Degree, Strength, And Clustering In Gene Networks

What Network Summaries Describe

A gene co-expression network reduces relationships among expression profiles to nodes and edges. Once the network has been constructed, node-level summaries can describe its local structure: how many other genes are linked to a gene, how much total edge weight those links carry, and how interconnected its neighbors are. These are descriptive properties of a specified network, not direct measurements of a gene’s biological importance.

The earlier material on nodes, edges, and co-expression network estimands distinguishes the network being estimated from the expression data used to estimate it. The earlier material on hard and soft thresholding and on signed, unsigned, and signed-hybrid networks also matters here: the summaries below depend on the adjacency matrix produced by those choices. For clarity, first consider an undirected network with no self-loops and nonnegative edge weights. Write \(a_{ij}=a_{ji}\) for the weight of the edge between nodes \(i\) and \(j\), with \(a_{ii}=0\). In an unweighted network, \(a_{ij}\) is 1 for an edge and 0 otherwise. In a weighted network here, \(a_{ij}\) lies between 0 and 1, and 0 means no edge.

Definition: The degree \(k_i\) of node \(i\) is the number of its neighbors: \(k_i=\sum_{j\ne i} I(a_{ij}>0)\). Its weighted strength \(s_i\) is the sum of its incident edge weights: \(s_i=\sum_{j\ne i}a_{ij}\). Degree counts links; strength totals their weights.

In an unweighted network, degree is the number of edges incident to a node, and strength equals degree if each existing edge has weight 1. In a weighted network, strength reflects both how many edges there are and how large their weights are. It is not generally a count, and its numerical size depends on the scale used for the weights.

Degree and Strength in a Small Network

Worked Example: Counting Neighbors and Summing Weights

Consider a hypothetical weighted network with nodes \(U,V,A,B,C,D\). The only edges incident to \(U\) or \(V\) are \(U\!-\!A\) with weight 0.9, \(U\!-\!B\) with weight 0.1, \(V\!-\!C\) with weight 0.4, \(V\!-\!D\) with weight 0.4, and \(V\!-\!A\) with weight 0.4. All listed weights are positive, so each listed edge counts as one neighbor.

For \(U\), there are two neighbors, \(A\) and \(B\). Thus \(k_U=2\), and its strength is \(s_U=0.9+0.1=1.0\). For \(V\), there are three neighbors, \(C,D,A\), so \(k_V=3\), and \(s_V=0.4+0.4+0.4=1.2\).

Here \(V\) has both greater degree and greater strength. That conclusion is only about this network representation: it does not establish that \(V\) is more biologically important. In another network, a node with fewer but stronger links could have greater strength than a node with more weak links.

Local Clustering: Are a Node’s Neighbors Connected?

Degree and strength describe a node’s direct connections. Local clustering asks a different question: among the node’s neighbors, how often are pairs connected to each other? A triangle involving \(i,j,h\) consists of the three edges \(i\!-\!j\), \(i\!-\!h\), and \(j\!-\!h\). The local clustering coefficient compares the number of triangles involving \(i\) with the number of neighbor pairs that could form such triangles.

Definition: In an unweighted network, the local clustering coefficient of node \(i\), when \(k_i\ge2\), is \(C_i=t_i/\binom{k_i}{2}\), where \(t_i\) is the number of edges among the neighbors of \(i\). Equivalently, \(C_i=2t_i/[k_i(k_i-1)]\). If \(k_i<2\), use the convention \(C_i=0\) and state that convention when reporting results.

The denominator counts all possible pairs among the \(k_i\) neighbors. If all those pairs are linked, every possible pair closes a triangle with \(i\), and \(C_i=1\). If none are linked, \(C_i=0\). Intermediate values give the fraction of possible neighbor-to-neighbor edges that are present. The coefficient is local to node \(i\); it is not the fraction of all edges in the entire network that belong to triangles.

Worked Example: Binary Local Clustering

Consider a hypothetical unweighted network on nodes \(A,B,C,D\) with edges \(A\!-\!B\), \(A\!-\!C\), \(B\!-\!C\), \(B\!-\!D\), and \(C\!-\!D\), and no other edges. For node \(B\), the neighbors are \(A,C,D\), so \(k_B=3\). There are \(\binom{3}{2}=3\) possible pairs among them: \(A\!-\!C\), \(A\!-\!D\), and \(C\!-\!D\). Two are present, \(A\!-\!C\) and \(C\!-\!D\). Therefore,

C_B=\frac{t_B}{\binom{k_B}{2}} =\frac{2}{\binom{3}{2}} =\frac{2}{3}\approx0.667.

