Why Network Construction Choices Need Testing
A co-expression network is not a direct observation of biological edges. It is constructed from expression measurements using choices about correlation, adjacency, and module detection. Two reasonable choices can yield different edge weights, gene groupings, or downstream interpretations. Sensitivity analysis asks whether the scientific conclusion persists when those choices are varied over defensible alternatives.
This question differs from asking whether a result is stable under resampling. Resampling changes which observations contribute to the analysis; sensitivity analysis changes the analysis specification. Both can be informative, but they answer different questions. Here the comparison holds the expression data, preprocessing, and sample set fixed while varying network-construction choices.
There is no universal requirement that every edge or module remain identical. The relevant question is whether the biological statement being made—such as a group of genes forming a module or a module showing a particular phenotype association—survives the perturbations. A robust conclusion is not necessarily true, and a sensitive conclusion is not automatically false; sensitivity identifies where the evidence depends on analyst choices.
Specify the Choices Before Comparing Results
Let \(X_{ig}\) denote the expression value for sample \(i\) and gene \(g\), with the same set of samples and genes used in each comparison. For each pair of genes, a correlation measure produces \(r_{gh}\). Pearson correlation measures linear association and can be strongly affected by outlying observations. Spearman correlation is the Pearson correlation of ranks and captures monotone association; it is less sensitive to the scale of extreme values but can respond differently when ties or nonlinear patterns occur. Robust correlation estimators, such as the biweight midcorrelation, reduce the influence of some outliers under their chosen estimation rules. None removes the need to inspect data quality or account for confounding.
The correlation calculation is meaningful only when the paired expression values refer to the same samples and have been processed comparably. Correlations can also reflect batch, cell-mixture, or other shared sample-level effects rather than gene-specific biological coordination. A sensitivity analysis should therefore keep preprocessing and any justified adjustment fixed; it should not use changes in these steps to conceal a result that is otherwise fragile.
An adjacency rule transforms a correlation into a nonnegative edge weight. Two common soft-threshold rules, for a positive power \(\beta\), are:
An unsigned adjacency gives strong positive and negative correlations high weights. A signed adjacency gives high weight to strong positive correlations and low weight to strong negative correlations. Increasing \(\beta\) accentuates the strongest correlations relative to weaker ones. A hard-threshold alternative sets an edge to 1 when a specified correlation criterion is met and to 0 otherwise; this discards gradations in strength and can make edges change abruptly near the cutoff.
Module construction adds further choices. In hierarchical clustering followed by dynamic tree cutting, parameters such as minimum module size and cut sensitivity can change how finely branches are divided. A subsequent merging rule can combine modules whose eigengenes are sufficiently similar. These choices should be varied over scientifically defensible settings, not arbitrary extremes selected to produce a desired result.
Worked Example: The Same Correlation Under Different Adjacency Rules
Suppose two hypothetical genes have Pearson correlation \(r_{gh}=0.70\) across the same expression samples, and the soft-threshold power is \(\beta=2\). The unsigned weight is
The signed weight is
The same positive correlation therefore gives different weights under the two rules. For a second hypothetical pair with \(r_{gh}=-0.70\), the unsigned weight is still \(|-0.70|^2=0.4900\), whereas the signed weight is \(((1-0.70)/2)^2=(0.15)^2=0.0225\). Thus, changing signedness can alter not just the numerical scale but which relationships contribute strongly to the network.
For comparison, a hard unsigned threshold at \(|r_{gh}|\geq 0.60\) would assign weight 1 to both pairs, whereas a hard positive-correlation threshold at \(r_{gh}\geq 0.60\) would retain only the positively correlated pair. Those rules answer different construction questions; neither is interchangeable with the soft-threshold calculations.
Measure Whether Modules and Claims Persist
Before comparing module labels, match modules across specifications. Labels such as “module 1” and “module 2” are arbitrary and may be permuted. For two matched gene sets \(A\) and \(B\), a simple agreement measure is the Jaccard index: the number of shared genes divided by the number of genes in either set. It ranges from 0 (no shared genes) to 1 (identical sets). A low value does not by itself show which module is preferable; it signals that membership depends on the construction choices.
A module-level overlap measure should be paired with measures relevant to the actual conclusion. For example, if the claim concerns phenotype association, compare the association estimate and its direction across matched modules, not only the gene lists. If the claim concerns a set of proposed hub genes, report whether those genes retain their roles under alternative constructions. A high average overlap can hide one important module that changes substantially, so summarize both overall agreement and the particular result being interpreted.
Worked Example: Quantifying Module Overlap
Two hypothetical specifications produce modules \(A=\{a,b,c,d,e,f\}\) and \(B=\{a,b,c,d,g,h\}\). The intersection is \(\{a,b,c,d\}\), containing 4 genes. The union contains \(a,b,c,d,e,f,g,h\), or 8 genes. Therefore,
Half of the genes appearing in either module are shared. The two modules are not identical, and the four genes found in only one of them should not be silently treated as stable membership. If the biological conclusion depends on a particular gene among \(e,f,g,h\), its changing assignment deserves explicit attention. This calculation measures agreement; it does not establish that either module is biologically valid.
Compare Choices in a Controlled Way
A useful first pass changes one class of choices at a time: correlation measure, adjacency rule or power, and module parameters. Keep the other settings fixed for each comparison. This identifies which changes coincide with the largest shifts and is easier to interpret than changing everything simultaneously. A small grid of combinations can then check whether choices interact—for instance, whether a module is sensitive to signedness only at one power. Record the full specification for every fit.
