Network biologists find "hub genes", the genes most connected in an inferred regulatory graph, and treat them as likely disease drivers. Every current method builds that graph with one node per gene. Many genes, however, make several transcript variants ("isoforms") with opposite jobs. Tumor myeloid cells switch FLT-1 from a soluble decoy receptor to a membrane signaling receptor. Liver tumors raise two MEIS2 isoforms while total MEIS2 looks unchanged. Merging such variants into one node blends opposing edges and can promote or bury the wrong gene. Single-cell long-read sequencing now reads full-length transcripts in individual cells, so the data to fix this exist. Nobody has yet built a network on it.
We propose a single-cell network model that gives each multi-isoform gene two kinds of node. One tracks how much the gene is transcribed. The others track how the transcribed RNA is split among its isoforms. Cell types are clustered at the same time as the network is estimated. Hubs can then be classified as transcriptional hubs, where the gene's output is wired into many others, or splicing hubs, where the isoform choice is wired into many others. Edges pointing into splicing nodes become candidate upstream triggers of isoform switches. These are exactly the regulators that isoform studies say are missing. The central empirical claim is testable: a meaningful set of splicing hubs sits on genes that gene-level methods rank as peripheral. This proposal succeeds or fails on whether those hubs survive perturbation-based checks.
Why this gap is real
Four lines of evidence motivate the node redefinition. Gene-level aggregation cancels signal. Isoform-level clustering of 160 exon-array datasets separated cancer from normal lines at 97.5% purity, against 75% at gene level, because reciprocal isoform changes cancel out (Isoform level expression profiles provide better cancer signatures than gene level expression profiles). MEIS2C binds CDC73 to drive Wnt signaling, and MEIS2D silences LATS1 to release YAP, yet the authors report the upstream signals that choose these isoforms as unknown (MEIS2C and MEIS2D promote tumor progression via Wnt/β-catenin and hippo/YAP signaling in hepatocellular carcinoma). AR-V7 runs a metastatic SOX9 program that full-length AR does not (Androgen receptor splice variants drive castration-resistant prostate cancer metastasis by activating distinct transcriptional programs). A single "AR" node would credit AR-FL with that program. The FLT-1 decoy-to-signaling switch in tumor myeloid cells is visible only after sorting (Identification of Reprogrammed Myeloid Cell Transcriptomes in NSCLC).
The methods we build on have a complementary blind spot. scGIR runs a PageRank-style walk on per-cell gene correlation networks and finds "dark genes": hubs without differential expression. VMPLN fits a mixture Poisson log-normal model that clusters cells and estimates cell-type-specific sparse precision matrices (the matrix of partial correlations, whose nonzero entries are network edges) in one step. Its preprint ( Single-cell gene regulatory network analysis for mixed cell populations with applications to COVID-19 single cell data) shows that clustering first and inferring networks afterward propagates clustering errors. ESPACE relaxes the sparsity penalty on known hubs. All three take one count per gene.
The data side has the opposite profile. The long-read colorectal atlas catalogues 125,205 isoforms and 394 dysregulated transcript structures in tumor epithelium across 12 patients (An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing). The ovarian HIT-scISOseq study captures 152,546 isoforms at 12,000 reads per cell. It also shows that gene-level counting overestimates protein-coding expression by about 20% (Detection of isoforms and genomic alterations by high-throughput full-length single-cell RNA sequencing in ovarian cancer). Neither study infers any network. VMPLN supplies our estimation machinery, and scGIR and ESPACE are the hub-scoring and prior-injection techniques we compare against or reuse. The isoform studies supply the data and the motivating cases.
Core Claim
Treating each isoform as its own gene does not solve the problem. Sister isoforms of one gene share a promoter, and they share ambiguous reads during quantification. Their proportions are also forced to sum to one. These properties produce strong within-gene correlations that are artifacts of bookkeeping, not regulation, and they inflate the degree of every multi-isoform gene. The key ingredient is to separate these two sources of variation explicitly.
For gene $g$ in cell $i$ with library size $s_i$, let the total count be $y_{ig} \sim \text{Poisson}(s_i e^{z_{ig}})$. Let the isoform counts given the total be multinomial with probabilities $\text{softmax}(0, u_{ig})$. Here $z_{ig}$ is latent transcriptional activity and $u_{ig}$ holds log-ratios of each non-reference isoform against a reference isoform. The latent vectors $(z_i, u_i)$ follow a mixture of Gaussians, with cluster $k$ having its own precision matrix $\Theta_k$:
$$ \max_{\{\Theta_k,\pi_k,\mu_k\}} \; \sum_i \log \sum_k \pi_k\, p(y_i, n_i \mid \mu_k, \Theta_k) \;-\; \lambda \sum_k \sum_{(a,b)\notin \mathcal{W}} |\Theta_{k,ab}|, $$
In this objective, $n_i$ are the isoform counts, $\pi_k$ are mixture weights, and $\mathcal{W}$ is the set of within-gene node pairs. Those pairs are estimated without penalty but excluded from hub scores. Hub scores are then computed separately on $z$-nodes and $u$-nodes. The within-gene exclusion is the piece that isoforms-as-genes lacks, and the ablations below isolate it. Fitting reuses VMPLN's variational ADMM, extended with a multinomial likelihood term.
Research Program
The first stage is simulation. We extend VMPLN's 48 scenarios, which vary mixing, dropout, and graph shape (hub, random, block, scale-free), in three ways. Genes receive two to four isoforms. A planted subset of edges attaches only to usage nodes, including cases where a gene's usage node is a hub while its total is flat. Reads are shared across isoforms at rates matching long-read ambiguity. Ground truth is therefore known for both edge types.
