When a gene is deleted in one cell type of a living tissue, the transcriptomic consequences show up everywhere — in the edited cells, but also in every neighboring wild-type cell that receives altered signals from them. Every existing method for reading those consequences mixes the two together. Bulk tissue RNA-seq followed by weighted gene co-expression network analysis (WGCNA — a method that groups genes into coordinated "modules" by raising pairwise expression correlations to a power and clustering the resulting network) reports one blended module structure, and the field cannot tell which part of a perturbed module was dismantled inside the knockout cells and which part is a reactive shift in bystanders. This is not a minor bookkeeping problem: it decides whether a therapeutic strategy should target the edited lineage or its neighbors.
The cleanest demonstration of the gap is the study that deleted Efnb2 specifically in parvalbumin-positive (PV+) interneurons of mouse cortex and showed anxiolytic behavior plus a collapse in excitatory synaptic drive onto those cells, with spontaneous excitatory post-synaptic current frequency falling from 0.5494 ± 0.0846 Hz in wild-type to 0.2572 ± 0.0566 Hz in the knockout RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons. Their bulk WGCNA of medial prefrontal cortex found twelve modules, one of which ("MEgreen", enriched for Ras and insulin signaling) tracked the behavioral phenotype — but because the tissue was homogenized, the authors explicitly could not say whether that Ras/insulin rewiring happened inside the Efnb2-deficient PV+ cells or in wild-type pyramidal neurons and glia reacting to them. The same confound runs through the wider literature: a coronary-artery-disease WGCNA study identified the hub gene RNF181 from seven modules covering 8,525 genes, validated it in mouse cardiotoxicity models, and matched it to human GWAS — yet never determined which cardiac cell type actually drives the module Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease.
Meanwhile, two computational advances have solved the technical reason nobody did this at single-cell resolution, but stopped short of biology. Direct WGCNA on raw single-cell matrices fails because dropouts and library-size variance manufacture spurious correlations; library-size-stabilized metacells fix this by aggregating transcriptionally similar cells before normalization, yielding reproducible cell-type-specific co-expression modules on human cortex data (Synapse syn18485175 and syn21261143) Library size-stabilized metacells construction enhances co-expression network analysis in single-cell data. And cluster-free differential co-expression methods can detect modules whose wiring changes between conditions, validated on morphogen-signaling mutants in embryonic skin DiCoLo: Integration-free and cluster-free detection of localized differential gene co-expression in single-cell data. Both are observational, applied to postmortem or developmental snapshots. Neither has ever been pointed at a conditional knockout with a known whole-animal phenotype. That is the opening this proposal takes.
Core Claim
If we sequence single nuclei (rather than bulk tissue) from a conditional-knockout organ, build co-expression modules separately per cell type using metacell aggregation, and exploit a Cre-reporter allele to distinguish recombined from non-recombined cells in the same tissue, we can attribute each rewired module to either the edited lineage (cell-autonomous) or wild-type bystanders (non-cell-autonomous). The attribution is not just descriptive: it selects the cell type and hub gene for a second, deliberately chosen conditional knockout, whose whole-animal phenotype then tests whether the attributed compartment causally carries the original behavioral and electrophysiological deficit. The claim is that this measure–attribute–re-perturb loop will relocate the mechanistic story that bulk WGCNA got wrong, and that the second knockout will show it.
Research Program
The anchor system is the PV-Cre Efnb2 knockout, because it comes with a published bulk-WGCNA answer to overturn, quantitative phenotypes to correlate against, and a stated single-cell limitation Efnb2 study. We cross the floxed Efnb2 line with a PV-Cre driver plus a Cre-dependent fluorescent reporter allele, so that every nucleus captured by single-nucleus RNA-seq of medial prefrontal cortex can be classified as recombined (Cre-positive, gene deleted) or non-recombined, alongside cell-type annotation (PV interneurons, other interneuron classes, pyramidal neurons, astrocytes, oligodendrocytes, microglia). Reporter transcripts survive nucleus preparation, and the design also directly measures the leaky recombination and mosaicism that the conditional-knockout literature flags as its main technical failure mode.
Within each cell type we construct co-expression modules from metacells rather than raw nuclei, following the library-size-stabilized construction whose simulations proved that raw single-cell correlations are inflated by library-size artifacts LSMetacell, and we run cluster-free differential co-expression between knockout and wild-type animals to catch modules that rewire without changing mean expression DiCoLo. The attribution statistic is deliberately simple. For a module $M$ and cell type $c$, let $W(M,c)$ be one minus the correlation between the knockout and wild-type topological-overlap matrices restricted to that cell type; a module counts as cell-autonomous when its $W$ in the Cre-positive target lineage exceeds its $W$ in every Cre-negative lineage, and as bystander-driven when the inequality runs the other way. Bystander modules in lineages that never express Cre cannot reflect direct deletion by construction, so their rewiring is non-cell-autonomous by elimination.
