Interpreting Network Rewiring Across Conditions

What Counts as Network Rewiring?

Two condition-specific gene networks can differ in several ways. They may contain different edges, the same edges with different weights, genes with changed connectivity, or modules whose relationships have changed. These are related but distinct descriptions. In particular, a network with more edges will often share more edges with another network simply because it has more opportunities for overlap. A useful comparison therefore describes both the structural difference and the networks’ overall density.

The aim here is interpretation, not a new test for differential co-expression. A difference between estimated networks is descriptive evidence of a structural change; by itself, it does not establish that the change is statistically reliable, biologically important, or caused by the condition. The earlier material on testing differential co-expression addresses inferential evidence for changes in individual gene relationships. Network summaries help organize and interpret such differences at edge, gene, and module levels.

Definition: Let $V$ be the same set of $p$ genes in two conditions, and let $E_1$ and $E_2$ be their edge sets. For an undirected network without self-loops, the possible number of edges is $M=\binom{p}{2}$. The edge overlap is $|E_1\cap E_2|$, the union size is $|E_1\cup E_2|$, and the Jaccard similarity is $$ J(E_1,E_2)=\frac{|E_1\cap E_2|}{|E_1\cup E_2|} $$ when the union is nonempty. In that case $J$ ranges from 0 (no shared edges) to 1 (identical edge sets). If both edge sets are empty, the formula is undefined because its denominator is zero; if a single summary value is required, one common convention is to define $J=1$ for this case and to report that convention.

These comparisons require a common node set and a consistent definition of an edge: the same expression preprocessing, association measure, treatment of sign, and edge-selection rule in both conditions. If one network is signed and the other unsigned, for example, their edge sets are not directly comparable without first defining how signs are handled. If the node sets differ, an apparent edge difference may instead reflect which genes were included. The networks should be built on the same eligible genes before interpreting structural differences.

Edge Overlap and the Density Benchmark

For a binary network with $m_c=|E_c|$ edges in condition $c$, its density is the fraction of possible edges that are present:

d_c=\frac{m_c}{M}.

Jaccard similarity describes how much the edge sets agree relative to their union, but does not by itself say whether the overlap is surprising given the network sizes. A simple reference calculation holds the two edge counts fixed and treats the $m_2$ edges of the second network as a uniformly selected subset of the $M$ possible edges. This is a descriptive benchmark, not a model of gene biology or a formal test.

Justification: Under that reference calculation, define an indicator $I_e$ equal to 1 if possible edge $e$ appears in both networks and 0 otherwise. Each of the $m_1$ edges in the first network has probability $m_2/M$ of appearing in the second, so $E(I_e)=m_2/M$. Since the overlap is the sum of these indicators over the first network’s edges, $$ E\bigl(|E_1\cap E_2|\bigr) =\sum_{e\in E_1}E(I_e) =m_1\frac{m_2}{M} =\frac{m_1m_2}{M}. $$

An observed overlap above this reference indicates more shared edges than expected under uniform random placement with the same edge counts. It does not, on its own, quantify uncertainty or demonstrate that the networks are statistically different. A formal inferential analysis needs an appropriate sampling or randomization model; the benchmark is useful for interpretation, not a substitute for that analysis.

Worked Example: Interpreting Edge Overlap Alongside Density

Suppose two hypothetical networks use the same six genes, so $M=\binom{6}{2}=15$ possible edges. The first network has 5 edges and the second has 7; they share 3 edges.

The densities are $d_1=5/15=0.333$ and $d_2=7/15=0.467$. The union contains $5+7-3=9$ edges, so the Jaccard similarity is $J=3/9=0.333$. Thus one third of the edges in the union are shared.

With edge counts held fixed, the reference expected overlap is $m_1m_2/M=(5)(7)/15=2.333$ edges. The observed overlap of 3 is about $0.667$ edge above that reference. This is a modest descriptive excess, not proof of significant preservation. The density difference also matters: the second network has more edges, so the raw shared-edge count should not be interpreted without the edge counts and the common set of 15 possibilities.

Edge overlap is not the only summary. The symmetric difference, $|E_1\setminus E_2|+|E_2\setminus E_1|$, counts edges present in exactly one network and makes turnover explicit. A normalized version can be useful when comparing networks of different sizes, but the denominator and interpretation should be stated. Jaccard is one such normalization; it is undefined for two empty networks unless a convention is adopted.

Connectivity Changes Are Not the Same as Edge Changes

For an undirected binary network, the degree $k_i^{(c)}$ of gene $i$ is the number of edges incident to that gene in condition $c$. Degree summarizes how many partners a gene has, not which partners they are. In a weighted network, the strength $s_i^{(c)}$ is the sum of the weights on edges incident to gene $i$. Strength therefore depends on both edge presence and edge weights. If weights can be negative, the analysis must specify whether strength sums signed weights or uses a nonnegative adjacency; these quantities answer different questions.

For an undirected binary network with $p$ genes and $m_c$ edges, the mean degree is $2m_c/p$, since each edge contributes one degree to each endpoint. A gene’s degree divided by mean degree is a descriptive relative-connectivity ratio only when the mean degree is positive. If a network has no edges, every degree and its mean are zero, so the ratio is undefined; report that it is undefined rather than assigning a relative-connectivity value.

Worked Example: Equal Density Can Hide Different Partners

Consider two hypothetical binary networks on genes 1 through 5. In condition 1 the edges are $\{12,13,14,23,45\}$; in condition 2 they are $\{12,13,15,25,34\}$. Both have $M=10$ possible edges and 5 observed edges, so each has density $5/10=0.5$.

