Biologists routinely group genes that rise and fall together into "modules." They then name the most connected gene in each module a "hub" and propose it as a master regulator worth knocking out. The standard tool is weighted gene co-expression network analysis (WGCNA). It turns pairwise correlations into a soft-weighted network, finds modules by clustering, and summarizes each module by its first principal component, the "module eigengene." Almost nobody checks whether hubs actually drive their modules. The data to check now exist: large CRISPR interference screens with single-cell readout (Perturb-seq) measure what happens to every other gene when one gene is knocked down, one gene at a time, for hundreds of genes.
This note proposes a test of hub causality and one specific way hub-picking should be fixed. The claim is that a gene's hub status becomes predictive of its causal influence only under two conditions. The network has to be built within cell states rather than on averaged bulk samples. And the gene has to stay a hub across several different stress conditions, not just the one condition a study happened to profile. The logic borrows from invariant causal prediction: causal relationships should persist across environments, while correlations that come from cell composition or from one stressor's generic "crisis" response should not.
Core Claim
Hubs from a WGCNA network built on one condition with bulk-like averaging mostly mark cell-composition and proliferation programs, not regulators. The research file supplies three motivating results. Single-Cell Co-expression Analysis Reveals Distinct Functional Modules… compared single-cell and bulk glioblastoma networks and found that under 10% of top co-expressed pairs were shared between them. Its bulk links reflected broad metabolism and chromosomal proximity, while its single-cell links were five times more enriched for physical protein interactions. Correlation Analysis Connects Cancer Subtypes showed that shared expression profiles can be passive byproducts of proliferation rather than shared drivers. Stress-biology work shows that single-condition profiling hides how gene-family members divide the labor. In cotton annexins, GhANN4 and GhANN11 split salt/drought versus cold/heat roles. In Arabidopsis, paralog pairs partition into distinct response classes across four stresses.
We define a gene's invariant hubness as the lowest intramodular connectivity it reaches across a set of contexts:
$$H_g = \min_{e \in E} k^{(e)}_g$$
Here $k^{(e)}_g$ is the correlation between gene $g$ and its module eigengene in context $e$, computed on library-size-stabilized metacells. $E$ is the set of contexts, which can be stress conditions, cell lines, or cell states. We use the minimum rather than the mean on purpose. A gene that is central under one stress and peripheral under another should not rank highly. The prediction is that $H_g$ ranks genes by the measured size of their knockdown effect better than single-context hubness, bulk hubness, or simple expression-based scores.
Research Program
The method is assembled from existing parts. Metacells come from LSMetacell, which groups similar cells while holding cumulative library size near a target and so avoids the spurious correlations that sparse single-cell data produce. We then run standard WGCNA within each context: soft-threshold power, topological overlap, modules, and eigengenes. Modules are matched across contexts by gene overlap, and $H_g$ is computed as above. To imitate "bulk," we build pseudo-bulk samples that deliberately mix cell states in varying proportions. These give a same-data control for heterogeneity masking.
Evaluation uses held-out interventional data only. Networks are built only from non-targeting control cells. Each knocked-down gene's ground-truth influence is the energy distance (E-distance) between its perturbed cells and control cells, a measure of how far apart two cell populations are, with the permutation E-test from scPerturb. A second ground-truth score measures module specificity: whether knocking down a hub moves its own module's eigengene more than it moves other modules. That second measure tests the WGCNA premise that modules are causal units, not only whether hubs matter.
Retrospective contexts come from the file's resources. The first are the K562 and RPE1 CRISPRi screens in CausalBench, with about 162,000 cells each and 622 and 383 interventions respectively. The second are the additional Perturb-seq lines used by scDEBART, HepG2 and Jurkat. The third are the stimulus- and stress-bearing screens harmonized in scPerturb, including Adamson's unfolded-protein-response screen and Papalexi.
These datasets vary cell type more than environmental stress. The retrospective phase therefore tests the invariance principle only under that limitation. The file lists no public screen that crosses gene knockdowns with a panel of environmental stresses, and that absence is Gap 2 in the file. We fill it with one small prospective screen, described below.
Baselines
The naive reference points are random ranking and ranking by mean expression in control cells. Highly expressed and essential genes tend to produce large knockdown effects whatever their network position, so expression level is the confound any hub score must beat. We add an essentiality proxy: ranking by how much knockdown reduces the number of cells recovered. Any method that fails to beat expression plus essentiality has learned nothing about regulation.
The standard practice baseline is ordinary WGCNA hub ranking by intramodular connectivity. It is run once on mixed-state pseudo-bulk and once on metacells from a single context. This is what published hub papers do, for example the Arabidopsis paralog study. The pooled baseline builds one network on all contexts concatenated. It isolates whether taking the minimum across contexts matters, as opposed to simply using more data.
