Pathway-Level Hierarchical Priors for Omics Effects

Why Put a Prior at the Pathway Level?

An omics analysis may estimate an effect for each gene, protein, metabolite, or other feature. A pathway database provides a way to group related features, but pathway membership does not imply that all members have identical effects. A useful prior can express a weaker claim: members of the same pathway may tend to have effects near a shared pathway-level center, while individual features retain room to differ.

This is partial pooling. A feature estimate is informed both by its own data and, through the shared center, by the estimates of other pathway members. The amount of borrowing depends on the assumed variability among features in the pathway. The prior describes a statistical structure; it does not establish that pathway membership causes an effect or that the pathway is active.

A Hierarchical Prior for Feature Effects

Let \(g\) index pathways and \(j\) index features within pathway \(g\). Let \(\beta_{gj}\) be the effect for feature \(j\) in pathway \(g\), measured here as a log-fold change. For a simple model, assign each feature to one pathway, and assume that pathway assignments are fixed before analysis. Let \(\theta_g\) be the pathway-level center. A Gaussian hierarchical prior is

The global center \(\mu_0\) is the prior average across pathway centers; \(\tau^2\) describes variation among pathway centers. The variance \(\omega^2\) describes feature-to-feature variation around a pathway center. Equivalently, write \(\beta_{gj}=\theta_g+\delta_{gj}\), where \(\delta_{gj}\sim N(0,\omega^2)\). The deviations are taken to be independent across features and independent of the pathway centers. Unless specified otherwise, the hyperparameters \(\mu_0,\tau^2,\omega^2\) are fixed in the calculations below.

The two variance components have distinct meanings. A small \(\tau^2\) concentrates pathway centers near the global center. A small \(\omega^2\) concentrates member effects near their own pathway center. Setting \(\omega^2=0\) makes all effects within a pathway identical; this is a hard equality constraint, not merely strong partial pooling. A very large \(\omega^2\) makes the pathway center weakly informative about individual features.

The base model assumes nonoverlapping pathway assignments, a common \(\omega^2\), and a common distribution for pathway centers. Features with no pathway assignment need a separately specified background prior. Overlapping pathways cannot be handled by assigning a feature to several independent copies of this same model: doing so would give that feature multiple effects. An explicit model of the overlapping memberships and their shared contributions would be needed instead.

Dependence Implied by the Prior

The shared center creates a specific covariance structure. For a feature in pathway \(g\), the law of total variance gives the marginal variance as the variance of its conditional mean plus the expected conditional variance. For two different features in that pathway, conditional independence given \(\theta_g\) means their covariance comes from the variation in that shared center.

For example, effects in a pathway are positively correlated under this prior when \(\tau^2>0\), but they are not identical when \(\omega^2>0\). The prior correlation is a property of the specified model, not a correlation estimated from the observed data. For features in different pathways, the base model gives covariance zero, assuming independent pathway centers and deviations.

Updating the Pathway Center and Feature Effects

Suppose the available data for feature \(j\) are a noisy effect estimate \(y_{gj}\), with known sampling variance \(v_{gj}\). For this conjugate example, assume the sampling model \(y_{gj}\mid\beta_{gj}\sim N(\beta_{gj},v_{gj})\), with estimates conditionally independent given the feature effects. This normal approximation is appropriate when the upstream effect estimates and their standard errors are adequately summarized by an approximately normal sampling distribution. The prior variances are treated as known here; uncertainty about them is not included in these formulas.

Integrating over a feature effect gives \(y_{gj}\mid\theta_g\sim N(\theta_g,\omega^2+v_{gj})\). Thus each estimate informs the center with variance equal to its measurement variance plus the feature-deviation variance. For a single pathway, define \(q_{gj}=\omega^2+v_{gj}\). Combining these normal observations with \(\theta_g\sim N(\mu_0,\tau^2)\) yields

These are the posterior variance and mean of the pathway center. Conditional on \(\theta_g\) and its own observation, the feature effect has posterior variance \(r_{gj}=(1/\omega^2+1/v_{gj})^{-1}\) and mean \(r_{gj}(y_{gj}/v_{gj}+\theta_g/\omega^2)\). If uncertainty in the pathway center is retained, the marginal posterior for the feature is also normal, with

Here \(\mathbf y_g\) denotes all observed estimates in pathway \(g\). The variance includes both the uncertainty remaining about the feature after conditioning on the center and the contribution of uncertainty in that center. Ignoring the second term would understate posterior uncertainty.

Conditional Shrinkage Is Not the Full Hierarchical Update

A simpler calculation holds \(\theta_g\) fixed. It illustrates how an individual estimate is pulled toward a specified center, but does not account for learning that center from the data. This distinction matters whenever a reported posterior is intended to represent uncertainty in both the feature effect and the pathway center.

Assumptions, Diagnostics, and Alternatives

The model is most interpretable when the effects are on a comparable scale, pathway assignments are relevant to the biological context, and the sampling variances are credible. Its conditional independence assumptions are also consequential: after accounting for a pathway center, feature deviations are assumed independent, and sampling errors are assumed independent given the effects. Correlated estimates or shared technical artifacts can create dependence not represented by this prior.

Warning signs include estimates for pathway members that are systematically more dispersed than the assumed \(\omega^2\) allows, highly influential pathway members shifting the center, and posterior summaries that change substantially under modest changes to \(\tau^2\) or \(\omega^2\). Examine the number of members and the uncertainty of their estimates; a center based on one or two noisy features may remain strongly influenced by its global prior. Also check whether pathway membership is ambiguous or overlapping. A mathematically precise posterior does not repair an unsuitable pathway definition or an incorrect sampling model.

The main alternatives answer different questions. Independent feature priors omit pathway borrowing and can be preferable when pathway labels are unreliable or irrelevant. A hard pathway constraint sets every member effect equal to the pathway center, eliminating feature-specific deviations. The hierarchical prior lies between these choices: it shares information but allows heterogeneity. A model with pathway-specific deviation variances can accommodate differing heterogeneity, but requires additional assumptions and information to estimate those variances reliably. No option is universally best; the choice should reflect the scientific claim and the available data.

The Gaussian specification is convenient and yields the closed-form updates shown here, but may be a poor description when effect distributions are strongly skewed, heavy-tailed, or dominated by outliers. The base model also treats pathway membership as known and does not represent pathway overlap, context-dependent effects, or uncertainty in the hyperparameters. These limitations should be made explicit rather than hidden behind the phrase “pathway-informed.”