In condition 1, the degrees for genes 1 through 5 are $(3,2,2,2,1)$. In condition 2, they are $(3,2,2,1,2)$. The mean degree is $2m/p=2(5)/5=2$ in each network. Gene 4’s degree falls from 2 to 1, while gene 5’s rises from 1 to 2. Their relative-connectivity ratios are, respectively, $2/2=1$ and $1/2=0.5$ for gene 4, and $1/2=0.5$ and $2/2=1$ for gene 5.

Genes 1, 2, and 3 retain the same degrees, but this does not mean the networks are identical: genes 4 and 5 have different partners across conditions. The example shows why a degree change is a useful summary but cannot replace edge-level comparison. Equal density also does not mean equal topology.

If the graph is weighted, compare strengths as well as, or instead of, degrees when edge weights carry meaningful information. The scales must be comparable across conditions: weights generated by different transformations or differently scaled estimates cannot be read as direct strength changes. Degree and strength are summaries of connectivity, not direct measures of gene activity, regulation, or causal influence.

Module Relationships Across Conditions

Modules provide a higher-level view of network organization. If module assignments are held fixed across conditions, one can compare the number or fraction of edges within each module and between each pair of modules. A shift from mostly within-module edges to more between-module edges can indicate a change in how groups of genes are connected. The comparison is meaningful only if the modules refer to the same gene groups in both conditions. When modules are detected separately, their labels are arbitrary; align them by gene overlap or another stated matching rule before comparing them.

Worked Example: A Change in Between-Module Connectivity

Suppose six hypothetical genes are assigned to two fixed modules, $A=\{1,2,3\}$ and $B=\{4,5,6\}$, in both conditions. In condition 1, the edges are $\{12,13,23,45,34\}$. Four are within modules and one, $34$, connects the modules. In condition 2, the edges are $\{12,45,14,25,36\}$. Two are within modules and three connect $A$ to $B$.

Both networks have 5 edges among $M=15$ possibilities, so their total densities are $5/15=0.333$. The fraction of observed edges that are between modules is $1/5=0.20$ in condition 1 and $3/5=0.60$ in condition 2. Because the total edge counts are equal, this particular summary is not explained by a difference in overall density. It describes a redistribution of edges toward between-module connections, conditional on the stated fixed module assignment. It does not establish that the modules themselves are biologically distinct or that the redistribution is statistically reliable.

Module eigengenes offer another summary: they are lower-dimensional representations of module expression, and their correlation describes how the module summaries vary together across samples within a condition. For example, a hypothetical eigengene correlation changing from $0.62$ in one condition to $-0.41$ in another describes a change from positive to negative association between those summaries. It is not the same quantity as the number of edges between modules, and it does not imply that one module causes changes in the other. Correlation estimates also depend on sample size and variability, so apparent differences should be considered alongside their uncertainty.

Diagnostics, Alternatives, and Limits

Before interpreting a rewiring summary, check that both networks use the same genes, edge definition, sign convention, and construction rule. Report edge counts and densities with overlap statistics. If one condition is much denser, the observed overlap may be high simply because it has many edges; the fixed-edge-count benchmark helps reveal this, while density-matched comparisons answer a different question and may discard a genuine global density change. Neither should be substituted for the other without stating the scientific target.

Binary edge comparisons discard weight differences: two networks can have identical edge sets but substantially different edge strengths. Weighted comparisons retain more information, but require comparable weight scales and a clear treatment of sign. Conversely, degree and strength compress many edges into one value per gene and can hide changes in partner identity. Module-level summaries are easier to interpret at a group level but depend on module definitions and may obscure gene-specific changes. These summaries complement rather than replace one another.

Warning signs include a very small or zero edge count, a large density imbalance, modules that cannot be matched clearly, or a striking summary based on only a few edges. A zero-edge network has density zero and mean degree zero; Jaccard is undefined if both networks are empty under its ratio definition, and relative degree is undefined in either zero-edge network. Such cases should be reported explicitly. Also distinguish a thresholded edge’s absence from evidence that the underlying association is exactly zero: estimates near a cutoff can switch edge status with small changes in data or construction choices.

Finally, different sample sizes and sample composition can affect estimated associations and their precision. A network difference may reflect technical batch effects, cell-mixture differences, or other confounding rather than condition-specific biology. Earlier material on batch, cell mixture, and spurious co-expression is relevant when assessing these possibilities. Descriptive network comparison cannot resolve them by itself; use the inferential results and study design to qualify claims. Prefer wording such as “the estimated network shows reduced connectivity” to an unsupported assertion that a gene has been biologically rewired.

Key takeaway: Interpret rewiring at several levels: which edges are shared or changed, how gene connectivity shifts, and whether module relationships change. Always place these summaries beside edge counts and densities, and treat them as descriptions of estimated networks unless an appropriate inferential analysis supports a stronger conclusion.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions below.

Two networks each have 4 edges and share all 4. What are their Jaccard similarity and densities if there are 10 possible edges?
Both edge sets are empty. What is the Jaccard ratio, and how should a reported value handle this case?
A network has 6 genes and no edges. What are its mean degree and a gene’s degree-to-mean-degree ratio?
Two networks have equal edge counts and equal degrees for every gene, but different edge sets. Can they still be rewired relative to one another? Explain.

Solutions. (1) The union contains 4 edges, so $J=4/4=1$; each density is $4/10=0.4$. (2) The ratio is $0/0$, hence undefined. If one value is required, state the convention, such as assigning $J=1$ when both sets are empty. (3) Mean degree is $2m/p=0$; the ratio is $0/0$ and is undefined. (4) Yes. Degrees count partners but do not identify them, so edge identities can differ even when every gene’s degree is unchanged.