We add a non-WGCNA observational competitor: out-degree in a gene regulatory network inferred from the same control cells by the observational methods in the CausalBench suite, with feedback loops allowed. CausalBench found that simple observational methods compete with causal discovery methods, and cyclic structure matters because biological networks contain feedback. Every method gets the same control cells, the same gene filter, and the same tuning budget. Hyperparameters, including WGCNA power and metacell size, are chosen on K562 and frozen for all other contexts.
The strongest learned competitor is scDEBART, pretrained on 6.28 million differential expression profiles. Its predicted response magnitude for each held-out gene serves as a ranking score. It reports a mean Enrichment Factor of 11.96 on Perturb-seq benchmarks and 71.4% top-1 reverse-knockout accuracy. Those are predictions of which genes respond to a knockdown, not rankings of which knocked-down genes have the largest effects. For the task here, ranking genes by knockdown E-distance from observational data, we found no reported result on Lacuna.
Beating these baselines requires two things. The first is a statistically higher Spearman correlation and precision-at-50 against E-test-significant knockdowns than both expression-plus-essentiality and single-context hubness, and this must hold in at least three of the four cell lines. The second is a gain in partial correlation after regressing out expression and essentiality. The key ablations remove one ingredient at a time: minimum-across-contexts versus pooled, and metacell versus pseudo-bulk.
Why not just…?
Why not just rank genes by differential expression or by expression level? This is the most serious threat. Hubs in any correlation network tend to be highly expressed, and expression level alone may explain most of the ranking signal. MONSTER argues that regulators often rewire their targets without being differentially expressed themselves, which would favor network scores. That argument comes from observational cohorts, not from knockdowns. We settle the question directly with partial correlations conditioned on expression and essentiality. If invariant hubness adds nothing after conditioning, we will report that co-expression hubness is an expression proxy. That negative result would itself be useful to the many hub-validation papers.
Why not just use a pretrained perturbation model like scDEBART? It may rank effect sizes well. However, it is trained on broad perturbation corpora that may overlap with the benchmark screens, and it yields no module structure that would explain why a gene matters. We run it head to head and report where it wins. The proposal's distinct contribution is the module-specificity test, which no response predictor addresses. If scDEBART dominates on ranking, the claim narrows to what invariant hubs say about module causality.
Why not just pool all conditions into one bigger network? Pooling adds samples and might recover the same hubs. It also adds between-condition variance, which creates modules that track the condition itself: the "generic crisis" signal that single-stress studies cannot distinguish from specialized responses. The pooled-versus-minimum ablation measures exactly this. If pooling matches the minimum, the invariance idea is unnecessary and we drop it.
Why not just deconvolve bulk data instead of using metacells? Deconvolution methods such as POPPY estimate cell composition. They do not recover within-cell correlation structure, and they depend on reference atlases. The pseudo-bulk arm shows directly how much mixing states degrades hub validity on the same cells.
Proposed Experiments and Decisive Checks
The retrospective phase runs the full comparison on K562, RPE1, HepG2, and Jurkat. Contexts are cell lines, and within each line, cell-cycle and stress-response metacell states. Scores are also computed on the Adamson and Papalexi screens as stimulus contexts.
The first decisive check is conditional. Does $H_g$ predict knockdown E-distance beyond expression and essentiality? The second is about module specificity. Do high-$H_g$ hubs shift their own module eigengene more selectively than expression-matched non-hubs? A clean negative on both checks falsifies the core claim.
The prospective phase is a pre-registered screen in RPE1 designed to follow the scPerturb power guidelines of at least 300 cells per perturbation and about 1,000 unique molecular identifier (UMI) counts per cell. It targets 60 genes. Twenty are top invariant hubs computed from existing data, 20 are hubs in a single context but not invariant, and 20 are non-hubs matched on expression level. Each gene is knocked down under five conditions: control, heat, sorbitol osmotic stress, hydrogen peroxide, and heat combined with osmotic stress. That design comes to roughly 90,000 cells.
Predictions are frozen before sequencing. Invariant hubs should produce larger and more module-specific effects in every condition, including the combined stress, which was never used to build $H_g$. Context-specific hubs should matter only where they were hubs. This screen directly measures the cross-stressor division of labor that single-stressor studies cannot see. It also turns WGCNA hub-picking from a literature-matching exercise into a prediction that can fail.
References
- Single-Cell Co-expression Analysis Reveals Distinct Functional Modules, Co-regulation Mechanisms and Clinical Outcomes
- Correlation Analysis Connects Cancer Subtypes
- Genome-Wide Identification of the GhANN Gene Family and Functional Validation of GhANN11 and GhANN4 under Abiotic Stress
- The differential expression patterns of paralogs in response to stresses indicate expression and sequence divergences
- Library size-stabilized metacells construction enhances co-expression network analysis in single-cell data
- scPerturb: Harmonized Single-Cell Perturbation Data
- A large-scale benchmark for network inference from single-cell perturbation data
- scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning
- Estimating drivers of cell state transitions using gene regulatory network models
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level