Gene Modules That Survive Every Definition of "Community"

When biologists analyze RNA sequencing data (measurements of how strongly each gene is expressed across many samples), they routinely group genes into "modules" that rise and fall together and then read biology off those groups. The grouping step is a community detection problem on a gene–gene correlation network, and network scientists have long known that this problem has no single correct answer. "Community" can mean a densely wired cluster, a group separated by few edges, a set of nodes with interchangeable connection patterns, or a region where a random walker stays trapped. These definitions give different partitions of the same graph. Transcriptomics mostly ignores this and reports whatever modules one pipeline produces, usually hierarchical clustering in WGCNA, which is itself fragile to its cut parameters.

We propose to treat the choice of definition as a source of uncertainty that can be measured, rather than a hidden default. The core claim is that groups of genes that stay together under all four standard community definitions, and under random reorderings of the input, form "lens-invariant cores." We expect these cores to replicate across independent donors and to predict held-out expression better than modules from any single method. We also expect them to do better than a same-size ensemble built from repeated runs of one method. Genes whose module membership changes with the definition are not noise to discard. They are candidates for bridge or transitional genes, and we test that directly.

Core Claim and What Is New

Network science supplies both the motivation and the main tools. Different approaches to community detection sorts community methods into four lenses: cut-based, clustering or density, stochastic block model (a statistical model in which nodes belong together if they connect to the rest of the graph in the same way), and dynamical or flow-based. It shows these lenses can disagree sharply. On a ring-of-rings graph, the Map Equation partition costs 2.67 bits/step against 4.13 for the modularity partition, and in a bipartite graph the density lens finds no communities while the block-model lens finds two perfect ones. Constant Communities in Complex Networks shows that heuristics like Louvain change their output when the vertex order is shuffled. It defines "constant communities," node sets that co-cluster under every ordering, and collapsing them cut modularity variance to near zero. We borrow that invariance idea, but we apply it across definitions rather than only across orderings of one definition.

On the transcriptomics side, three papers bound what is already known. COGENT splits samples in half, rebuilds the network on each half, and scores agreement of edges and node degrees. It stops short of asking whether modules replicate, and that module-level replication is our primary metric. Core Structure Detection replaces WGCNA's fragile tree cut with cores that survive increasing correlation thresholds. It reports that 85% of known cell-cycle regulators fall in its cores on the Spellman yeast data. Its notion of robustness covers thresholds within a single density-style definition, whereas ours spans definitions. MuDCoD finds gene communities in single-cell data across donors and time by smoothing spectral embeddings. It commits to one block-model or spectral notion of community and never asks how another definition would change the modules. We reuse its datasets and treat it as a strong competitor.

The cell-clustering ensemble GRACE is the closest methodological relative. It averages five cell clustering methods and keeps pairs on which a majority agree. We differ in three ways. We cluster genes, not cells. We pick component methods to span distinct definitions on purpose. And we require agreement under every lens rather than a majority, which turns disagreement into a signal instead of averaging it away.

To make this precise, let $p^{\ell}_{ij}$ be the fraction of runs of lens $\ell$ (across vertex orderings and donor resamples) in which genes $i$ and $j$ land in the same module. The lens-invariant affinity is

$$s_{ij} = \min_{\ell \in \{\text{cut},\ \text{density},\ \text{block},\ \text{flow}\}} p^{\ell}_{ij},$$

and a core is a connected set of at least three genes whose pairwise $s_{ij}$ all exceed a threshold $\tau$. Taking the minimum, rather than the mean, is the key ingredient, and the ablations below isolate it. A gene's lens sensitivity is how much its co-assignment with its core partners varies across lenses.

Research Program

Network construction follows a design meant to sidestep single-cell dropout, the frequent zero counts caused by missed molecules. For the single-cell datasets we form pseudobulk profiles per donor, time point, and cell type, then build signed soft-thresholded co-expression graphs as in WGCNA. We keep the network fixed across lenses so that only the community definition varies. Each lens gets a standard instantiation. The density lens uses Louvain modularity optimization and, separately, WGCNA's default tree cut. The cut lens uses normalized-cut spectral partitioning. The block lens uses degree-corrected stochastic block model inference. The flow lens uses Infomap. Ollivier–Ricci curvature, which flags bridge edges via optimal transport, is added as a fifth geometric lens on sparsified k-nearest-neighbor graphs only. Its $O(EV^2)$ cost reported in Ollivier-Ricci Curvature-Based Method to Community Detection in Complex Networks rules out dense graphs. Every lens is run under many vertex orderings and donor resamples to estimate $p^{\ell}_{ij}$.

