Interaction-Based Priors for Molecular Network Edges

Why Put Interaction Knowledge in an Edge Prior?

An integrated molecular network represents relationships among measured entities, such as genes, proteins, and metabolites. The data may support an edge between two entities, but the strength of that evidence can vary. A known molecular interaction can provide useful prior information; it should not, however, force the model to include the edge. A probabilistic prior expresses the distinction between an edge that is supported before examining the current integrated data and an edge that is newly inferred from those data.

The goal is to let interaction knowledge influence the starting probability of each candidate edge while preserving the possibility that the data will either weaken that edge or support an edge not listed in a database. This is a soft-prior approach: the model can revise its beliefs as it sees the data. The earlier material on representing pathway databases as model priors and calibrating prior strength from interaction confidence provides the general background; here the prior is attached to individual network edges.

The candidate-edge set and meaning of an edge must be specified. For example, an undirected edge may represent a modeled association between two entities, not a causal effect or a directed regulatory relationship. The prior concerns edge inclusion under that model. It does not establish biological truth, mechanism, or direction.

From Interaction Evidence to a Probability

A database may report a confidence score, evidence category, or number of supporting experiments. Such a score is not automatically a probability that an edge exists. A score can be used to define an edge-specific prior only after its interpretation is made explicit. One defensible strategy is to calibrate scores against independent reference data, estimating how often edges in a score range are verified. Another is to use a prespecified mapping from evidence categories to prior probabilities, with the mapping justified and checked for sensitivity.

Calibration data should be independent of the omics data used for inference, and the reference set should be relevant to the candidate edges. If calibration is performed using the same observations later included in \(Y\), the evidence can be counted twice. Database absence also needs care: an unlisted interaction may be untested, poorly measured, or absent from the database for reasons unrelated to whether it exists. Unless database coverage supports a meaningful negative interpretation, absence should not be treated as proof of no edge.

A simple prior assigns separate probabilities to database-supported and unsupported candidate edges. A richer mapping can allow probabilities to vary with calibrated score. In either case, probabilities should remain strictly between zero and one if the model is intended to allow data-driven inclusion and exclusion. Assigning probability zero or one makes the edge impossible to revise.

Updating Edge Probabilities with Integrated Data

Let \(Y\) denote the observed data across the included omics layers. For a candidate edge, compare two models: \(M_1\), in which \(Z_{ij}=1\), and \(M_0\), in which \(Z_{ij}=0\). The data evidence is summarized by a Bayes factor, the ratio of marginal likelihoods under these models. Each marginal likelihood integrates over the model’s unknown parameters rather than evaluating them at a single convenient value.

The odds update follows directly from Bayes’ rule. For brevity, write \(Z=Z_{ij}\) and \(\pi^0=\pi^0_{ij}\). Then

If the posterior probability is denoted by \(\pi^1_{ij}\), converting the posterior odds back to a probability gives

This formula makes the separate roles clear: interaction knowledge sets the prior odds, while the current data contribute the Bayes factor. A strong prior can be overcome by sufficiently strong contrary data, and a low prior can be overcome by sufficiently strong data in favor of inclusion.

Distinguishing Supported Edges from Newly Inferred Edges

An edge can be called prior-supported when external interaction evidence gave it an elevated prior relative to the baseline used for unsupported candidate edges. A newly inferred edge is an edge not supported by the included prior source, but whose integrated-data evidence raises its posterior probability. These labels describe how the model arrived at its probability; neither label alone is a final decision rule.

A probability does not by itself determine which edges should be reported. A decision threshold depends on the consequences of false inclusion and false exclusion. If the cost of including a nonexistent edge is \(C_{FP}\), and the cost of excluding a real edge is \(C_{FN}\), the expected losses are \(C_{FP}(1-\pi^1_{ij})\) for inclusion and \(C_{FN}\pi^1_{ij}\) for exclusion. Inclusion has lower expected loss when

This threshold is a decision rule under stated costs, not a universal standard for declaring biological interactions. When many edges are considered, uncertainty and multiplicity across the network also matter; reporting posterior probabilities and the decision rule is more informative than presenting a selected edge list without its criteria.

Model Assumptions and Diagnostics

The odds update is exact for the specified two-model comparison when the prior probability and marginal likelihoods are coherent. Applying it edge by edge requires additional care. In a network with many edges, one edge can affect the likelihood for another. The relevant Bayes factor may therefore be conditional on the status of other edges, or the model may estimate the full graph jointly. Treating all edge indicators as independent Bernoulli variables is a simplifying prior, not a general fact about biological networks.

A defensible analysis should make the following assumptions explicit:

• The candidate entities, candidate-edge set, and meaning of edge inclusion are defined before interpreting results.
• External interaction evidence is available independently of the current omics observations, or any dependence is modeled so that evidence is not counted twice.
• The mapping from evidence to prior probability is calibrated or justified, and its uncertainty is not hidden when it is substantial.
• The likelihood distinguishes the edge-present and edge-absent models appropriately and accounts for relevant measurement and model uncertainty.
• The data used to calculate the Bayes factor are compatible with the modeled entities and edge definition.

Diagnostics should examine whether conclusions depend heavily on the chosen prior mapping, whether known-supported edges are almost automatically retained, and whether unsupported edges can still emerge when the data are informative. Review the distribution of prior and posterior probabilities, compare prior-supported and unsupported candidates, and inspect whether results are driven by a small number of database entries. Sensitivity to plausible alternative priors is especially important when calibration data are sparse. A posterior probability close to one can still be misleading if the likelihood is misspecified or the prior evidence has leaked information from the same data.

Alternatives, Limitations, and a Common Mistake

A uniform prior assigns the same inclusion probability to every candidate edge. It is simpler and avoids relying on external interaction databases, but it does not use differences in prior evidence. A hard constraint fixes selected edges as present or absent; this can be appropriate only when the scientific model truly requires those statuses, because the data cannot revise a probability of zero or one. Edge-specific soft priors occupy the middle ground: evidence can influence the model without making an edge mandatory.

Interaction-based priors inherit limitations of their sources. Coverage may favor well-studied molecules, confidence may not be comparable across evidence types, and a reported interaction may not apply in the biological setting under study. A prior can also make a network appear to confirm knowledge that was already encoded in it. These concerns do not make priors unusable; they make calibration, provenance, sensitivity checks, and clear separation of prior support from data evidence essential.

Common mistake: “This edge is absent from the database, so its prior probability should be zero; any posterior support must therefore indicate a model error.” This reasoning confuses not being recorded with being impossible. With \(\pi^0_{ij}=0\), the posterior remains zero for every finite Bayes factor, as the edge has been excluded by assumption. For an unlisted but plausible edge, use a nonzero baseline prior and let the data update it. Reserve a zero prior for a genuine structural impossibility in the model, not for missing database coverage.