Did the Network Rewire, or Did the Cell Mix Change? A Composition Null for Reconsolidation Networks

When a memory is recalled, the cells that hold it can briefly become changeable before they settle again. Experiments show this happens only in the specific pathway that was reactivated, not in its neighbours. A natural way to look for this change in molecular data is to compare gene-interaction networks between animals that recalled the memory and animals that did not. Those networks are usually estimated with partial correlations: the correlation between two genes after removing what every other measured gene explains. The plan has a statistical trap. If recall merely increases the fraction of cells sitting in an activity-induced state, the pooled partial-correlation network changes even when no cell's internal wiring changed at all. Hidden mixtures of cell types act as unmeasured confounders, so averaging over them produces edges that differ between conditions and look like rewiring.

This note proposes a test and an estimator that separate these two explanations. The central object is a "composition-only null": the hypothesis that two conditions share the same per-state conditional networks and differ only in how many cells occupy each state. Rejecting this null is evidence that some state's dependence structure was modified. That is the cellular counterpart of the distinction the latent-cause theory of reconsolidation draws between updating an existing memory and creating a new one. We build the test, characterize when simpler fixes are enough, and validate it on semi-synthetic single-cell data where the ground truth is known.

Core Claim

Consider data that are a mixture of cell states with weights $\pi_k$, means $\mu_k$ and covariances $\Sigma_k$. The pooled covariance is $\Sigma = \sum_k \pi_k \Sigma_k + \sum_k \pi_k (\mu_k - \bar\mu)(\mu_k - \bar\mu)^\top$, where $\bar\mu$ is the pooled mean. For two states this reduces to

$$\Sigma = \pi \Sigma_1 + (1-\pi)\Sigma_2 + \pi(1-\pi)\,\delta\delta^\top, \qquad \delta = \mu_1 - \mu_2 .$$

The pooled partial correlations come from $\Sigma^{-1}$. Through the low-rank term, they depend on $\pi$ even when $\Sigma_1$ and $\Sigma_2$ are fixed. A pathway-specific recall event raises $\pi$, the fraction of reactivated cells. It therefore creates differential edges among exactly the genes that define the activated state, which are the genes a biologist would most want to interpret.

Our claim has two parts. First, standard pooled differential partial-correlation analysis is systematically miscalibrated in this regime. Second, a mixture of Gaussian graphical models can test whether the composition shift explains the data. The test fits component precision matrices shared across conditions under the null, lets only one component's precision differ under the alternative, and compares the two by likelihood ratio. A Gaussian graphical model is a model in which zeros of the inverse covariance mark pairs that are independent given everything else.

The test answers a well-posed question that the existing tools in the file do not ask. Pooled shrinkage estimators such as DRAGON give calibrated partial-correlation edges for bulk tumours but assume a single population. Cell-level network methods such as scGIR use pairwise, expression-weighted correlations and do not condition on other genes.

Why Reconsolidation Is the Right Regime

The reconsolidation literature supplies the structure that makes this question sharp, not just a place to apply it. Song et al. showed that reactivating a fear memory makes only the reactivated amygdala synapses labile, while neighbouring traces in the same structure stay intact. Any tissue sample therefore mixes a small changed population with a large unchanged one. Kastner et al.'s synaptic model adds that lability requires active stimulation, which unbinds molecular stabilizers; idle synapses do not respond to protein-synthesis inhibitors. This predicts a within-state change confined to stimulated cells, not a global one. Liu et al. found that a DNA methyltransferase inhibitor in insular cortex or basolateral amygdala impairs a morphine-withdrawal memory only when paired with reactivation. That points to a methylation-to-expression dependence that should change only in the reactivated subpopulation.

