Deconvolving Covariance, Not Samples: Cell-Type-Resolved Time-Varying Gene Networks from Bulk Cohorts

Most long-running biological cohorts, such as aging mice, diabetic donors, or post-infarction patients, are profiled with bulk RNA sequencing. Bulk sequencing reports one expression vector per tissue sample, averaged over every cell in it. When researchers ask how gene regulation changes over time, they usually compute gene–gene correlations across these samples. Those correlations are contaminated in a specific way. If neurons are lost with age, every pair of neuron-marker genes becomes correlated across samples simply because samples differ in how many neurons they contain. That signal says nothing about how neurons regulate their genes. A network that "rewires" between ages may therefore only be recording a change in cell composition.

We propose to estimate one gene network per cell type and per time point directly from the second moments of the bulk data. We never try to reconstruct each sample's cell-type-specific expression, which is the ill-posed step in current pipelines. The core observation is simple. If a bulk sample is a proportion-weighted sum of independent cell-type contributions, then the covariance of bulk residuals across samples with known proportions is a proportion-squared-weighted sum of the cell types' own covariances. This splits into one tiny least-squares problem per gene pair, and every pair can be solved in parallel. Once the per-cell-type covariances are recovered at each time point, sparse networks come from node-wise neighborhood selection, again parallel per gene. Changes over time are detected with an exact change-point penalty, also per gene pair. The result is a method that scales to genome-wide networks, separates composition drift from real rewiring, and says which cell type hosts each rewired edge.

Core Claim

Let $y_i \in \mathbb{R}^p$ be the log-normalized bulk profile of sample $i$ at time $t_i$. Let $\pi_{ic}$ be its estimated proportion of cell type $c$ out of $C$ types, and let $r_i$ be its residual after regressing out a proportion-weighted mean model (plus covariates such as genotype and sex). Assume the cell-type contributions deviate from their means independently across types, with covariance $\Sigma^{c}(t)$. Then

$$ \mathbb{E}\left[r_{ij}\, r_{ik} \mid \pi_i, t_i = t\right] \;=\; \sum_{c=1}^{C} \pi_{ic}^{2}\, \Sigma^{c}_{jk}(t), $$

where $j,k$ index genes. For each pair $(j,k)$, the vector of cell-type covariances $\Sigma^{c}_{jk}(t)$ is the coefficient vector of a regression of residual products on squared proportions across the samples at time $t$. Each such regression has only $C$ unknowns per time point.

Two hypotheses give different predictions. Under pure composition drift, $\Sigma^{c}(t)$ stays constant while the bulk covariance changes through $\pi(t)$. Under rewiring, some $\Sigma^{c}_{jk}(t)$ jumps. We identify such jumps by solving, for each pair, an $\ell_0$ problem that penalizes the number of changes across ordered time points, rather than smoothing with a kernel.

The new element is the combination: deconvolving second moments across samples, decomposing entry by entry across gene pairs, and fitting change-points per pair. The claim is that this recovers cell-type-attributed dynamic edges more accurately than deconvolving samples first and inferring networks second. We also expect it to stay tractable at $p \approx 10^4$.

Relation to Prior Work

The motivating evidence comes from studies that hit the composition confound without resolving it. The aging-cortex study Deconvolution reveals cell-type-specific transcriptomic changes in the aging mouse brain ran CIBERSORTx on 141 bulk cortices at 3, 12 and 24 months. It then built WGCNA correlation modules on the inferred cell-type profiles. Its networks inherit whatever errors deconvolution makes, and they are marginal correlations rather than conditional-independence graphs. We use the same cohort, but estimate cell-type covariances without per-sample reconstruction and follow with conditional (neighborhood-based) graphs.

SCDC gives accurate proportions by weighting multiple single-cell references. It deconvolves each sample independently and infers no networks. We reuse it as the source of $\pi_{ic}$.

Symphony is the closest joint model in the file. It infers cell-type-specific regulatory networks while deconvolving bulk chromatin accessibility. It is static, fits all cells and genes together by Bayesian inference, and reports degradation below 5% abundance. We keep its lesson that deconvolution and network inference should constrain each other. We give up its generative richness in exchange for a moment-based, per-pair decomposition that adds time.

