When a researcher sequences a whole chunk of heart tissue, the readout is an average over millions of cells — thickening muscle cells, expanding scar-forming fibroblasts, infiltrating immune cells. In a heart that is progressively failing, the mix of those cells changes drastically from week to week, and the average hides which cells are actually driving the disease. This is not a cosmetic problem: two analyses in the literature show that computing gene co-expression networks on such averages manufactures false "hub genes" that simply track cell-population shifts rather than regulatory control. The pressure-overload field has the right experiment and the right perturbation tool but has never connected them through a cell-resolved lens. Studies of the transverse aortic constriction (TAC) model — a surgical tightening of the aorta that pushes a mouse heart from compensatory hypertrophy, to heart failure with preserved ejection fraction (HFpEF), to decompensated failure over eight weeks — have identified hub genes from bulk ventricular tissue, but never resolved them by cell type and never validated them genetically. Conversely, studies that do knock genes out conditionally in specific cardiac cell types pick their targets by hypothesis, not by network analysis. The proposal here is to close that loop: deconvolve a staged TAC timecourse into cell-type-resolved profiles, build cell-type-specific networks whose hubs escape the known biases of global degree centrality, and validate the top predicted hubs by cell-type-restricted Cre-loxP knockout in mice subjected to TAC — then reprofile the knockouts to see, for the first time, both the autonomous and the bystander network consequences of deleting one gene in one cell type.
Core Claim
The claim is that the hub genes reported from bulk pressure-overload transcriptomics are substantially a compositional artifact, and that a deconvolution-first pipeline will surface a different, largely non-overlapping set of cell-type-restricted drivers — a subset of which will be causally validated by conditional knockout while bulk-derived hubs will not. This claim is testable at every stage: against the published bulk analysis, against a positive control gene already validated by conditional knockout, and against the explicit tension raised by work in cancer showing that phenotype-relevant drivers can be anti-correlated with global network centrality.
The motivation comes from three converging lines in the file. Xiang Li et al. profiled bulk ventricular tissue at 2, 4, and 8 weeks post-TAC and built protein-interaction and competing-endogenous-RNA networks to find stage-specific hub genes, reporting "199 DEGs were found in the 2W-post-TAC group, while 376 were identified in the 8W-post-TAC group" — but, as the authors themselves note, all of it on homogenized tissue, with no cell-type resolution and no genetic validation Differential mRNA Expression and Circular RNA-Based Competitive Endogenous RNA Networks in the Three Stages of Heart Failure in Transverse Aortic Constriction Mice. Separately, SNAIL demonstrated that standard normalization of heterogeneous bulk RNA-seq inflates false-positive co-expression links to 8.6% at a correlation threshold of p=0.3, and that on Mouse ENCODE spike-ins the correlation-coefficient error is 0.03856 RMSE — artifacts that a fibrotic, leukocyte-rich heart is maximally positioned to trigger Adjustment of spurious correlations in co-expression measurements from RNA-Sequencing data. And SubNetDL found that in cancer, genes predictive of therapy response are inversely correlated with standard degree centrality, because high-degree hubs are generic pan-tissue infrastructure rather than context-specific regulators A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer. Nobody has tested whether the same inversion holds in pressure-overload cardiomyopathy once cell types are separated out.
Research Program
The program has four phases, each grounded in datasets and tools documented in the file.
The first phase builds the cell-resolved reference that the deconvolution field currently lacks for this disease. The open problem for every deconvolution method is reference-atlas coverage: if the diseased or transitional cell state is absent from the reference, attribution degrades — a limitation stated explicitly for both POPPY and Count Bridges. The available single-cell heart-failure reference, GSE183852 (220,752 nuclei and 49,723 cells across 14 major cell types, used by the PathFinder study), is human and end-stage, not staged murine pressure overload Decipher macrophage-fibroblast-cardiomyocyte signaling interactions associated with heart failure using deep graph neural network models and single-cell RNA-seq data. So we will generate a matched single-nucleus RNA-seq reference from C57BL/6J mice at sham, 2-week, 4-week, and 8-week post-TAC timepoints, matching the design and staging of the public bulk timecourse GSE182985, including the fibroblast and myeloid activation states that a remodeling heart contains and healthy atlases lack. Using the same sub-strain and surgical staging as the bulk data also directly addresses the sub-strain variability (C57BL/6J versus C57BL/6N) that Li et al. flagged as a generalizability limit.
