Biologists build gene networks from single-cell RNA sequencing (scRNA-seq), which measures which genes are active in each individual cell, by linking genes whose activity rises and falls together across cells. These networks have a small number of hubs: genes connected to hundreds of others. Hubs then dominate whatever is computed downstream, including candidate drivers, interaction signatures, and cell clusters. We propose that many of these hubs are artifacts of mixing. When cells from different types, cell-cycle phases, or transitional states are analyzed together, any gene that differs between those groups appears correlated with every other gene that also differs between them. The result is a hub that reflects which cells were sampled rather than any regulatory interaction. Bulk tissue averaging hides cell types entirely. Pooling cells hides them one level down, inside the co-expression statistics.
The field currently treats hubs in one of two ways. Some methods keep them as biology. Others shrink them with a uniform penalty that ignores why a gene became a hub. We propose a third option: split each gene's connectivity into three parts. One is explained by variation in cell state. One is explained by technical detection patterns. The remainder is dependence that persists among cells in the same state. Hub correction then becomes a measurement instead of a tuning choice. A gene whose connections disappear once cell state is held fixed is a "composition hub" and should be discounted. A gene whose connections survive is a candidate true regulator and should be kept, even though a degree penalty would have suppressed it.
Core Claim
The claim is that a substantial share of hub interference in single-cell co-expression networks comes from cellular heterogeneity inside the pooled sample. Correcting for it requires conditioning on continuous cell state, not penalizing degree. For a gene $g$, write its degree in the pooled network as $d_g$. We define
$$ d_g = d_g^{\text{state}} + d_g^{\text{tech}} + d_g^{\text{within}}, $$
where each term is defined by nested estimates. $d_g^{\text{within}}$ is the number of significant partners after conditioning on a continuous latent cell state and on a detection-matched null. $d_g^{\text{tech}}$ is the additional degree lost when only the detection-matched null is removed, with state left unconditioned. $d_g^{\text{state}}$ is the remainder. Because the terms are nested differences, the split depends on ordering, and we will report both orderings.
The prediction is that genes with the largest $d_g$ are mostly high-$d_g^{\text{state}}$ genes. We further predict that ranking candidates by $d_g^{\text{within}}$ removes the correlation between degree and predicted importance without losing recall of experimentally supported edges. That correlation is the degree bias reported for knowledge graphs, where it reaches $R^2$ of 0.7 to 0.8 in Implications of topological imbalance for representation learning on biomedical knowledge graphs.
This addresses a contradiction in the literature. Hub-penalizing methods treat high-degree nodes as noise sinks. These include Prize-Collecting Steiner Forest (PCSF) solvers with explicit degree penalties (Omics Integrator, PCSF), the normalized random walks of SLN-SRW, and SubNetDL, whose top genes are inversely correlated with centrality. By contrast, scPGI-finder presents its lung adenocarcinoma interaction networks as organized around hubs such as MLF2 and CTLA4 partners. scPGI-finder scores gene pairs by co-activation linked to proliferation scores, and proliferation is itself a cell-state axis. Our decomposition asks directly which of those hubs survive once cell-cycle and lineage state are held fixed.
Research Program
We would build a conditional co-dependence estimator with three stages.
Continuous cell-state coordinate. We first learn a continuous cell-state coordinate for every cell using a low-dimensional embedding. The embedding is recomputed with gene $g$ held out whenever we score edges involving $g$, so that a gene cannot condition away its own signal. We use a continuous coordinate rather than clusters because intermediate states are real and common. Examples include hybrid epithelial–mesenchymal cells (Intermediate cell states in epithelial-to-mesenchymal transition) and interneuron subtypes that lie on gradients (Diversity of Interneurons in the Dorsal Striatum). Discrete clusters would leave composition variation inside each cluster.
Local dependence testing. Second, for each gene pair we test dependence within local neighborhoods in state space. This reuses the local-window idea of the Cell-Specific Network (CSN) method (Cell-specific network constructed by single-cell RNA sequencing data), which builds a separate gene network for every cell. The difference is that our windows are defined over cell state rather than over expression scatter alone.
Detection-matched null. Third, the null distribution permutes values within neighborhoods while preserving each gene's pattern of zero counts. This targets the concern that genes detected in many cells, or genes sharing zero patterns, gain spurious dependencies. scRNA-seq imputation work such as scBFP treats these technical zeros ("dropouts") as a central problem.
The output is a per-gene triple ($d_g^{\text{state}}, d_g^{\text{tech}}, d_g^{\text{within}}$) and a within-state network.
Program hubs. One conceptual risk must be handled explicitly. A true master regulator that drives a state change will have much of its degree assigned to the state component. We will therefore report high-state, low-within genes as "program hubs," not as artifacts. We then test separately whether they carry causal support in perturbation data. The decomposition sorts hubs into categories. It does not automatically delete any of them.
Downstream re-analyses. We will re-run two published pipelines with the within-state network substituted for the pooled one. For CSN, we examine whether its degree matrix, which it uses in place of expression for clustering, is inflated by composition and detection hubs. For scPGI-finder, we examine whether its hub-centered signatures survive.
Baselines
Naive reference. Pooled Spearman co-expression across all cells, thresholded to the same edge density as every other method. This measures how much hub structure appears with no correction at all.
Two-step cluster-then-network. Leiden clustering followed by per-cluster co-expression and GENIE3. This is the standard practice. It is also the most natural correction for composition hubs, which makes it the most direct competitor for our first stage.
Joint mixture model. VMPLN (Single-cell gene regulatory network analysis for mixed cell populations) clusters cells and estimates per-type sparse precision matrices together. Its authors report that it beats two-step pipelines under heavy mixing, although the file gives no numeric pAUPRC. It is the strongest existing method that models heterogeneity inside network inference. Its limitation for our question is that it assumes discrete types.
