Tens of thousands of post-mortem brain samples have been profiled with bulk RNA sequencing, which grinds a piece of tissue into one averaged measurement per gene. Researchers recover which cell types were present using deconvolution: they write the bulk profile as a weighted mixture of reference cell-type profiles taken from single-cell data and solve for the weights. In neurodegeneration this approach has a specific blind spot. When a bulk sample shows less of the signal belonging to a vulnerable neuron type, two explanations fit. The neurons may have died. Or they may still be alive but stressed, and have switched off the very marker genes the method uses to recognize them. The two outcomes carry very different clinical meaning, because a live stressed neuron can still be rescued. Batch effects make the problem worse. When cohorts from different labs are pooled, correction methods that force the cohorts to look alike can erase genuine cell loss when disease stage differs between cohorts, or they can report technical differences as neuron death.
We propose a deconvolution model that estimates two quantities separately for each fine neuronal subtype in each sample: how many cells are present (abundance) and how far the surviving cells have drifted from their healthy expression program (state shift). The model also estimates cohort batch effects under a constraint that stops them from absorbing either quantity. We will test whether this separation can be identified in practice on semi-synthetic bulk mixtures built from a large multiregional Alzheimer's atlas, where the truth is known because we inject it. The central empirical question is simple: can bulk data tell "20% of L2/3 intratelencephalic neurons are gone" apart from "all L2/3 neurons are present but have downregulated their markers by 20%", even when the cohorts differ technically?
Core Claim
The two signals can be separated because they act on genes differently. Losing cells scales every transcript of that cell type by the same factor. Stress-induced silencing changes a subset of genes, a low-dimensional program. The ALS evidence motivating this is direct. In $SOD1^{G93A}$ mice, marker transcripts in spinal inhibitory interneurons decline before the neurotransmitter transporters GlyT2 and Gad67 fall at late stages, and the authors concede that marker-based definitions cannot tell death from downregulation (Spinal inhibitory neurons degenerate before motor neurons and excitatory neurons in a mouse model of ALS).
We write the expected bulk count of gene $g$ in sample $s$ from batch $b(s)$ as
$$ \mathbb{E}[y_{gs}] = \beta_{g,b(s)} \sum_{k} n_{ks}\, \mu_{gk}\, \exp\!\big(\textstyle\sum_{r} W_{gkr} z_{ksr}\big). $$
Here $n_{ks}$ is the abundance of cell type $k$ in sample $s$, and $\mu_{gk}$ is the healthy reference expression. $W_{gk\cdot}$ is a small set of $r$ stress programs per cell type, learned from disease-stage variation inside a single-nucleus atlas. $z_{ks\cdot}$ is how strongly sample $s$ expresses those programs. $\beta_{g,b}$ is a gene-by-batch multiplicative technical factor.
Three constraints make the parameters identifiable. First, each stress program is centered to have zero mean on a panel of "anchor" genes: genes whose within-cell expression stays stable across disease stage in the atlas. Without this, $z$ could imitate a change in $n$. Second, the batch factors are shared across cell types. They are fitted with disease stage and pathology as explicit covariates, so a cohort that happens to contain more advanced cases does not have its neuron loss absorbed into $\beta$. Third, counts follow a negative binomial likelihood rather than a continuous approximation, in the spirit of the discrete-count argument in Count Bridges enable Modeling and Deconvolving Transcriptomic Data.
The model has a known failure mode. If stress causes a near-global shutdown of transcription within a cell type, the state shift and loss of abundance become truly confounded. The model should then report wide uncertainty rather than an answer. We will check whether its uncertainty intervals behave this way.
Relation to Prior Work
The closest work is EPIC-unmix. It builds an empirical Bayes prior from single-nucleus references, updates it on the bulk ROSMAP and MSBB Alzheimer's cohorts, and estimates both proportions and per-sample cell-type-specific expression. It already recovers expression shifts, but it does so for broad classes (neurons, astrocytes, microglia) and restricts itself to consensus marker genes. Those are exactly the genes that drift under stress. It also does not constrain the shift to be distinguishable from a change in proportion. We keep its idea of a reference-informed prior. What we change is the resolution (fine neuronal subtypes), the structure of the shift (low-rank, anchored), and the explicit comparison of loss against silencing.
GTM-decon resolves sub-cell-type states in prefrontal cortex with guided topic models. Two properties make it a weaker fit here: it has no batch layer, and it requires sparsifying the bulk data.
Three lines of work address batch effects between references and bulk cohorts. SCDC ensembles proportions estimated from multiple references, weighting them by reconstruction error. RNA-Sieve models sampling noise across platforms with a likelihood and reports calibrated intervals. POPPY aligns bulk and single-cell embeddings contrastively along the Cell Ontology, and its authors flag neurodegeneration as untested. None of the three separates loss from state change, and none guards against batch correction absorbing disease-driven differences in composition.
That last risk is a constraint on our design, not a hypothetical. Alignment methods built on distribution matching, such as BERMUDA and scAdapt, assume the populations shared across batches should overlap. That assumption is false when one cohort is biologically depleted. scBatchProx shows how easily unconstrained latent correction drifts. We therefore avoid distribution matching entirely and use a covariate-adjusted multiplicative batch term.