Module eigengenes — the first principal component of each module's expression, a single summary score per animal per cell type — are then correlated across animals with the two published phenotypes: elevated-plus-maze behavior and the PV+ sEPSC frequency, whose wild-type and knockout values are already on record Efnb2 study. The decisive step is the second-generation perturbation. If, contrary to the bulk reading, the Ras/insulin-enriched rewiring attributes to pyramidal neurons or astrocytes rather than PV+ cells, we delete the top intramodular hub gene of that bystander module specifically in the bystander cell type (a new Cre line crossed to a floxed allele of the hub), leaving Efnb2 intact. If the bystander attribution is causal, this second knockout should reproduce or partially rescue the phenotype without ever touching the originally edited gene. The multi-scale loop closes exactly the way the cross-scale literature demands: a cell-level network claim must survive whole-animal physiology, because cell-culture and bulk-tissue readouts are known to diverge from intact-organism outcomes, as the in-vitro-versus-in-vivo discordance literature repeatedly shows Investigation on Trichothecene-Stimulated Lipid Peroxidation and Toxic Effects of Trichothecenes in Animals, Bone marrow-derived mesenchymal stem cells ameliorate chronic high glucose-induced β-cell injury through modulation of autophagy.
A secondary, deliberately cheap experiment audits the in-silico alternative. Perturbation predictors such as DYNAMO, which represents knockouts as learned linear operators over gene-network embeddings and reaches zero-shot prediction correlation $\rho_\Delta = 0.283$ on unseen genes in cultured K562 cells Gene-Embedding Perturbation Operators for Zero-Shot and Transferable Prediction of Transcriptional Responses, and scDEBART, pretrained on 6.28 million expression-change profiles with a mean enrichment factor of 11.96 scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning, have never been tested against an in-vivo conditional knockout. We feed them the Efnb2 perturbation and score their predicted expression shifts against our measured single-nucleus responses. Given that virtual cell models inflate effect sizes roughly 23-fold and collapse to causal-network F1 scores of 0.041–0.048 across random seeds on the PerturbCausal benchmark Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery, we expect in-vivo performance below their in-vitro numbers, but the measurement itself is the contribution: no reported result exists on Lacuna for any virtual cell model evaluated against in-vivo tissue.
Baselines
The do-nothing reference point is the existing published answer: bulk RNA-seq of medial prefrontal cortex with standard WGCNA, exactly as run in the original study, which produced twelve modules and a stress-correlated MEgreen module enriched for Ras and insulin signaling Efnb2 study. We will re-run this on our own cohort with matched animals, sequencing depth, and analysis parameters, so the only difference is resolution. We beat this baseline if the single-cell attribution changes the compartmental conclusion — for example, if bulk analysis implies PV+ origin while per-cell-type networks show the rewiring concentrated in Cre-negative lineages — and if that changed conclusion is confirmed by the second knockout. Reproducing the bulk result without changing the interpretation would count as a null outcome.
The strongest simple competitor is per-cell-type differential expression on the identical single-nucleus data: aggregate nuclei of each type into pseudobulk samples, run standard differential-expression testing between knockout and wild-type, and rank genes by effect size. This uses the same data, the same animals, and no tuning advantage, so the comparison is structurally fair. It beats us if the top differentially expressed genes in the correct cell type predict the second knockout's outcome as well as our module hubs do. We beat it if the module-eigengene correlation with continuous phenotypes (EPM metrics, sEPSC frequency) is higher cross-validated across animals, and — more decisively — if the hub gene chosen by network centrality, which need not be differentially expressed at all, turns out to matter in the second knockout while the top differentially expressed gene does not. The premise that coordination, not marginal per-gene change, carries the signal is the standard argument for co-expression analysis over single-gene testing Session Introduction.
The key ablation isolates the metacell ingredient: run WGCNA directly on the raw single-nucleus count matrix, skipping library-size stabilization. This is expected to fail by producing inflated, non-reproducible correlations, which is precisely what the metacell simulations demonstrated LSMetacell; the ablation makes that failure empirical rather than cited. A second ablation replaces the cross-condition differential co-expression step with simple module-eigengene comparison, testing whether the cluster-free machinery from DiCoLo earns its complexity on our data.
Why not just…?
Why not just use bulk RNA-seq with standard WGCNA? It is cheaper and deeper, and it is what the anchor study did. But bulk averaging provably destroys the answer this proposal exists to obtain: the Efnb2 authors themselves state they cannot localize the MEgreen rewiring to PV+ cells versus neighbors Efnb2 study, and bulk correlations between genes expressed in the same fluctuating cell type are confounded by tissue composition — the same averaging that rendered rare clonogenic Sca1+ cardiac progenitors invisible in bulk myocardium until single-cell profiling found them PDGFRα demarcates the cardiogenic clonogenic Sca1+ stem/progenitor cell in adult murine myocardium. We will nevertheless run bulk WGCNA on matched tissue as the baseline above, so if bulk somehow reproduces our attribution, the proposal's added value collapses honestly — the second knockout adjudicates which attribution is right.
