Hub Claims and Degree Bias in Omics Networks

Why a High Degree Is Not Yet Evidence of a Biological Hub

A node with many estimated connections may be biologically important, but its degree is also shaped by how the network was measured and fitted. A feature measured more reliably can have more detectable associations; a small sample can produce unstable edge selections; and a different penalty or candidate-edge set can change which connections appear. A degree ranking is therefore a property of an estimated network under specified choices, not direct evidence that one molecule regulates many others.

Here, consider an undirected network of \(p\) omics features. An edge represents the conditional association specified by the network model, not necessarily a causal effect. The graph may represent, for example, conditional dependencies in a Gaussian graphical model. The same caution applies to other estimated graphical models, although their edge-selection mechanisms differ. The earlier material on sensitivity analysis for network structure motivates comparing conclusions across reasonable modeling choices; this tutorial focuses on interpreting node degree.

The target graph is usually unknown, so its degree cannot simply be read off the data. The estimated degree combines correctly selected edges with false positives and missed connections. Moreover, if some possible edges are excluded before fitting, the reported degree cannot count them. Comparisons should therefore state the candidate set, estimation method, tuning choices, and whether degree is raw or normalized.

How Edge Errors Translate into Degree Bias

Assume that the candidate set \(C_i\) is fixed for the analysis. For each eligible true edge, define its selection probability \(s_{ij}=P(\hat A_{ij}=1\mid A_{ij}=1)\). For each eligible absent edge, define its false-selection probability \(q_{ij}=P(\hat A_{ij}=1\mid A_{ij}=0)\). These probabilities refer to repeated samples under the specified data-generating process and fitting procedure; they are not probabilities that an edge is biologically real after observing it.

Write the target degree within the candidate set as \(d_i(C)=\sum_{j\in C_i}A_{ij}\). Because \(\hat A_{ij}\) is binary, its expectation is its probability of being selected. Summing those expectations gives the expected estimated degree:

Subtracting \(d_i(C)\), which counts the candidate true edges, yields a useful bias decomposition. The first sum loses \(1-s_{ij}\) for each true edge that is missed; the second adds \(q_{ij}\) for each absent candidate edge that is selected.

This identity does not require edge selections to be independent. Independence would be needed for a simple variance calculation, but not for the expectation. Crucially, both sums are over eligible edges in \(C_i\): an excluded edge has no selection probability in this setup and contributes to neither sum. If the scientific target is degree in the full graph rather than in the candidate graph, omitted true edges are an additional source of undercounting.

Sample Size, Variability, and the Opportunity to Detect Edges

With limited sample size, estimates of associations and conditional associations can be noisy. That can lower selection probabilities for real but modest edges and raise false-selection probabilities for absent edges. Increasing sample size often improves estimation under suitable model and regularity conditions, but it does not guarantee a larger or smaller estimated degree in every dataset. The effect depends on signal strengths, the estimator, the tuning rule, and whether the target is defined consistently.

A further issue is that nodes may have different numbers of eligible connections. A node with 20 candidate neighbors has more opportunities to accumulate selected edges than one with five. Raw degree alone is not a fair comparison when candidate sets differ. The proportion \(\hat d_i(C)/|C_i|\) can describe the fraction of eligible edges selected, but it answers a different question from raw degree and does not correct for differences in edge detectability.

Feature variability matters through the quality and scale of measurements. In a Gaussian graphical model, conditional dependence is often represented by off-diagonal entries of a precision matrix, or by partial correlations derived from it. A pure change of measurement units should not change a correctly standardized partial-correlation target. But low biological variability relative to measurement error can reduce reliability, making associations harder to estimate; zero or near-zero variance can also make standardization unstable or impossible. For count and other non-Gaussian measurements, prevalence, dispersion, and model fit likewise affect how much information is available for a particular edge.

For example, suppose two latent standardized biological measurements have correlation \(0.60\), and each observed value is the biological value plus independent measurement error. If their reliabilities are \(0.138\) and \(0.962\), the observed correlation is approximately \(0.60\sqrt{0.138\times0.962}=0.219\), under the classical independent-error model. A threshold of \(0.30\) would not select this association, even though the latent correlation is \(0.60\). This illustration concerns a marginal correlation, not a general formula for partial correlations in every network estimator; it shows why measurement quality can change edge detectability without changing the underlying biology.

Regularization and the Meaning of a Hub

Sparse network estimators use regularization to control complexity. For example, a Gaussian graphical lasso estimates a precision matrix while penalizing off-diagonal entries; entries shrunk to zero correspond to absent edges in the fitted graph. Stronger penalization usually yields a sparser graph, but individual node degrees need not change monotonically along a tuning path. Edges can enter or leave as the penalty changes, and node rankings can consequently shift.

Regularization does not affect every node equally in practice. Differences in feature scale, missingness, measurement reliability, correlation patterns, or node-specific penalties can alter selection behavior. Even when a common penalty is used, a node’s estimated degree depends on the joint fitting problem and on the other features included. A high-degree node may reflect a broad set of modest associations, a cluster of correlated measurements, or model misspecification; degree alone cannot distinguish these explanations.

Degree is also a coarse summary: it gives equal weight to every selected edge and ignores the magnitude and uncertainty of the corresponding association. A weighted degree, such as the sum of absolute edge weights, can retain strength information, but it inherits scale and estimation issues and can be dominated by a few uncertain large estimates. Any such summary should specify how weights are defined and whether they are comparable across feature types and layers.

Diagnostics and Alternatives to a Single Degree Count

Before describing a node as a hub, check whether the conclusion persists under reasonable analytical choices and whether the node had comparable opportunities to acquire edges. Useful checks include:

• State the graph and candidate set. Report whether the degree is full, within-layer, cross-layer, or restricted by prior filtering. A degree computed from a restricted set is not a full-network degree.
• Examine tuning sensitivity. Compare degree and ranking over defensible regularization values, while noting that the edge sets need not be nested.
• Assess edge uncertainty. Use edge uncertainty with bootstrap and subsampling or stability selection for sparse networks, and report the individual edges contributing to the hub’s degree.
• Check model fit and measurement quality. Poor fit, outliers, missingness, near-constant features, or layer-specific reliability differences can make selected edges misleading.
• Calibrate selection errors. Consider whether edge-selection and false-discovery calibration are appropriate for the estimator and candidate universe. A degree can be high because many opportunities each have a modest false-selection rate.
• Compare like with like. For nodes with different candidate counts, report raw degree alongside a clearly defined opportunity-aware summary and layer-specific results.

These checks are complementary rather than interchangeable. A normalized degree adjusts for the number of eligible edges but not their detectability. Bootstrap selection frequency describes resampling stability but is not the probability that an edge is true. A weighted degree summarizes estimated strengths but does not automatically account for uncertainty. Formal hub testing can compare a node’s degree with an appropriate null or reference distribution, provided that the reference preserves relevant constraints such as layer structure, candidate availability, and the fitting procedure.

Common Mistake

Wrong reasoning: “Node A has the largest estimated degree, so it must be the most biologically central feature.” This overlooks the possibility that A has more eligible edges, better-measured values, or more unstable false selections, and that its ranking may depend on regularization.

Correction: Describe the degree precisely as an estimate from a specified candidate network and fitting procedure. Then assess whether the relevant edges are stable, whether the node had comparable opportunities for selection, and whether the result survives reasonable modeling choices. Even a robust conditional-association hub is not, by itself, evidence of causal control.