The conceptual link comes from Gershman, Monfils et al.. Their model explains boundary conditions on reconsolidation by an expectation-maximization procedure that decides whether a retrieval belongs to an old latent cause, which is then updated, or to a new one, which yields a separate extinction-like trace. Their model operates on scalar associative weights and treats the brain as one agent. Our composition null asks the same question of cell populations. A shift in state occupancy with unchanged per-state networks resembles "a new cause was added or reweighted." A change inside one state's precision matrix resembles "the existing cause was modified." Thome et al. report that clinical reconsolidation protocols often fail to replicate because small procedural changes push the system from updating to new-trace formation. A molecular readout that distinguishes the two outcomes would therefore be useful, and a pooled analysis cannot provide one.

Research Program

We will build the estimator, the test, and a testbed with known answers.

The estimator is an expectation-maximization mixture of Gaussian graphical models fit jointly to cells from both conditions. Each cell has a soft assignment to $K$ states. The E-step computes these assignments. The M-step re-estimates each state's precision matrix by shrinkage. We use per-block shrinkage in the style of DRAGON so that gene modules with different noise floors are not penalized identically. Under the null, all $\Theta_k$ are shared across conditions and only the weights $\pi_k^{(c)}$ vary with condition $c$. Under the alternative, one designated state, the candidate reactivated state, gets a condition-specific $\Theta_k^{(c)}$. The test statistic is the likelihood ratio. We calibrate it by parametric bootstrap from the fitted null, because shrinkage and soft assignments make asymptotic chi-square calibration unreliable. Edges reported as rewired are those whose condition contrast in the alternative state survives a bootstrap threshold.

Two practical constraints shape the design. Single-cell counts are discrete and zero-inflated. Raw inverse covariance on them produces artifacts, and the file notes that continuous conditional-independence tests break under discretization (sample-efficient CI test under discretization). We will therefore run everything on rank-based Gaussian-copula transforms and report sensitivity to that choice. The second constraint is that the reactivated state may be rare and may overlap heavily with resting cells. Atlas-based methods such as POPPY cannot represent a transient state that is absent from their reference. We therefore allow an optional weak anchor: a prior that cells with high scores on a user-supplied activity module are more likely to belong to the candidate state. We report results with and without it.

The testbed is semi-synthetic. We take excitatory-neuron cells from the Tasic mouse visual cortex scRNA-seq data (used in scGIR) and fit a background mixture to them. We then mark a fraction of cells as "reactivated" and inject one of three ground truths into the second condition. Under composition-only, the reactivated fraction rises from $\pi_0$ to $\pi_1$ while per-state networks are untouched. Under rewiring-only, $\pi$ is fixed and a block of the reactivated state's precision is altered, mimicking stabilizer unbinding in stimulated cells only. In the combined case, both happen. We resample counts with a Poisson observation layer so that dropout and discreteness are realistic. As a second, spatial testbed, we use MERFISH mouse brain spots deconvolved into single-cell profiles with Count Bridges. This checks whether deconvolution error itself creates false rewiring calls.

Baselines

The naive reference point is pooled differential Pearson co-expression: correlation matrices computed separately per condition and edges ranked by the change. This is what a WGCNA-style analysis would report. We expect it to flag both indirect and composition-driven edges. It sets the floor that every other method must clear.

The strongest existing pooled method is DRAGON-style shrinkage partial correlation, fit separately to each condition, with edges compared by their analytic p-values (DRAGON). It is the strongest pooled competitor because it handles unequal noise across feature blocks and is sample-efficient. The paper reports an edge-recovery AUC above 0.8 with fewer samples than GeneNet or the graphical lasso, and an AUC gain of about 0.14 over standard single-penalty models in unbalanced simulations. Those numbers are for static, single-population network recovery. We found no reported result on Lacuna for any method on differential-network attribution under a mixture with a shifting composition, so we cannot quote a target number for this task. On edge recovery within a single population, our estimator should match DRAGON's reported AUC above 0.8 when $K=1$; otherwise the shrinkage step is mis-implemented. Beating it means a lower false rewiring-call rate in the composition-only scenario and higher AUROC for true rewired edges in the combined scenario.

