Mice that repeatedly win or repeatedly lose fights for 20 days change collagen gene expression across the brain. Collagens are structural proteins in the extracellular matrix (ECM), the scaffold that holds synapses in place. The changes are region-specific: most collagen genes go up in the hypothalamus and striatum and down in the ventral tegmental area (VTA), the midbrain dopamine hub. The authors who reported this say they cannot tell whether the collagen shift drives the behavioral pathology or simply follows changes in dopamine and serotonin signaling (Aberrant Expression of Collagen Gene Family…). Their study looked at a single day, with three mice per group. Neither of those limits can be fixed by better statistics on the existing data.
We propose a sampling design and analysis built to answer that one question. Animals are sampled at six points along the conflict and deprivation course. Each animal contributes five brain regions and a daily behavioral record. The analysis combines two kinds of evidence. The first is timing: which change comes first. The second is conditional dependence between animals: whether collagen variation in one region still tracks behavior or dopamine-related genes after accounting for everything else measured, including circulating stress hormone. We call the resulting graph a partial correlation network, which keeps only associations that survive conditioning on all other measured variables. Timing alone cannot tell a local cause from a shared systemic driver. Conditioning alone cannot give direction. Together, and checked against one targeted intervention, they can decide between the "driver" and "downstream" accounts.
Core Claim
The specific claim is testable. If ECM remodeling is downstream of monoamine signaling, three things should hold. The regional collagen change should lag the monoaminergic gene change. After conditioning on the monoamine and glucocorticoid nodes, collagen should have no direct edge to behavior. And silencing the upstream dopamine or serotonin change during stress should abolish the collagen change.
If instead collagen is an active participant, three different things should hold. A direct collagen–behavior edge should survive conditioning, most importantly during deprivation, when behavior rebounds. Collagen should remain altered after monoamine genes return toward baseline, as a "scar." And the intervention should leave part of the collagen change intact.
A third outcome is also informative. If cross-region collagen co-variation disappears once plasma corticosterone is conditioned on, the brain-wide ECM pattern is a systemic hormonal effect rather than a circuit effect.
What Is Actually Different Here
Prior work supplies the phenomenon, the methods, and the constraints, but none of it combines them.
The stress papers motivate the question. Smagin et al. profiled 49 collagen genes in five regions using only univariate tests, principal components and clustering, all at day 20. Redina et al. added a 14-day deprivation state and found 1,111 of 2,629 nucleus accumbens changes persisting. However, they compared three conditions pairwise and ignored the ECM. Chen et al. built WGCNA co-expression modules, which rest on marginal correlations and keep indirect links, using a single cross-sectional endpoint.
The method papers supply the machinery but have never been applied to this kind of data. The Time-Varying Graphical Lasso (Hallac et al.) estimates one sparse precision matrix (inverse covariance) per time step. It shares strength across time with a penalty on how much the network changes between steps, but it was tested on finance and vehicle sensors. The Weighted Fused Pathway Graphical Lasso (Xiao et al.) fuses networks across unordered disease states and restricts edges to curated pathways.
One design fact matters more than any method choice: brain tissue can be collected only once per animal. There is no within-animal expression time series. This makes lglasso, which models repeated measurements on the same subject, and lagged autoregressive models such as TV-DBN, mismatched to the expression data. The only truly longitudinal within-animal signal is the daily behavioral record, and we use it as a covariate trajectory rather than as a network node sampled over time.
Research Program
Animals and sampling. Male C57BL/6J mice go through the standard 20-day agonistic paradigm as winners or losers, with time-matched controls. Animals are sampled on days 3, 10 and 20 of conflict. Winners are also sampled at 7 and 14 days of fighting deprivation, and losers after 7 and 14 days of recovery; untouched controls serve as day 0. Each group-by-time cell has eight animals, which is still modest.