Why not just predict the knockout effect with a virtual cell model? Because the evidence says the current ones do not carry joint distribution structure: on the PerturbCausal benchmark, no deep virtual cell model beat a sparse linear ElasticNet comparator on causal-network recovery, effect sizes were inflated roughly 23-fold, and the STATE model's apparent F1 of 0.462 on one seed collapsed to 0.041 and 0.048 on others PerturbCausal study. DYNAMO's own performance drops from $\rho_\Delta = 0.283$ to 0.108 for genes lacking regulatory-network coverage, and it was trained and evaluated only on immortalized cell lines DYNAMO. This is a real threat if the models transfer better than expected, which is why we test rather than assume: our in-vivo audit is exactly the experiment that would settle it, and if a model matches our measured single-cell response well, that would be a striking and publishable finding in its own right.
Why not just do differential expression per cell type and skip modules entirely? This is the most serious alternative, and it is cheap. We run it as a fair baseline on identical data, as described above. What it cannot do by construction is rank genes that rewire without changing abundance, or compress thousands of coordinated genes into one phenotype-correlatable score per animal; the Efnb2 result itself illustrates the value, since the phenotypic signal surfaced as a module–stress correlation rather than as a leading single-gene effect Efnb2 study. The second knockout is the honest arbiter: if a differentially expressed gene predicts the causal compartment as well as the hub gene does, the module machinery is not earning its keep, and we will say so.
Why not just run in-vitro Perturb-seq on PV-like cultured neurons? Throughput is incomparable, but the multi-scale literature documents systematic divergence between culture and organism: trichothecene lipid peroxidation seen in isolated preparations failed to track in-vivo toxico-pathology because of clearance and systemic buffering Trichothecene study, and the BM-MSC autophagy mechanism only became credible when the in-vitro rescue was confirmed by restored fasting glucose and β-cell mass in diabetic rats BM-MSC study. A PV+ interneuron has no in-vitro equivalent with its in-vivo afferent drive, which is the very quantity the knockout disrupts (sEPSC frequency halved). Culture data can supplement our attribution, but it cannot test whether bystander rewiring in intact circuitry carries a behavioral phenotype.
Proposed Experiments and Decisive Checks
Experiment one: single-nucleus RNA-seq of medial prefrontal cortex from PV-Cre Efnb2 knockout animals, Cre-reporter littermate controls, and wild-types, with elevated-plus-maze phenotyping of every sequenced animal, plus slice electrophysiology for the sEPSC comparison against the published values of 0.5494 ± 0.0846 Hz (wild-type) and 0.2572 ± 0.0566 Hz (knockout) Efnb2 study. Decisive check: the Ras/insulin rewiring attributes to a specific compartment with a cell-type ranking, and the ranking is stable across a held-out animal cohort.
Experiment two: module-eigengene correlation with continuous phenotypes across animals, single-nucleus data only, cross-validated; compared against the per-cell-type differential-expression baseline and the raw-matrix WGCNA ablation on identical data and compute. Decisive check: network hubs or eigengenes predict phenotype better than differentially expressed genes out of sample.
Experiment three: the second-generation conditional knockout of the attributed compartment's hub gene, with the same behavioral and electrophysiological readouts. Decisive check: reproducing or rescuing the anxiolytic phenotype and shifting sEPSC frequency toward the wild-type value, without deleting Efnb2 — this is the only result that elevates the attribution from description to mechanism, and failure of this check, though disappointing, would itself bound how far single-cell co-expression attribution can go.
References
- RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
- Weighted Co-Expression Network Analysis Identifies RNF181 as a Causal Gene of Coronary Artery Disease
- Library size-stabilized metacells construction enhances co-expression network analysis in single-cell data
- DiCoLo: Integration-free and cluster-free detection of localized differential gene co-expression in single-cell data
- Investigation on Trichothecene-Stimulated Lipid Peroxidation and Toxic Effects of Trichothecenes in Animals
- Bone marrow-derived mesenchymal stem cells ameliorate chronic high glucose-induced β-cell injury through modulation of autophagy
- Gene-Embedding Perturbation Operators for Zero-Shot and Transferable Prediction of Transcriptional Responses
- scDEBART: Predicting in silico Single-Cell Perturbation Responses via Large-Scale Differential Expression Learning
- Virtual Cell Models Inflate Perturbation Effect Sizes and Undermine Causal Gene Regulatory Network Recovery
- Session Introduction
- PDGFRα demarcates the cardiogenic clonogenic Sca1+ stem/progenitor cell in adult murine myocardium