The second stage is real data. We use the CRC atlas and the HGSOC cohort, restricted to genes with at least two isoforms detected in a meaningful fraction of cells. The ovarian cohort has only 2,571 cells, so it will likely need pooling into small metacells. We report results with and without pooling, because pooling blurs cell states.
The third stage turns splicing-hub edges into trigger hypotheses. For each usage hub, we list the incoming edges from transcription-factor and RNA-binding-protein activity nodes. We also check whether the edge persists after conditioning on the target gene's own $z$. That check separates promoter-level triggers, which move usage through transcription, from splicing-factor triggers, which move usage directly.
Baselines
The naive reference is gene-level hub ranking: VMPLN, and separately scGIR, run on summed gene counts from the same long-read cells. That ranking is the status quo the claim argues against. On gene-level data scGIR reports an adjusted Rand index (ARI) of 0.68 on the Chu-time differentiation dataset, against 0.64 for raw expression (scGIR). We rerun Chu-time with our model's gene-activity nodes only, as a check that the extension does not degrade gene-level behavior. Beating it there is not the goal.
The strongest existing competitor is VMPLN with every isoform entered as a separate gene. It inherits joint clustering, and its preprint reports AUPRC (area under the precision-recall curve) up to 21% higher than Glasso, GENIE3, PIDC and others in highly mixed simulations ( Single-cell gene regulatory network analysis for mixed cell populations with applications to COVID-19 single cell data). No reported result was found on Lacuna for any network method on single-cell long-read data. The comparison is therefore new, and we run all methods ourselves. Every method gets the same cells, the same gene and isoform filter, the same number of clusters, and the same sparsity-tuning grid chosen by the same held-out likelihood or stability criterion.
Beating this baseline means two things. First, higher AUPRC on planted usage edges in simulation. Second, and more important, lower precision for spurious within-gene-driven hubs on multi-isoform genes. We also include ESPACE-style penalty relaxation on candidate hubs as a third comparator.
Two ablations isolate the key ingredient. One removes the within-gene exclusion $\mathcal{W}$. The other replaces the multinomial usage model with log isoform counts, which removes the compositional handling. If neither ablation hurts, the decomposition is not doing the work.
Why not just…?
Why not just run a gene-level hub method and then test hub genes for isoform switching? This post-hoc approach only looks for switches within genes already ranked central. MEIS2-type cases, where total expression is flat and isoform choice drives two pathways, would rarely enter the candidate list. The decisive measurement is the fraction of our validated splicing hubs that fall outside the gene-level top ranks. If that fraction is near zero, the post-hoc pipeline is adequate and our claim fails.
Why not just treat each isoform as a separate gene in VMPLN or scGIR? This is the real threat, and it is the main baseline above. We predict it inflates degree for multi-isoform genes through shared-read and sum-to-one artifacts. The simulation with planted shared-read noise tests that prediction directly. If isoforms-as-genes matches our model on usage-edge AUPRC and spurious-hub rate, we will report that the decomposition is unnecessary.
Why not just deconvolve bulk RNA-seq? Deconvolution tools such as POPPY and Count Bridges recover cell fractions and gene-level profiles. Both depend on a reference, and neither models isoform structure. Long-range coupled splicing, which accounts for 299 of 394 CRC dysregulated structures, is not recoverable from short fragments. Bulk deconvolution could still be a downstream consumer of our cell-type splicing hubs, not a replacement for them.
Decisive Checks
Three checks settle the proposal. The first is perturbation recovery. Robust stratification of breast cancer subtypes using differential patterns of transcript isoform expression knocked down YBX1, MAGOH and related factors in MCF7 cells, shifting isoform ratios in 495 of 694 subtype genes. Using those target sets as labels, splicing-factor usage edges from our model should recover them with higher precision than isoforms-as-genes edges or gene-level edges. The knockdowns are bulk and in a cell line, so this checks edge direction, not cell-type specificity.
The second is known-case recovery. Wherever FLT-1, CD44, CASP2 or BIN1 isoforms are detectably expressed in the long-read cohorts, their usage nodes should be the ones carrying cell-type-restricted edges. The third is the headline count: how many splicing hubs appear on gene-level-peripheral genes, with stability across bootstrap resamples of cells and patients. Too few patients to replicate across cohorts is the main risk. We will report CRC-to-HGSOC hub overlap honestly, even if it is low.
References
- Isoform level expression profiles provide better cancer signatures than gene level expression profiles
- MEIS2C and MEIS2D promote tumor progression via Wnt/β-catenin and hippo/YAP signaling in hepatocellular carcinoma
- Androgen receptor splice variants drive castration-resistant prostate cancer metastasis by activating distinct transcriptional programs
- Identification of Reprogrammed Myeloid Cell Transcriptomes in NSCLC
- scGIR: deciphering cellular heterogeneity via gene ranking in single-cell weighted gene correlation networks
- Single-cell gene regulatory network analysis for mixed cell populations
- Single-cell gene regulatory network analysis for mixed cell populations with applications to COVID-19 single cell data
- Single-cell gene regulatory network analysis for mixed cell populations with applications to COVID-19 single cell data
- Enhanced construction of gene regulatory networks using hub gene information
- An isoform-resolution transcriptomic atlas of colorectal cancer from long-read single-cell sequencing
- Detection of isoforms and genomic alterations by high-throughput full-length single-cell RNA sequencing in ovarian cancer
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Robust stratification of breast cancer subtypes using differential patterns of transcript isoform expression