Tissue and measurements. From each animal we dissect the five regions of the Smagin study: hypothalamus, VTA, hippocampus, striatum and midbrain raphe. We also collect terminal plasma for corticosterone, following the plasma-marker work of the CSDS plasma study. Expression is measured with a targeted qPCR panel of about 120 genes covering four groups:
- the 49 collagen genes;
- matrix-degrading and matrix-binding genes;
- monoaminergic genes (Th, Dat1, Drd1, Drd2, Ppp1r1b, Tph2, Slc6a4);
- glucocorticoid-response and cytokine genes.
A bulk RNA-seq subset at days 10 and 20 checks panel scores against genome-wide profiles. Daily attack or defeat counts are recorded for every animal.
Keeping the network small. With about eight animals per cell, a gene-level network is statistically hopeless. We therefore infer a network of about 18 nodes. Each region contributes three nodes: an ECM score, a monoamine score and a stress/immune score. The remaining nodes are plasma corticosterone and two behavioral summaries: cumulative agonistic intensity and the most recent three days of behavior.
Region scores are computed on variance-stabilized counts. This responds to the count-likelihood warning in GPcounts, and aggregating several genes into one score makes the Gaussian approximation much less strained. A nonparanormal rank transform serves as a robustness check.
The estimator. We fit sparse precision matrices jointly over the group-by-time grid. One fusion penalty links adjacent time points within a group; we compare a smooth penalty with TVGL's perturbed-node penalty, which lets a single hub rewire. A second fusion penalty links winners, losers and controls at the same time point. Penalty strengths are chosen by the extended BIC, with stability selection over animal subsamples.
We deliberately do not restrict edges to curated pathways. geneRNIB reports that database priors omit context-specific links, and brain collagen regulation under social stress is poorly annotated.
Timing. In parallel, each node's group-level trajectory is fit with a Gaussian-process model to estimate onset time, the first day it departs from controls, with bootstrap intervals.
Intervention check. The network should make a concrete prediction, for example that striatal ECM depends directly on the VTA monoamine node in losers. We test that prediction by chemogenetically silencing VTA dopamine neurons with DREADDs (designer receptors activated only by a designer drug) during days 1–10 of defeat. We then measure whether the striatal collagen change is abolished.
Baselines
Naive reference: timing only. The simplest reference is per-gene differential expression at each time point against controls, plus onset times, with no multivariate modeling. This is the natural extension of the Smagin analysis. It tells us which changes come first but cannot separate a local cause from a shared hormonal driver. The proposal beats it only if the conditional analysis changes a conclusion that timing alone would reach. The clearest case would be a collagen–monoamine co-occurrence that becomes conditionally independent once corticosterone is included.
Strongest simple competitor: marginal correlation networks. We build WGCNA-style marginal correlation networks on the same 18 node scores, using the same data and the same stability-selection thresholds. This is the method of Chen et al.. geneRNIB found plain Pearson correlation competitive with far more complex network inference, so this is a serious competitor.
The only cross-domain number favoring the precision-based approach comes from fMRI connectivity. There, precision measures reached an identifiability index above 2.1 versus about 1.5 for correlation (Liu et al.). No reported result was found on Lacuna for either method on stress or ECM transcriptomic networks.
Strongest existing temporal method: TVGL as published. TVGL is fit separately for each group, with no group fusion. Hallac et al. report a 92% F1 improvement over static baselines on synthetic data. It is run with an identical tuning grid and identical subsamples.
Other temporal baselines. We also fit independent graphical lasso at each group-by-time cell, and one static network pooled over all time points.
Ablations. The first ablation isolates the key ingredient, conditioning. It compares the full conditioning set against a set without corticosterone and a set without behavior. The second removes each fusion penalty in turn.
What counts as winning. On simulations matched to our design, the joint model must recover edges more accurately, measured by F1 against the true graph. On real data it must have higher held-out likelihood on left-out animals and more stable edges under subsampling. Most importantly, its intervention prediction must be confirmed where the marginal network's competing prediction fails.
Why not just…?