The program then has three parts. First, we map how much the lenses disagree on real co-expression graphs, measured as pairwise NMI between lens partitions and as the fraction of genes in lens-invariant cores. Second, we test whether invariant cores replicate and predict better than the alternatives. Third, we ask whether lens-sensitive genes are biologically special. Our hypothesis is that they sit on curvature-negative bridges and change module membership over time in the T-cell and iPSC series. Signed networks are where the bipartite paradox should appear biologically. Two anti-correlated programs look like one block to the block model but two clusters to the density lens, so we track that case explicitly.

Baselines

The naive reference point is size-matched random gene sets and, separately, the raw WGCNA default pipeline with stock power, cut height, and minimum module size. If invariant cores do not beat WGCNA defaults on cross-donor replication, the project has no practical value. WGCNA is also the de facto standard that any biologist would use.

The strongest existing methods are Core Structure Detection and MuDCoD. CSD directly targets module robustness and has the one concrete gene-level number in our file, 85% of known cell-cycle regulators recovered in cores on the Spellman data. MuDCoD is the only cited method built for multi-donor, multi-time single-cell gene networks. On the iPSC and CD4+ T-cell data it reports higher cross-donor NMI than competitors, but no reported numerical result found on Lacuna for MuDCoD's real-data NMI. On its synthetic MuS-Dynamic-DCBM benchmark it reports up to 20% higher ARI than PisCES and static spectral clustering. All methods receive the same pseudobulk networks, the same donor splits, and an equal hyperparameter search budget, tuned only on training donors.

Beating CSD means recovering at least as many Spellman regulators in fewer total core genes, that is, with higher precision. Beating MuDCoD means higher held-out-donor module NMI on the T-cell and iPSC series at matched module counts.

Two ablations isolate the key ingredient. The first is a same-lens ensemble that runs only Louvain, or only WGCNA, with as many orderings and resamples as the four-lens method uses in total, and applies the same minimum rule over runs. It tests whether crossing definitions adds anything beyond plain stability. The second replaces the minimum over lenses with the mean, which is a GRACE-style majority consensus. It tests whether demanding unanimity across definitions matters.

Why not just…?

Why not just use WGCNA with sensible defaults? The CSD paper shows WGCNA's modules reshuffle substantially under small changes to cut height and minimum cluster size, so a single default run is a single draw from an unstable process. The benchmark here does not argue the point. WGCNA defaults are a baseline, scored on whether their modules replicate in held-out donors, and if they replicate as well as invariant cores we will say so.

Why not just bootstrap one method and keep its stable modules? This is the real threat. Lens-invariance might reduce to ordinary resampling stability, which COGENT-style splitting and constant-community preprocessing already provide. The matched-budget same-lens ensemble is designed to settle it. If that ensemble matches four-lens cores on replication and prediction, we will drop the cross-definition claim and report the narrower result that module-level resampling stability is what matters. The lens-sensitivity analysis would then be the only remaining contribution.

Why not just pick whichever method gives the most GO/KEGG-enriched modules? Gene Ontology and KEGG annotations are incomplete and skewed toward well-studied genes. Evaluation of Community Mining Algorithms in the Presence of Attributes shows that treating noisy node attributes as ground truth misleads method comparison and recommends using them only as guidance. Selecting by enrichment would also reward methods that rediscover already-annotated pathways. We therefore report enrichment as a secondary metric only and never use it for tuning.

Why not just run a GRACE-style majority ensemble over gene clusterings? Majority voting hides the disagreements we want to study. GRACE's own gains (average ARI 0.72 against 0.60 for the best single method) are on cell labels, not gene modules. The mean-versus-minimum ablation tests directly whether unanimity buys better replication.

Proposed Experiments

On synthetic data, we plant modules of each structural type into one graph: dense clusters, block-equivalent groups, and flow-trapping groups. We do this with LFR graphs and MuDCoD's MuS-Dynamic-DCBM generator. We then check two things. Single lenses should recover only their own planted type. Invariant cores should be high-precision and recover all planted types, though not completely, and lens-sensitive nodes should concentrate at planted boundaries.

On real data, the primary test splits donors in the CD4+ T-cell activation series (0, 16, 40 hours, 5 days) and the HipSci iPSC dopaminergic series (days 30 and 52) into disjoint halves. We build modules on one half and score three things on the other: NMI between independently rebuilt modules, the fraction of held-out expression variance explained by module eigengenes (the first principal component of each module's expression), and stability under vertex reordering. The secondary check is regulator recovery precision on the Spellman yeast cell-cycle data against CSD.

The decisive outcome is threefold. Four-lens invariant cores must beat both the matched same-lens ensemble and the mean-consensus ablation on held-out-donor replication. They must also beat CSD on regulator precision. And lens-sensitive genes must be enriched among curvature-negative bridge genes and among genes whose module membership changes across activation time points. If the first comparison fails, the cross-definition hypothesis is falsified, and we will report the result as such.