The strongest heterogeneity-aware competitor is cluster-then-estimate. We cluster cells on normalized expression, which scGIR reports gives an ARI of 0.64 on Chu-time versus 0.68 for its gene-importance representation. We then fit DRAGON-style shrinkage within each cluster and condition and compare. We run it with both expression-based and scGIR-based clusters, using the same transformed data, the same shrinkage routine, and the same bootstrap budget for edge thresholds as our method.

A further baseline is covariate adjustment: pooled partial correlation with an activity-module score or a cluster indicator added as an extra conditioning variable.

Two ablations isolate our key ingredient, the sharing of per-state networks across conditions under the null. One fits independent mixtures per condition and compares matched components; this removes the sharing. The other gives our method the true state labels; this oracle separates assignment error from the test itself.

All methods see the same cells. Each gets the same number of hyperparameter settings on a held-out simulation seed and the same compute.

Why not just…?

Why not just cluster the cells first and run a graphical model inside each cluster? This is the real threat. If the reactivated state forms a clean, well-populated cluster, per-cluster networks are already free of the composition term, and our method adds little. The case for a joint mixture is the regime the biology suggests: a small reactivated fraction that grades into resting cells. There, hard clustering mislabels boundary cells. Because the boundary moves when $\pi$ changes, the mislabelling differs between conditions, which reintroduces a composition artifact inside "pure" clusters. We will not argue this point; we will measure it. The sweep over separation strength and $\pi$ will locate the crossover. If cluster-then-estimate is as good across the realistic range, we will report that, and the contribution shrinks to the calibrated composition-only test, which cluster-then-estimate does not provide.

Why not just add the activity score or cluster label as a covariate? Conditioning on the state indicator removes the rank-one mean-difference term in the pooled covariance. It still assumes all states share one within-state covariance, so a genuine rewiring confined to reactivated cells gets averaged with the unchanged majority and diluted by roughly the factor $\pi$. Covariate adjustment should therefore be well calibrated under composition-only but low-powered under rewiring-only. The rewiring-only scenario tests exactly this.

Why not just use the best pooled estimator, since partial correlations already remove indirect effects? Partial correlations remove indirect effects only through observed variables. The large-scale fMRI benchmark of Liu et al. shows that precision-based measures beat Pearson correlation for fingerprinting: an identifiability index above 2.1 versus about 1.5. But that is a single-population, resting-state setting. Hidden state membership is an unobserved variable, and the displayed identity shows pooled precision moving with $\pi$ alone. The composition-only scenario directly measures how often pooled DRAGON calls rewiring that never happened.

Proposed Experiments and Decisive Checks

The first decisive check is calibration under composition-only. Across a grid of baseline fractions $\pi_0$, condition shifts $\pi_1 - \pi_0$, and state separations $\|\delta\|$, we measure the type I error of the composition-only test at nominal 0.05 and the number of edges each baseline falsely calls rewired. We expect pooled methods' false-call counts to grow with $\pi(1-\pi)\|\delta\|^2$, the size of the rank-one term. Our test should stay near nominal. If it does not, the null is misspecified and the project stops there.

The second check is power and attribution in the rewiring-only and combined scenarios. We measure the AUROC for recovering truly rewired edges among all candidate edges, separately for edges inside and outside the activity module. This is where we expect separation from covariate adjustment, which dilutes rewiring, and from pooled methods, which confuse rewiring with composition.

The third check is the oracle-label ablation and the independent-mixtures ablation. These show how much of any gain comes from sharing networks across conditions rather than from better clustering.

The fourth check is robustness: rank-copula versus log-normalized inputs, Poisson versus overdispersed resampling, and the Count Bridges–deconvolved MERFISH testbed, where deconvolution error is an extra source of spurious differences.

Success means three things together: calibrated type I error, a clear reduction in false rewiring calls relative to pooled DRAGON, and a reported region of separation strength and fraction in which the method beats cluster-then-estimate. A natural follow-up, outside this note's scope, is applying the test to paired methylation and expression data from reactivated versus non-reactivated tissue. Mirroring the Liu et al. intervention, the question would be whether the methylation-to-expression partial correlations, which DRAGON found to be negative 92% of the time in bulk breast tumours, change only within the reactivated state.