that link scale; exponentiating gives the corresponding multiplicative ratio. If the factor coordinates are correlated, the loading remains a conditional model coefficient. It need not equal the marginal association between that feature and the factor's observed estimates.

Comparisons, Diagnostics, and Limitations

A loading signature is one of several useful summaries. A ranked list of loadings is compact and preserves direction, but it can hide whether remaining features have moderate loadings. A heat map of loadings across features, factors, and views can show sign patterns and concentration, but its colors are misleading if each view uses an undocumented scale. A thresholded signature is easy to communicate but discards information near the cutoff. Whichever summary is used, state the scale, sign orientation, and any threshold.

Before assigning substantive meaning, examine whether the interpretation is supported by the fitted model and the data. Useful checks include:

• Confirm that the feature's observation model and link match the recorded data and the intended interpretation.
• Check the preprocessing and units. Centering, scaling, transformations, and offsets affect what a loading means.
• Look for unusually large loadings that are driven by a small number of samples or by a feature with extreme values.
• Check whether the leading features and their directions are reasonably stable under refitting or modest data perturbations.
• Consider technical structure, such as batch or assay effects, that could produce a coherent signature unrelated to the scientific interpretation.
• For count and binary features, translate coefficients at relevant baseline means or probabilities rather than presenting a link-scale number as an observed-scale change.

Instability, poor fit, or a signature dominated by technical structure weakens the interpretation. Loadings can summarize conditional patterns under the fitted model, but the model may be misspecified, features may be correlated, and a reported loading may have substantial uncertainty. Furthermore, a factor's signs and scales depend on how its coordinate is represented. Interpret direction relative to the stated orientation, and do not compare raw magnitudes across views unless the units and scaling make that comparison meaningful. Shared signatures across views can motivate biological hypotheses, but they do not establish shared mechanism or causation.