Do Deconvolution Methods Erase the Effects of Chronic Stress?

Most of what we know about gene expression in rodent models of depression comes from bulk RNA sequencing: a piece of brain is ground up and the RNA of hundreds of thousands of cells is read out as one averaged profile. Computational deconvolution tries to undo that averaging. It uses a reference atlas of single-cell profiles to estimate, for each sample, how much of each cell type is present and what each cell type was expressing. These tools are now good enough to find cell-type-specific signals in large human brain cohorts. They have not been tested on the case where they are most likely to fail. Chronic stress does not mainly kill cells. It changes what surviving cells express, and it can push astrocytes, microglia or neurons into states that a healthy reference atlas never contained. Methods that pull estimates toward the reference may then treat the stress effect as noise and shrink it away.

We propose to measure this failure directly. Bulk and single-nucleus data will be collected from the same stressed and control mice, so the true cell-type-specific stress effect is known. We will then report how much each method attenuates it, and test a simple correction that frees each cell type to depart from the reference along a direction shared by stressed animals. In two sentences: reference-guided unmixing should shrink stress-induced cell-state changes toward zero, and the shrinkage should grow with how far the stressed state lies from the healthy atlas. We will quantify this attenuation against paired single-nucleus ground truth in chronic social defeat, and show whether a calibrated, shift-aware unmixing step recovers known astrocyte-specific effects before anyone uses these tools to re-mine the existing bulk stress archives.

Core Claim

Two stress paradigms dominate the field. In chronic unpredictable mild stress (CUMS), rodents get a randomized sequence of mild stressors over weeks so they cannot habituate. In chronic social defeat (CSDS), a mouse faces a new dominant aggressor every day. These paradigms produce changes that look like changes of cell state rather than cell number. Examples include NLRP3 inflammasome activation with abnormal astrocyte activation (Yu Fang, Haoran Wu et al.), mitochondrial fragmentation in prefrontal neurons through SENP3/FIS1 (Shaoda Lai et al.), and persistent astrocyte-only upregulation of the transcription factor Zbtb7a in orbitofrontal cortex after CSDS (Sasha L. Fulton et al.).

The current unmixing tools differ in how hard they pull toward the reference. EPIC-unmix uses a single-cell prior updated by the bulk data (Chenwei Tang et al.). scProjection projects mixtures onto per-cell-type variational autoencoder manifolds learned from an atlas (Projecting RNA measurements...); its authors warn that off-manifold disease states will be pulled back toward the healthy reference. GTM-decon uses Dirichlet priors from the reference (Lakshmipuram S. Swapna et al.).

Our claim is quantitative and falsifiable. For a gene $g$ in cell type $k$, let $\beta_{gk}$ be the true stress log-fold-change measured by single-nucleus sequencing and $\hat\beta_{gk}$ the estimate from unmixing bulk data. The attenuation ratio $\hat\beta_{gk}/\beta_{gk}$ should fall below one. It should fall further as the stressed cells' distance from the reference manifold increases. Without stress-matched references, the strongest-prior methods should attenuate most. If instead attenuation is near one across methods, the field can use existing tools on stress archives as they are. That result would also be worth publishing.

Research Program

The work has three parts.

The first part is a paired ground-truth cohort. Male and female C57BL/6J mice undergo CSDS and are classified as susceptible or resilient, alongside controls. A second arm undergoes CUMS. For each animal, one hemisphere's orbitofrontal cortex goes to standard bulk RNA-seq. The other hemisphere goes to single-nucleus RNA-seq (snRNA-seq) at moderate depth. Pseudobulk per cell type per animal then gives a measured $\beta_{gk}$ in the same individuals. We choose orbitofrontal cortex because Sasha L. Fulton et al. supplies an independently validated, astrocyte-restricted positive control in Zbtb7a. That study separated nuclei only into neuronal and non-neuronal pools, so our snRNA-seq also tests whether the effect is confined to a particular astrocyte sub-state.

The second part is a controlled in-silico stress series. Starting from control snRNA-seq and the Allen Brain Map mouse references used by scProjection, we build synthetic bulk mixtures. In these mixtures one cell type's profile is shifted along the empirically measured stress direction, scaled from zero to beyond the observed magnitude. Cell proportions are varied independently. This lets us plot attenuation against distance from the reference manifold. It also lets us check whether each method separates a change in astrocyte state from a change in astrocyte proportion. Bulk analysis cannot make that separation, as Ying Sun et al. acknowledge for parvalbumin interneurons in mPFC.

The third part is the shift-aware correction. We model bulk expression of gene $g$ in animal $i$ as

$$ y_{gi} = \sum_k p_{ik}\,\big(\mu_{gk} + s_i\,\delta_{gk}\big) + \varepsilon_{gi}, $$

where $p_{ik}$ is the proportion of cell type $k$, $\mu_{gk}$ is the reference expression, $s_i$ is the stress indicator (or a continuous susceptibility score), $\delta_{gk}$ is a cell-type-specific stress shift, and $\varepsilon_{gi}$ is count noise. The stress shift $\delta_{gk}$ is shrunk toward zero with a strength chosen on a small held-out set of paired animals, rather than fixed by the atlas.