Thus two-thirds of the possible neighbor-to-neighbor links for \(B\) are present. For \(A\), its neighbors \(B,C\) are connected, so \(C_A=1/1=1\). This means all possible pairs among \(A\)’s neighbors are connected; it does not mean that \(A\) has many neighbors.

A Weighted Local Clustering Coefficient

A binary clustering coefficient treats every retained edge alike, even if the underlying weights differ substantially. One common weighted alternative uses the geometric mean of the three weights in each triangle. For this definition, weights must be nonnegative and on a specified common scale, taken here to be from 0 to 1. It is often called the Onnela weighted clustering coefficient.

Definition: For \(k_i\ge2\), define the weighted local clustering coefficient by \(C_i^{(w)}=\frac{1}{k_i(k_i-1)}\sum_{j\ne i}\sum_{h\ne i,j}(a_{ij}a_{ih}a_{jh})^{1/3}\), where the sums run over ordered pairs of distinct neighbors \(j,h\) of \(i\). If \(k_i<2\), set \(C_i^{(w)}=0\) by convention.

The geometric mean is small if any of a triangle’s three edges is weak, and it is zero if one of those edges is absent. This coefficient combines the triangle’s connectivity and its edge intensities. Its denominator remains the number of ordered pairs of neighbors, \(k_i(k_i-1)\), so it is an average contribution across potential neighbor pairs. If all edges have weight 1, it equals the binary clustering coefficient.

Worked Example: Weighted Clustering Around One Node

Suppose node \(i\) has three neighbors \(j,k,l\) with weights \(a_{ij}=0.8\), \(a_{ik}=0.5\), and \(a_{il}=0.4\). The weights among the neighbors are \(a_{jk}=0.6\), \(a_{jl}=0.1\), and \(a_{kl}=0.3\). All six weights are positive, so \(k_i=3\). The three unordered neighbor pairs contribute the geometric means \((0.8\cdot0.5\cdot0.6)^{1/3}\), \((0.8\cdot0.4\cdot0.1)^{1/3}\), and \((0.5\cdot0.4\cdot0.3)^{1/3}\). Each unordered pair appears twice in the ordered sum, which cancels with the factor of 2 in the denominator:

C_i^{(w)} =\frac{0.24^{1/3}+0.032^{1/3}+0.06^{1/3}}{3} \approx\frac{0.6214+0.3175+0.3915}{3} \approx0.4435.

The binary coefficient for this node is 1 because all three possible links among its neighbors are present. The weighted coefficient is lower because the triangle weights are not all strong. Its value summarizes this chosen weighting rule; it is not a probability or a measure of the statistical certainty of the network.

Why the Weighted Formula Has This Form

The formula follows directly from averaging triangle intensity over possible neighbor pairs. For an unordered neighbor pair \(\{j,h\}\), its triangle with \(i\) contributes \(q_{ijh}=(a_{ij}a_{ih}a_{jh})^{1/3}\). The ordered sum counts that pair twice, once as \((j,h)\) and once as \((h,j)\). There are \(\binom{k_i}{2}=k_i(k_i-1)/2\) unordered pairs. Therefore the ordered-sum average is the same as the average of \(q_{ijh}\) over unordered pairs:

\frac{\sum_{j\ne i}\sum_{h\ne i,j}q_{ijh}}{k_i(k_i-1)} =\frac{2\sum_{\{j,h\}}q_{ijh}}{2\binom{k_i}{2}} =\frac{\sum_{\{j,h\}}q_{ijh}}{\binom{k_i}{2}}.

If every edge has weight 1, each present triangle contributes 1 and every absent neighbor-to-neighbor edge gives a contribution of 0. The numerator then counts triangles, so this expression reduces to \(t_i/\binom{k_i}{2}\), the binary coefficient. This also shows why the weighted and binary coefficients answer related but different questions: one counts triangle closure, while the other averages triangle intensity under a particular rule.

Worked Example: Comparing Degree, Strength, and Clustering

Return to the hypothetical \(U,V,A,B,C,D\) network. Suppose the additional edges among \(U\)’s neighbors and \(V\)’s neighbors are \(A\!-\!B\) with weight 0.5, \(C\!-\!D\) with weight 0.2, and \(C\!-\!A\) with weight 0.8; \(D\!-\!A\) is absent. For \(U\), its only neighbor pair is \(A,B\), which is connected. Thus \(C_U=1\) and \(C_U^{(w)}=(0.9\cdot0.1\cdot0.5)^{1/3}\approx0.3557\).