The second phase applies count-based deconvolution to the public staged bulk data. Count Bridges — a generative model that operates on integer-valued expression counts using Poisson birth-death dynamics — is the natural tool here because it "outperformed established methods like CIBERSORTx and MuSiC in recovering cell-type proportions" on bulk RNA-seq Count Bridges enable Modeling and Deconvolving Transcriptomic Data; its treatment of counts as integers matters for the small, stage-specific bulk cohorts in GSE182985. POPPY's ontology-aware cross-modal alignment provides a second, independent deconvolution path, with reported gains of up to 73% in silhouette-score integration and up to 714% in composition-similarity correlation over standard bulk encoders Cross-modal transfer learning for mapping bulk transcriptomes at cellular level. Running both, and requiring that hubs recovered by the two independent deconvolvers agree, is a built-in robustness check rather than a separate experiment.
The third phase constructs cell-type-specific co-expression networks on the deconvolved profiles, with SNAIL-style spurious-correlation adjustment applied before any edges are drawn, and ranks candidate hubs by betweenness centrality within cell-type sub-networks rather than by global degree. The methodological precedent is DeepCSCN, which built cell-type-specific co-expression networks directly from scRNA-seq and outperformed GENIE3, CLR, and CS-CORE on eight benchmark datasets, identifying 280 cell-specific modules across a GTEx atlas of over 140,000 cells Deep Learning Driven Cell-Type-Specific Embedding for Inference of Single-Cell Co-expression Networks. Our departure from DeepCSCN is that its networks came from resting human tissue; ours are wired onto a longitudinal in vivo disease trajectory. For cross-cell-type signaling edges we adapt the Graphormer-based ranking approach of PathFinder, which reported AUC scores reaching 0.92 for cardiomyocytes and 0.83 for M2 macrophages in classifying healthy versus failing cells on GSE183852 — but PathFinder analyzed a static end-stage snapshot of human dilated cardiomyopathy, whereas we ask how the ranked cross-talk hubs move between 2, 4, and 8 weeks of controlled pressure overload. A key analytical commitment, following the SubNetDL finding, is to score hubs both ways — global degree versus cell-type-restricted betweenness — and treat disagreement between the two rankings as a primary result, not a nuisance.
The fourth phase is the causal loop, and it is where this proposal departs from everything in the file. We will select the top predicted hub for cardiomyocytes and the top predicted hub for fibroblasts or myeloid cells at the 4-week HFpEF transition — the stage where intervention plausibly still matters — and generate floxed alleles crossed to Myh6-Cre and Lyz2-Cre or a fibroblast Cre driver respectively, then subject the conditional knockouts and littermate controls to 4-week TAC with echocardiographic, histological (WGA, Masson's trichrome), and molecular endpoints, following the exact phenotyping template of the DUSP9 study, which used cardiomyocyte-specific knockout, transgenic overexpression, and TAC surgery to establish causality for a single candidate gene Dual-specificity Phosphatase 9 protects against Cardiac Hypertrophy by targeting ASK1. Two design details distinguish our validation from that study. First, our target is chosen by an unbiased, cell-resolved network pipeline rather than by prior liver-metabolism literature. Second, we then reprofile the conditional-knockout hearts post-TAC with single-nucleus RNA-seq and rebuild the networks, so we can attribute which downstream module shifts occurred inside the deleted lineage and which occurred in non-ablated bystander cells — the autonomous-versus-non-autonomous attribution that Ying Sun et al. explicitly could not make when they combined parvalbumin-neuron-specific Efnb2 knockout with bulk-tissue WGCNA and had to concede that their behavior-correlated MEgreen module might reside in surrounding cells rather than the targeted interneurons RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons. Sun et al. bring this design to the heart, to pressure overload, and to a hub chosen computationally rather than by hypothesis.
A built-in positive control makes the pipeline falsifiable before any new mouse is bred: Dusp9. Jiang et al. showed that DUSP9 protein rises nearly eight-fold at 4 weeks post-TAC and that cardiomyocyte-restricted deletion worsens hypertrophy, ejection fraction, and fibrosis under TAC. Our pipeline must recover Dusp9 as a cardiomyocyte-restricted hub at the 4-week stage from the deconvolved GSE182985 data. If it does not, the pipeline is wrong, not the literature. Symmetrically, a negative control comes from the bulk analysis itself: a top bulk-degree hub that our cell-resolved pipeline demotes to a compositional artifact should, when its cell-type-specific attribution predicts it is not a driver in any single lineage, show no conditional-knockout phenotype — a prediction we will test for one such gene.