The method components come from decoupled structure learning. Three works shape the design:

  • FASJEM shows that multi-condition graphical models can be split into entry-wise closed-form subproblems costing $O(Kp^2)$, where $K$ is the number of conditions. It requires pure, pre-sorted samples for each condition. Our first stage plays the same role but starts from mixtures.
  • Scalable Inference of Sparsely-changing Gaussian Markov Random Fields solves an exact $\ell_0$ change penalty per edge by shortest path on a small auxiliary graph. It reports under an hour on 500 million variables and structure learning with one sample per period. We apply its per-edge change-point machinery to the deconvolved covariance sequences.
  • On Time Varying Undirected Graphs and TV-DBN supply the kernel-smoothed node-wise alternative. The first proves that kernel smoothing slows convergence to $n^{-1/3}$ and assumes smooth change. Both assume one homogeneous population.

TV-DBN's directed lag model also needs the same individual measured repeatedly. Most bulk cohorts, including the aging and islet cohorts, are cross-sectional: different animals or donors at each time. We therefore estimate undirected graphs that vary over time and state plainly that we do not infer temporal direction.

One result constrains the design. geneRNIB found that simple decoupled methods, such as correlation and GRNBoost2, beat foundation models and multiomic pipelines on perturbation-based evaluation. This argues for keeping the estimator simple and moment-based. It also means correlation-based pipelines are serious competitors, not strawmen.

Research Program

The estimator has three decoupled stages. The first stage fits the mean model. Proportions come from SCDC, and each gene's mean is proportion-weighted with covariates, fit gene by gene. For every gene pair, the first stage then solves a weighted least-squares regression of residual products on the squared proportions at each time point, or at each kernel window along a continuous axis such as HbA1c. It uses heteroskedasticity-aware weights, because products of residuals have heavy tails.

The second stage takes each pair's sequence of $C \times T$ covariance estimates, with $T$ the number of time points. It fits a piecewise-constant path under an $\ell_0$ penalty on changes using the shortest-path method. This yields a denoised $\Sigma^{c}(t)$ and a set of change events per cell type. We project each $\Sigma^{c}(t)$ to be positive semidefinite before the third stage.

The third stage runs graphical-lasso-style neighborhood selection per gene on each $\Sigma^{c}(t)$. It chooses sparsity by StARS stability selection, as implemented in XMRF. Every stage is embarrassingly parallel across genes or gene pairs.

The analysis half of the program concerns identifiability. Cell-type covariances are recoverable only if the squared proportions vary enough across samples within a time point. We will derive the variance of the per-pair estimator as a function of the empirical design matrix of $\pi^2$. This makes the rare-type limit explicit. A type at 5% abundance enters with weight 0.0025, so we expect a detection floor at least as harsh as the 5% limit reported by Symphony. We will report that floor rather than hide it.

We will also test the independence-across-types assumption. Shared individual-level factors can induce cross-type covariance. We will add cross terms $\pi_{ic}\pi_{id}$ for a few selected pairs of types $(c,d)$ and check whether the fit changes. Finally, counts are not Gaussian, as Count Bridges and XMRF stress. We work on variance-stabilized data and treat a count-native version as out of scope.

Baselines

The naive reference applies kernel-smoothed node-wise neighborhood selection from On Time Varying Undirected Graphs directly to bulk profiles. It ignores composition and produces one network per time point. Every method must beat it on attribution by construction. The more telling comparison is on the composition-only simulations: there it should report many spurious "rewired" edges, and a correct method should report almost none.

The strongest existing competitor is the sequential pipeline used in the aging-cortex study. It deconvolves each sample into cell-type expression, then infers networks per type. We will run it in its strongest form: SCDC or CIBERSORTx for per-sample cell-type expression, followed by the same neighborhood selection, the same $\ell_0$ temporal fusion and the same StARS tuning as our method. Only the deconvolution step then differs. Both methods get identical proportions, gene sets, compute budget and hyperparameter search.

No number for cell-type-resolved dynamic edge recovery from bulk is reported. We searched Lacuna for such a result on this task and found none. The closest reported figures concern the component steps:

  • SCDC reports Pearson $r = 0.991$ for proportions on the 6:3:1 three-cell-line mixture.
  • FASJEM reports AUC 0.9332 on synthetic multi-condition networks built from pure samples.
  • TV-DBN reports its dynamic-structure F1 only graphically.