Uniform hub penalty. Degree-penalized alternatives applied to the pooled network: PCSF-style $\mu$ penalties and Laplacian-normalized propagation as in SLN-SRW. SLN-SRW reports an AUC of 0.79 on its integrated network. These test whether any uniform downweighting of hubs matches our targeted correction.
Covariate regression. Regressing out standard covariates (cell-cycle scores, sequencing depth, detection rate) before computing co-expression. This is the cheapest plausible competitor.
Foundation-model inference. UGRN (Towards Universal Gene Regulatory Network Inference) extracts regulatory links from a frozen scGPT model. It reports an average AUPRC of 0.768 on scBenchmark, against 0.584 for zero-out perturbation. We will re-run UGRN on the same BEELINE splits rather than import that figure, because the benchmarks differ.
Ablations. Three ablations isolate the key ingredient: state conditioning without the detection null, the detection null without state conditioning, and discrete-cluster conditioning in place of continuous state.
Fairness conditions. All methods receive the same cells, the same gene set, and the same final edge count. Each receives an equal hyperparameter search budget, tuned on held-out cell types.
Criterion for beating the baselines. We claim success if our method matches or exceeds the best baseline's AUPRC against perturbation ground truth. It must also cut the Spearman correlation between gene degree and gene score by at least half relative to the naive reference.
Why not just…?
Why not just cluster first and build networks within each cluster? This is the real threat. Within-cluster networks may already remove most composition hubs. VMPLN's authors show that the two-step pipeline propagates clustering errors when types overlap, but they did not measure hub counts. Our ablation comparing discrete clusters with continuous state settles this directly. The comparison is made on continuum-rich data (EMT-like and cell-cycle-dominated populations) and on cleanly separated PBMC data. If clusters perform as well on both, the contribution narrows to the detection-matched null and the diagnostic decomposition. We would report that outcome as such.
Why not just penalize hubs, as PCSF or Laplacian normalization does? A uniform penalty shrinks every high-degree gene equally, including pleiotropic regulators. The Omics Integrator line itself concedes that the penalty is set heuristically. Our perturbation-grounded evaluation will count true regulator edges lost under each penalty strength, compared with those lost under state conditioning.
Why not just regress out cell cycle and depth? Regression removes only the axes someone thought to name. It cannot remove lineage mixing or unnamed transitional states. The in-silico pooling experiment below measures how many pooled-network hubs remain after regression but vanish in pure populations.
Why not just use VMPLN or UGRN? VMPLN assumes discrete types and linear latent dependence. UGRN reports no degree-bias analysis and is evaluated against STRING and ChIP-seq references that carry citation bias. State of the interactomes measures that bias as $r_s = 0.80$ between literature citations and interaction coverage. Both methods enter as baselines under the degree-bias metric.
Proposed Experiments
Simulated mixtures. We simulate cells with known within-type networks and vary three factors: the number of states, the spacing between states (from discrete to continuous), and the dropout rate. Composition hubs are known by construction here, so we can directly measure the precision and recall of hub classification.
Controlled pooling of labeled PBMC data. We use the Kang (7,217 cells, 10 types) and Zheng (7,424 cells, 6 types) blood-cell datasets used by VMPLN. We compute networks on each labeled type alone and on the pooled cells. A composition hub is a gene that is a hub in the pooled network but in no pure network. We then ask what fraction of pooled hubs each method removes, and what fraction of pure-type edges it keeps.
Accuracy and degree bias on BEELINE. On the BEELINE datasets (hESC, hHEP, mDC, mESC, mHSC), we score AUPRC and early precision against loss- and gain-of-function ground truth. We prefer that ground truth over STRING because it is less exposed to literature circularity. We report the degree–score correlation of every method alongside accuracy.
Re-analysis of the lung adenocarcinoma atlas. We re-analyze the Salcher et al. atlas (about 161,000 cells) with scPGI-finder's pipeline. We compute the decomposition for MLF2, the CTLA4 partners, and the top hub genes. We then ask whether signatures rebuilt from within-state interactions retain the reported bulk performance: a mean AUROC of 0.974 across eight lung adenocarcinoma cohorts, and AUROC above 0.80 for the six-pair T-cell signature on immunotherapy cohorts. If corrected signatures keep that performance while dropping composition hubs, the hubs were not necessary. If performance collapses, the hubs mark proliferative state rather than genetic interaction. That is itself a substantive finding about what scPGI-finder measures.
CSN clustering check. We repeat CSN's clustering evaluation on the Buettner dataset (ARI 0.99 from its degree matrix versus 0.48 from expression) using within-state degrees. This tests whether CSN's advantage depended on state-driven degree. On a cell-cycle dataset, that dependence would be expected.
References
- Implications of topological imbalance for representation learning on biomedical knowledge graphs
- Network-Based Interpretation of Diverse High-Throughput Datasets through the Omics Integrator Software Package
- PCSF: An R-package for network-based interpretation of high-throughput data
- Predicting disease-related genes using integrated biomedical networks
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Charting cell-type-specific positive genetic interaction at single-cell resolution for lung adenocarcinoma
- Intermediate cell states in epithelial-to-mesenchymal transition
- Diversity of Interneurons in the Dorsal Striatum Revealed by Single-Cell RNA Sequencing and PatchSeq
- Cell-specific network constructed by single-cell RNA sequencing data
- Single-cell RNA sequencing data imputation using bi-level feature propagation
- Single-cell gene regulatory network analysis for mixed cell populations
- Towards Universal Gene Regulatory Network Inference: Unlocking Generalizable Regulatory Knowledge in Single-cell Foundation Models
- State of the interactomes: an evaluation of molecular networks for generating biological insights