Proposed Experiments
The computational experiments run first and cheaply. On GSE182985 (with GSE182912 circRNA microarray data as a secondary, orthogonal input), we deconvolve each staged sample into cardiomyocyte, fibroblast, endothelial, and myeloid profiles with Count Bridges and POPPY, validate deconvolution accuracy against our matched single-nucleus reference, then build SNAIL-adjusted, cell-type-specific co-expression networks and rank hubs by within-cell-type betweenness. The decisive quantitative comparisons are, first, hub-list overlap with the bulk-derived stage hubs of Li et al. (we predict low overlap at 8 weeks, where fibrosis and infiltration are worst), and, second, positive-control recovery of Dusp9 at 4 weeks. For cross-cell-type signaling hubs, we compare our trajectory-ranked targets against PathFinder's end-stage-ranked targets (BMP4, IGF1) on our data, asking whether trajectory information reorders the list that a static snapshot produces. Metrics are hub-recovery AUROC against the Dusp9 positive control, module-trait correlation with ejection fraction and fibrosis fraction across stages, and the betweenness-versus-degree rank inversion statistic. The in vivo experiments then test the two selected hubs: conditional knockout versus floxed littermate controls under 4-week TAC, with pre-registered endpoints of ejection fraction change, fibrotic area, and the single-nucleus reprofile measuring the fraction of downstream module disruption occurring inside versus outside the deleted lineage.
Baselines
The do-nothing reference is the published bulk pipeline itself: smooth-quantile-normalized bulk RNA-seq of staged TAC ventricles with WGCNA and protein-interaction degree centrality, exactly as Li et al. ran it on GSE182985, reporting 199 differentially expressed genes at 2 weeks and 376 at 8 weeks. It is the strongest competitor on this question because it is the incumbent: it is what the field currently believes the hub genes of pressure overload to be, and beating it means showing that its hub list fails causal validation where ours succeeds. We will rerun it on the same public data with the same compute and the same gene-universe filters as our pipeline, and score both pipelines on identical criteria: Dusp9 recovery, stage-specificity of hubs, and — the decisive tiebreak — conditional-knockout phenotype. A hub counts as beaten only if the bulk pipeline promotes it and it fails the knockout, or our pipeline promotes a gene the bulk pipeline misses and it passes.
The strongest existing method for the deconvolution stage is classical linear regression deconvolution with CIBERSORTx and MuSiC run on the same GSE182985 data with the same single-nucleus reference, the same holdout scheme, and the same tuning budget as Count Bridges and POPPY. The file reports that Count Bridges outperformed both on cell-type proportion recovery, but gives no reported result found on Lacuna for a numeric CIBERSORTx or MuSiC accuracy value, so we treat this as an open empirical comparison on this dataset rather than a settled one; proportion-recovery error against our matched reference is the metric. For the network-ranking stage, the strongest existing competitor is PathFinder, which reported AUC scores reaching 0.92 for cardiomyocytes and 0.83 for M2 macrophages on GSE183852. We will run PathFinder on our staged single-nucleus data unchanged, so the comparison isolates exactly one ingredient: whether longitudinal trajectory information changes which cross-talk hubs are prioritized relative to PathFinder's static-snapshot ranking.
The key ablation strips the proposal's central ingredient in two steps. Run the full pipeline but skip deconvolution, building networks on the raw bulk staged data (this quantifies how much of the hub-list change comes from cell separation alone versus from SNAIL-style correlation adjustment). Then run the full deconvolved pipeline but rank hubs by global degree centrality instead of within-cell-type betweenness, which isolates the SubNetDL-motivated claim that degree and betweenness select different genes. If the degree-ranked deconvolved list matches the betweenness-ranked list, the centrality claim collapses and the proposal reduces to "deconvolve first," which the ablation will have honestly established.
Why not just…?