The model itself is close to existing interaction-based approaches such as TCA and bMIND. The new parts are the calibration of shrinkage against paired ground truth and a residual test. That test flags genes whose bulk residuals in stressed animals cannot be explained by any reference cell type, which marks off-reference states instead of silently absorbing them. We then apply the calibrated pipeline to published bulk stress datasets, including hypothalamic chronic social conflict cohorts (Anna G. Galyamina et al., И. Л. Коваленко et al.). The goal is to ask which cell classes carry the reported Th/Gad2 and cytokine hub changes. This mirrors how Yingxue Ren et al. found aging-related mitochondrial and immune genes concentrated in unexpected cell types.

Baselines

The naive reference point is bulk differential expression with no unmixing, with stress effects assigned to whichever cell type expresses the gene most highly in the atlas. It sets the floor and shows how often bulk analysis puts the effect in the wrong cell type.

The second baseline is CIBERSORTx proportions followed by bMIND cell-type expression. This is the pipeline used on the 141-mouse aging cortex cohort, which reported about 90% validation accuracy against independent scRNA-seq (Yingxue Ren et al.). It was also used in psychiatric blood, where concordance with reference single-cell expression ranged from R² 0.61 to 0.84 (Cell-type deconvolution of bulk-blood RNA-seq...). Both of those settings involve a different tissue or condition from ours.

The strongest existing competitor is EPIC-unmix, which reported 4.6% to 109.8% higher accuracy than TCA and bMIND in inferring cell-type-specific expression on ROSMAP and MSBB, with 83% to 100% directional agreement with scRNA-seq differential genes (Chenwei Tang et al.). We also run scProjection, which reported 27% higher projection accuracy than Tangram on Allen Brain Map synthetic mixtures (Projecting RNA measurements...), and GTM-decon. For any stress paradigm we found no reported result on Lacuna for these methods.

For fairness, all methods receive the same reference atlas, the same bulk matrices, the same gene filtering, and the same hyperparameter search budget. Only the calibrated shift-aware variant uses the held-out paired animals, and those animals are excluded from all evaluation.

Two ablations isolate our ingredients. The first is the shift-aware model with atlas-fixed shrinkage, which separates the value of the explicit stress term from the value of calibration. The second gives every method a stress-matched reference that includes the stressed snRNA-seq; this is an upper bound. Beating EPIC-unmix means two things on held-out animals: a median attenuation ratio closer to one for the cell types and genes that snRNA-seq calls stress-responsive, and higher rank correlation between $\hat\beta_{gk}$ and $\beta_{gk}$. Calibration must not raise false cell-type assignments on null genes, so the comparison is reported at matched false-discovery rate.

Why not just…?

Why not just run snRNA-seq on every stressed mouse? For a single, well-funded experiment, that is the right answer, and our paired cohort does exactly that. The problem is scale and history. Designs that cross sex, susceptibility, paradigm, region and time point quickly exceed what single-cell budgets allow. Chenwei Tang et al. found that 57.4% of cell-type-specific differential genes recovered from large bulk cohorts were missed by smaller single-cell studies. And no amount of new sequencing revisits the existing bulk archives. The proposal treats snRNA-seq as the ruler for deciding when bulk unmixing can be trusted, not as a competitor to it.

Why not just add stressed cells to the reference atlas? This is the most serious threat. If a stress-matched reference removes the attenuation, much of our correction is unnecessary wherever such references exist, so we run it as an explicit upper-bound ablation. Its limits are practical. Stress-matched atlases do not exist for most regions, paradigms, sexes and time points in the historical datasets, and building one requires the very single-cell experiment we are trying to avoid repeating. The decisive measurement is how much of the gap between atlas-only and stress-matched references our calibrated variant closes using only a handful of paired animals. If it closes little, we will report that matched references are required and narrow the claim accordingly.

Why not just use FANS sorting, or simple marker-based linear deconvolution? FANS in the closest prior study separated only neuronal from non-neuronal nuclei (Sasha L. Fulton et al.), and it cannot be applied retrospectively to existing data. Marker-based linear models estimate proportions, not per-animal expression. They also rely on marker genes that can themselves change under disease (Lakshmipuram S. Swapna et al.). Their weakness may make them useful here, though: having little prior pull, they may attenuate less. So we include a marker-based proportion estimate combined with per-cell-type interaction regression as a baseline. If it matches the reference-prior methods on attenuation, that is itself evidence for the core claim.

Proposed Experiments

The first decisive check is the Zbtb7a positive control. In the paired CSDS orbitofrontal cortex cohort, each method's unmixed estimates are scored on whether the astrocyte-specific upregulation appears, how large it is relative to the snRNA-seq value, and whether it leaks into other cell types. If EPIC-unmix or scProjection returns an attenuation ratio near one here, the central concern is weakened for this effect.

The second check is the dose-response curve from the in-silico series. Attenuation is plotted against distance of the shifted cells from the reference manifold. That distance is measured as the reconstruction error of scProjection's per-cell-type autoencoder on the shifted cells. We predict a monotone decline in recovered effect for the prior-heavy methods and a flatter curve for the calibrated variant. Separately, we test whether each method mislabels a pure proportion change as a state change, or the reverse.

The third check compares susceptible and resilient mice, analyzed separately by sex, given the non-overlapping sex-specific networks reported by Madeline L. Pfau et al.. It asks whether the residual test flags off-reference genes and whether those genes match stress-specific clusters in the snRNA-seq.

Only after the calibrated pipeline passes all three checks do we apply it to the published hypothalamic social-conflict bulk cohorts. There we report which cell classes carry the Th, Gad2 and cytokine changes, with the residual flags shown so that readers can see where the reference cannot be trusted.