Beating the pipeline means higher area under the precision–recall curve (AUPRC) for cell-type-attributed edges and higher change-point F1 on semi-synthetic data at matched sample sizes. It also means better held-out agreement on real data, as described under the experiments.

Symphony serves as a static joint-model reference. It will be run on a gene subset small enough for its inference to finish. The goal is to check whether our moment estimator loses much accuracy relative to full joint modeling at one time point.

Two ablations isolate the key ingredients. The first replaces the $\pi^2$ regressors with $\pi$, which a naive linear analogy would suggest. This tests whether the correct second-moment structure matters. The second swaps the $\ell_0$ change path for kernel smoothing, which tests the abrupt-transition claim.

Why not just…?

Why not just deconvolve each sample and run a time-varying network method on each cell type? This is the real threat. With abundant cell types and accurate references, it may match us. Its weakness is that per-sample cell-type expression is underdetermined. Tools like CIBERSORTx tend to return profiles pulled toward the reference, and those shrink or distort exactly the between-sample covariance we care about. The aging study offers no way to measure this, because it validated marker genes rather than edges. The matched-pipeline baseline settles the question directly. If it ties us on the semi-synthetic benchmark across abundance levels, our contribution reduces to scalability and we will say so.

Why not just regress cell proportions out of the bulk data and build networks on the residuals? This removes composition-driven correlation in the mean. On the question "is there rewiring beyond composition?" it may perform nearly as well as our method. It cannot say which cell type an edge or change belongs to. Its residual covariance also still mixes $\sum_c \pi_c^2 \Sigma^c$ across types, with weights that shift as composition drifts. It therefore should still report spurious changes when one type's weight grows. We include this as an explicit baseline and report pooled change detection separately from attribution. That keeps the comparison honest about where our advantage lies.

Why not just bin samples by time point and run FASJEM on deconvolved profiles? Binning is what our per-time regressions already do for cross-sectional cohorts, so binning itself is not the objection. The objection is that FASJEM needs pure per-condition data. Its fused penalties are also convex surrogates that shrink true jumps. Running FASJEM on pipeline-deconvolved profiles is a variant of the first baseline, and we will include it.

Why not just profile single cells over time? Where affordable, this is better. But the cohorts at issue are already bulk-only: 141 aging mice, 77 islet donors, and public infarction blood series. SCDC also documents that droplet platforms lose fragile cell types that bulk data retains. Single-cell data enters our study as references and as validation, not as a replacement.

Proposed Experiments

Semi-synthetic benchmark (the decisive test). We will simulate cell-type-specific networks for three to five types with the BEELINE-style simulator used in the PMF-GRN work. Some edges change abruptly at known times and others stay fixed. We will mix the cell-level profiles into pseudo-bulks, with proportions that drift over time. The designs will match the real cohorts: three time points with roughly 47 samples each, as in the aging study, and a continuous axis with 77 samples, as in the islet cohort. We will vary the minimum abundance from 30% down to 2%. We will also include a pure-composition-drift condition with no rewiring at all. Here the correct answer is zero change events, and the naive bulk method should fail visibly. Metrics are per-type edge AUPRC, change-point F1 and attribution accuracy. The decisive check is a curve of attribution accuracy against abundance, for us and for the matched pipeline.

Real perturbation-based check. We will build pseudo-bulk mixtures from the multi-cell-type OPSCA perturbation data in geneRNIB. We will score the inferred per-type networks with geneRNIB's perturbation metrics and compare against the same metrics computed on the single-cell data itself. This tells us how much accuracy is lost by seeing only bulk.

Longitudinal application. On the aging cortex cohort, we test whether the immune-module changes reported by the aging study survive as within-type rewiring once composition drift is modeled. Predictions are checked against independent single-cell cortex datasets, used the same way that study used them. On the islet cohort, we use HbA1c as a continuous axis with SCDC proportions and ask whether beta-cell networks change or only beta-cell abundance does.

Scalability. We time the full estimator at $p$ = 2,000, 10,000 and 20,000 genes on CPU and GPU, alongside FASJEM's reported ability to handle $p = 10{,}000$.