Why not just run standard bulk WGCNA across the TAC timepoints? Because the file contains direct quantitative evidence that this manufactures false hubs: on GTEx, standard normalization inflated false-positive co-expression links to 8.6% at a correlation threshold of p=0.3, which SNAIL cut to 0.58%, and on Mouse ENCODE spike-ins the correlation RMSE fell from 0.03856 to 0.01516 Adjustment of spurious correlations in co-expression measurements from RNA-Sequencing data. A heart undergoing fibrosis is the worst case for this artifact, since fibroblast and myeloid fractions change most where the disease changes most. That said, this alternative is a real threat to the magnitude of the claimed effect, and the ablation above — bulk pipeline versus deconvolved pipeline on identical data — is precisely the experiment that will measure how large the gap actually is rather than assert it.
Why not just use CIBERSORTx for the deconvolution step? It might well suffice for coarse proportion estimates, and it is cheaper and more familiar. The file reports that Count Bridges outperformed CIBERSORTx and MuSiC in recovering cell-type proportions, but with no reported numeric value found on Lacuna, so we do not claim a settled advantage; instead we run all three on the same data with the same reference and let the proportion-recovery numbers decide. If CIBERSORTx matches Count Bridges on this dataset, we will say so and use the simpler tool — the proposal's causal claims about hub genes do not depend on which deconvolver wins, only on the cell separation happening at all.
Why not just pick the highest-degree nodes from the protein-interaction network and knock those out? Because the one study in the file that directly tested this logic found the opposite: SubNetDL showed that phenotype-relevant drivers were inversely correlated with global degree centrality, and that shuffling mutations across sub-population architectures collapsed predictive performance, establishing that functional relevance depends on context-specific topology rather than aggregate connectedness A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer. Our negative-control knockout of a bulk-degree hub is designed to settle this in the cardiac setting specifically: if the degree-selected gene also shows a knockout phenotype, SubNetDL's inversion does not transfer to pressure overload and we will report that.
Why not just use a pharmacological inhibitor or a whole-body knockout instead of breeding conditional Cre-loxP lines? Pharmacology bathes every cell type at once, so any observed effect cannot be attributed to the cell type where the hub acts — the same confound Shi et al. identified when they noted that ligand-based treatments obscure whether a behavioral defect arises in GABAergic or glutamatergic circuits, a question only cell-type-specific ablation could resolve Shi et al.. Whole-body knockouts risk embryonic lethality and conflate multi-organ effects, as Zhang et al. showed when only macrophage-restricted Med1 deletion could localize the ROS driver of intimal hyperplasia that global models blur MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation. And since our core claim is that hubs are cell-type-restricted, a perturbation that cannot respect cell-type boundaries cannot test it. This is also where our proposal differs most sharply from the morusin study, which reached single-cell resolution but validated only with a systemic compound in a chemical-injury aorta model, never a conditional knockout in pressure overload Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection.
Decisive Checks
Three findings would each independently establish the core claim: the deconvolved hub list diverges substantially from the bulk-derived list at 8 weeks, the pipeline recovers Dusp9 as a cardiomyocyte hub at 4 weeks without ever being told about it, and the conditional knockout of a pipeline-selected hub — but not of a bulk-degree negative control — alters ejection fraction and fibrosis under TAC. Conversely, the clearest falsification would be high overlap between bulk and deconvolved hub lists with the degree-ranked ablation matching the betweenness-ranked one, which would reduce this program to a confirmation that, in the pressure-overloaded heart, averaging over cells costs less than the broader literature suggests it does elsewhere.
References
- Differential mRNA Expression and Circular RNA-Based Competitive Endogenous RNA Networks in the Three Stages of Heart Failure in Transverse Aortic Constriction Mice
- Adjustment of spurious correlations in co-expression measurements from RNA-Sequencing data
- A network-based deep learning model integrating subclonal architecture for therapy response prediction in cancer
- Decipher macrophage-fibroblast-cardiomyocyte signaling interactions associated with heart failure using deep graph neural network models and single-cell RNA-seq data
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- Deep Learning Driven Cell-Type-Specific Embedding for Inference of Single-Cell Co-expression Networks
- Dual-specificity Phosphatase 9 protects against Cardiac Hypertrophy by targeting ASK1
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- Hippocampal µ-opioid receptors on GABAergic neurons mediate stress-induced impairment of memory retrieval
- MED1 Deficiency in Macrophages Accelerates Intimal Hyperplasia via ROS Generation and Inflammation
- Integrated Network Pharmacology, Single‐Cell Transcriptomics Unveil the Mechanistic Role of Morusin in Aortic Dissection