For \(V\), the neighbor pairs are \(C,D\), \(C,A\), and \(D,A\); two of the three are connected. Hence \(C_V=2/3\approx0.667\). The weighted coefficient has two nonzero contributions, \( (0.4\cdot0.4\cdot0.2)^{1/3}\approx0.3175\) and \((0.4\cdot0.4\cdot0.8)^{1/3}\approx0.5040\), and one zero contribution. Therefore \(C_V^{(w)}=(0.3175+0.5040+0)/3\approx0.2738\).

This comparison shows why the measures should not be substituted for one another. \(V\) has degree 3 and strength 1.2, compared with \(U\)’s degree 2 and strength 1.0. Yet \(U\) has greater binary and weighted local clustering. The summaries describe different features: number of links, total weight, and connectivity or intensity among neighbors.

Assumptions, Diagnostics, and Limits

These calculations assume a single specified network representation: nodes, edge weights, and any thresholding or transformation used to obtain them are fixed. The formulas above assume an undirected network, no self-loops, and nonnegative weights. For a signed network, do not insert negative weights into this particular weighted coefficient and interpret the result as if it were a nonnegative triangle intensity. A signed-network analysis needs an explicitly chosen signed clustering definition or a separately defined positive- or negative-edge network.

Several checks help catch errors before interpreting a result:

Check the neighbor set. Degree counts edges present under the chosen adjacency rule. In a weighted network, weak but positive edges still count unless a threshold or other rule removes them.
Check the clustering denominator. A node with \(k_i=3\) has three possible neighbor pairs, not six. The factor of two appears only when using the ordered-pair formula.
Check low degree. With fewer than two neighbors, there is no neighbor pair to evaluate. Report the convention \(C_i=0\) (and \(C_i^{(w)}=0\)) rather than implying a measured absence of triangles.
Check weight scale and convention. Strength changes if all weights are rescaled. Weighted clustering also depends on the weight scale and on the chosen formula. Report how weights were constructed and which coefficient was used.
Check sensitivity to edge retention. A threshold can change degree, strength, and the set of possible neighbor pairs. If conclusions change sharply under plausible network-construction choices, describe that sensitivity rather than presenting one summary as definitive.

Binary clustering is straightforward to interpret but discards differences in edge intensity once edges are retained. Weighted clustering keeps intensity but requires a stated definition; other weighted coefficients exist and need not give the same values. Strength preserves edge weights in a simple sum, but it does not reveal whether a node’s neighbors are connected to one another. None of these summaries by itself accounts for uncertainty in estimated co-expression edges, confounding, or biological function.

A high degree or strength is not evidence that a gene is a causal regulator, a validated target, or biologically important. A high clustering coefficient is not evidence that a group of genes forms a functional pathway. These summaries describe the topology of an estimated network; biological claims require additional evidence and appropriate analyses. In particular, a structural ranking can reflect network construction choices as well as the expression patterns.

Key takeaway: Degree counts neighbors, strength sums their edge weights, and local clustering measures how connected those neighbors are to one another. Interpret each in the context of the exact network and formula used; none alone establishes biological importance.

Check Your Understanding

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

A node has four neighbors, and five of the six possible pairs among those neighbors are connected. What is its binary local clustering coefficient?
A node has three incident edge weights, 0.2, 0.5, and 0.7. Find its degree and strength.
Why can a node have a higher degree than another node but a lower strength?
What convention is used for local clustering when a node has fewer than two neighbors, and why should it be reported?
What does a high local clustering coefficient establish, and what does it not establish biologically?

Solutions. (1) There are \(\binom{4}{2}=6\) possible neighbor pairs, so \(C_i=5/6\approx0.8333\). (2) All three weights are positive, so \(k_i=3\); the strength is \(s_i=0.2+0.5+0.7=1.4\). (3) Degree counts links without regard to their weights, whereas strength sums weights. For example, three edges of weight 0.1 give degree 3 and strength 0.3, while two edges of weight 0.8 give degree 2 and strength 1.6. (4) The convention is to set the coefficient to 0 for \(k_i<2\), because there are no neighbor pairs over which to calculate closure; reporting it distinguishes the convention from an observed set of unconnected pairs. (5) It establishes that a large fraction of the node’s possible neighbor-to-neighbor links are present under the specified network definition. It does not establish that the node or its neighbors are functionally important, causal, or part of a validated pathway.