Why not just run a dense time course with per-gene differential expression and read off which genes change first? This is the most serious alternative, and we include it as the reference. It will likely answer "does collagen lag dopamine genes?" on its own. What it cannot answer is whether two changes that co-occur share a cause. Chronic social stress raises glucocorticoids throughout the body, and a single hormonal driver could move hypothalamic, striatal and hippocampal collagen together while monoamine genes also shift. The decisive comparison is the corticosterone-conditioning ablation. If removing corticosterone from the conditioning set leaves every conclusion unchanged, the network adds little beyond timing, and we will report that.
Why not just use correlation modules like WGCNA? Marginal correlation keeps edges produced by common drivers, and Chen et al. is exactly that kind of analysis. Still, geneRNIB shows correlation can match sophisticated methods, so we do not assume it loses. Two tests settle it. The first is edge recovery in design-matched simulations with a planted common driver. The second is whether the correlation network or the partial correlation network correctly predicts the outcome of VTA silencing.
Why not just fit a separate network at each time point, or one pooled network? Eight animals per cell makes separate networks unstable, and pooling blurs the transition into deprivation. TV-DBN found static models detected 2 to 5 times fewer enriched modules than time-varying ones. The fusion ablation tests whether sharing across time and groups actually helps at our sample size, or merely smooths away real changes at the deprivation boundary.
Why not skip network inference and go straight to intervention? Interventions are the gold standard but costly. There are many possible upstream nodes and regions, and the network's job is to choose which single experiment to run. If the network's chosen intervention fails its prediction, that failure is itself evidence against the method.
Proposed Experiments and Decisive Checks
Simulation first. Before collecting animals, we simulate data matched to the design: 18 nodes, eight samples per cell, 18 cells. The simulations include a planted systemic driver and a planted single-hub rewiring at the deprivation boundary. They answer whether the joint estimator recovers edges materially better than the marginal and independent baselines at this sample size. If it does not, we increase animals per cell before any wet-lab work.
Replicating the known result. On real data, the first check is whether the day-20 region pattern from Smagin et al. replicates at larger n: hypothalamus and striatum up, VTA down.
The decisive analysis. We then examine three things jointly:
- the onset ordering of ECM versus monoamine nodes in each region;
- whether a direct ECM–behavior edge survives conditioning during deprivation, when Redina et al. observed behavioral rebound alongside persistent gene changes;
- whether cross-region ECM edges vanish once corticosterone is conditioned on.
The final test. The DREADD silencing experiment confronts the network's top directional prediction with a manipulation.
Reading the outcomes. Three outcome patterns are possible, and each is publishable. In the "downstream" pattern, collagen lags monoamine genes, has no direct edge to behavior, and its change is abolished by silencing. In the "participant" pattern, collagen has a persistent direct edge and its change partly survives silencing. In the "systemic" pattern, the cross-region collagen edges are explained by corticosterone. Any one of them resolves the open question that the single-endpoint studies left unanswered.
References
- Aberrant Expression of Collagen Gene Family in the Brain Regions of Male Mice with Behavioral Psychopathologies Induced by Chronic Agonistic Interactions
- Neurotransmitter Genes in the Nucleus Accumbens That Are Involved in the Development of a Behavioral Pathology After Positive Fighting Experiences and Their Deprivation: A Conceptual Paradigm for Data Analysis
- Identification and experimental validation of hub genes underlying depressive-like behaviors induced by chronic social defeat stress
- Network Inference via the Time-Varying Graphical Lasso
- Weighted Fused Pathway Graphical Lasso for Joint Estimation of Multiple Gene Networks
- lglasso: Longitudinal Graphical Lasso
- Time-Varying Dynamic Bayesian Networks
- Blood Plasma Markers in Depressed Mice under Chronic Social Defeat Stress
- Non-parametric modelling of temporal and spatial counts data from RNA-seq experiments
- geneRNIB: a living benchmark for gene regulatory network inference
- Benchmarking methods for mapping functional connectivity in the brain