State what must persist: a matched module, a core gene set, a phenotype-association direction, or another specific biological interpretation.
Set out plausible correlation measures, adjacency settings, and module parameters before examining which setting gives the most attractive result.
Change one class at a time for attribution, then examine a manageable set of combinations if interactions could affect the conclusion.
Report overlap and the target biological results for each specification, highlighting the choices associated with the largest changes.
A short justification for one-at-a-time comparisons is that they isolate a contrast. If the baseline specification is \(S_0\) and \(S_c\) changes only the correlation measure, then a difference between their outputs is attributable to that change within the fixed analysis pipeline. If \(S_c\) also changes the adjacency rule and cut parameters, the observed difference cannot be assigned to any single choice. One-at-a-time contrasts do not reveal interactions; combined settings are needed to assess those.
Worked Example: Identifying the Most Influential Choice
Consider hypothetical outputs for a baseline module and alternatives, with the module matched to the baseline in each fit. The table reports Jaccard overlap with the baseline and the correlation between the matched module eigengene and a fixed phenotype. The phenotype association is an illustrative summary, not evidence of causation.
| Specification | Jaccard overlap | Eigengene–phenotype correlation |
|---|---|---|
| Baseline: Pearson, unsigned, \(\beta=6\), minimum size 10 | 1.00 | 0.60 |
| Spearman correlation only | 0.86 | 0.58 |
| Biweight correlation only | 0.82 | 0.62 |
| Signed adjacency only | 0.55 | 0.31 |
| Unsigned adjacency, \(\beta=4\) only | 0.72 | 0.49 |
| Unsigned adjacency, \(\beta=8\) only | 0.84 | 0.57 |
| Minimum module size 15 only | 0.78 | 0.56 |
Among the listed alternatives, changing correlation measure gives overlaps from 0.82 to 0.86, a range of \(0.86-0.82=0.04\), and phenotype correlations from 0.58 to 0.62, a range of 0.04. Changing the adjacency setting gives overlaps from 0.55 to 0.84, a range of \(0.84-0.55=0.29\), and phenotype correlations from 0.31 to 0.57, a range of 0.26. The minimum-size change gives overlap 0.78 and correlation 0.56, relatively close to the baseline values.
In this hypothetical comparison, adjacency choices have the largest observed effect, especially signedness. The positive association direction persists in the displayed alternatives, but its magnitude and the matched membership change notably under signed adjacency. A careful report would say that this result is sensitive to adjacency specification, rather than claiming an invariant module–phenotype relationship. Because these are one-at-a-time contrasts, the table does not tell us whether signedness interacts with power or module parameters.
Diagnostics, Alternatives, and Limitations
Sensitivity analysis is most useful when its summaries make dependence visible. Check that every comparison uses the same genes, samples, preprocessing, and downstream definition of the target claim. Confirm that modules have been matched rather than compared by arbitrary labels. Look for a conclusion that reverses direction, loses key genes, splits into several modules, or becomes unassignable under reasonable settings. Such changes are evidence of dependence on construction choices, not a reason to select the specification that supports a preferred story.
One-at-a-time comparisons are transparent but can miss interactions. A factorial or grid comparison varies multiple settings in a planned combination and can reveal that one choice matters only in the presence of another. It also creates more results to summarize and should use a manageable, justified set of settings. A single “best” network selected by a topology diagnostic is not a substitute for sensitivity analysis: it reports a selection decision, not whether the biological conclusion persists.
The comparison is limited to the alternatives actually examined. If every specification shares the same unaddressed confounding or measurement problem, apparent agreement does not resolve it. Jaccard overlap can also be misleading when modules differ greatly in size or when a broad shared core is accompanied by unstable peripheral genes. Report module sizes and inspect the genes that determine the substantive interpretation. Sensitivity summaries complement, rather than replace, checks on data quality, biological plausibility, and uncertainty.
Common Pitfall
Correction: A module may persist while its membership, eigengene association, or key genes change. Match modules explicitly, quantify overlap, and compare the specific biological claim. Also state which alternatives were tested; agreement across a narrow set of choices is not evidence of robustness to every possible construction.
The central reporting principle is to show the dependence, not conceal it. Name the baseline and alternatives, give a small number of interpretable comparison measures, identify which choice had the largest impact, and state whether the conclusion is stable, qualified, or specification-dependent.
Check Your Understanding
Answer each question, then compare your reasoning with the worked solutions below.
- For \(r_{gh}=-0.60\) and \(\beta=2\), calculate the unsigned and signed soft-threshold adjacencies.
- Two matched modules share 6 genes; their union contains 10 genes. Calculate their Jaccard index and interpret it.
- Why is it useful to vary one construction choice at a time before examining combined settings?
- A module’s eigengene–phenotype correlation is positive under most settings but negative under one defensible setting. What should be reported?
Solutions. (1) The unsigned adjacency is \(|-0.60|^2=0.3600\). The signed adjacency is \(((1-0.60)/2)^2=(0.20)^2=0.0400\); the negative correlation therefore receives much less weight under the signed rule. (2) \(J=6/10=0.600\). The sets share 60% of the genes in their union, so they overlap but are not identical. (3) Holding other settings fixed makes the output contrast attributable to the changed choice within the pipeline. Combined settings can reveal interactions, but by themselves make it harder to identify which choice produced a difference. (4) Report the direction change and the specification under which it occurs, along with the corresponding matched-module results. Do not describe the association direction as robust across the tested settings; investigate whether the conclusion depends on that construction choice.