Several papers supply the biology we test against. An integrated, multiregional, and multimodal cell atlas of Alzheimer's disease documents selective depletion of L2/3 and L5 IT excitatory neurons and of Vip, Sst and Pvalb interneurons. It also shows different timing of pathology between limbic and neocortical regions, and it provides the more than 7 million nuclei from 84 donors that we use as our reference and simulation substrate. Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures shows that closely related lineages get misassigned when composition becomes extreme. We expect the mirror image in late-stage disease tissue, where residual neuronal signal gets misassigned once vulnerable neurons are nearly gone.
Research Program
The work proceeds in three stages. In the first stage we learn the reference quantities from the Alzheimer's atlas. We estimate healthy profiles $\mu$ from low-pathology donors. We select anchor genes by ranking genes on the stability of their within-cell-type expression across pathology stage. We learn the stress programs $W$ by factorizing, within each cell type, the variation in expression across donors that correlates with pathology after the anchors are held fixed.
A caveat applies to this stage. Single-nucleus isolation under-samples fragile degenerating cells, so the programs may reflect only the survivors that were healthy enough to be captured. We treat this as a known bias. We test sensitivity to it by fitting $W$ on subsets of donors at different pathology stages.
In the second stage we fit the full model to bulk data by maximum likelihood with weak priors on $n$ and $z$, and we produce posterior intervals for both. In the third stage we apply the model to the real ROSMAP and MSBB bulk cohorts. We ask whether the inferred abundance and state shift of vulnerable subtypes follow the region-specific pathology timing that the atlas describes.
Baselines
The naive reference point is non-negative least-squares deconvolution against healthy fine-subtype profiles, run after ComBat correction by cohort. It shows what a routine pipeline reports: every marker decline read as cell loss, and batch treated as a nuisance with no disease covariate.
The strongest existing competitor for this question is EPIC-unmix. It is the only method in the file validated on post-mortem Alzheimer's bulk cohorts, and it already estimates per-sample cell-type expression, so it could in principle absorb silencing. On ROSMAP and MSBB it reports "4.6% to 109.8% higher accuracy" in cell-type-specific unmixing than TCA and bMIND (EPIC-unmix). For fine neuronal subtypes we found no reported result on Lacuna.
For cross-protocol robustness the reference numbers are these. RNA-Sieve reports a mean $L_1$ error of 6.9% against 8.6% for the next-best method on Smart-seq2 versus 10x mixtures, with 95% intervals covering the truth 90.3% to 96.7% of the time (RNA-Sieve). SCDC reports a Pearson correlation of 0.991 on a cell-line mixture (SCDC). Neither number was measured on brain tissue.
All baselines receive the same fine-subtype reference drawn from the same atlas donors, the same gene universe, and the same bulk inputs. Each tunable method gets an equal hyperparameter search budget, selected on a held-out simulation split. EPIC-unmix is run twice: once with its default marker-gene selection and once on our full gene set, so that gene choice alone cannot explain any gap.
Two ablations isolate our key ingredients. In the first, we remove the state term ($z=0$), which reduces the model to discrete-count deconvolution with a batch term. In the second, we keep the state term but drop the anchor centering, so that any gain can be credited to the identifiability constraint rather than to extra capacity.
Beating the baselines means two things on the injected-silencing scenarios. Abundance error for vulnerable subtypes must be lower than EPIC-unmix, with non-overlapping bootstrap intervals over simulated cohorts. The model must also separate loss from silencing better than chance while the baselines, which have no separation mechanism, cannot. On pure-loss scenarios we must match EPIC-unmix, because the extra parameters must not cost accuracy when there is no silencing.
Why not just…?
Why not just use a disease-stage reference, so stressed neurons appear as their own reference states? This is the most serious alternative. The atlas spans the Alzheimer's spectrum, so one could split each subtype into healthy and stressed reference profiles and deconvolve against both. That approach turns a continuous shift that varies by donor into a few discrete bins. It also inherits the isolation bias against degenerating nuclei. Most importantly, it still does not constrain the stressed-state fraction to differ from loss. We include it as a baseline, fed to non-negative least squares and to EPIC-unmix. If it matches our model on the continuous-silencing simulations, the low-rank state term is unnecessary, and we would say so.
Why not just estimate proportions from disease-stable genes only? If anchor genes really do not change under stress, deconvolving on anchors alone gives abundance directly, with no state model at all. This is a real threat to the abundance half of the claim, and we run it as an explicit baseline. We expect it to lose precision, because anchors are few and weakly specific for fine subtypes. It also cannot report the state shift, which is the clinically relevant quantity. The experiment settles it: if anchors-only matches our abundance error, the contribution shrinks to estimating state.
Why not just batch-correct with disease stage as a covariate in ComBat, then deconvolve? Covariate-protected ComBat does address the confounding between batch and disease, and we test exactly that pairing. It does nothing about marker silencing, however. The confounded-design experiment below shows whether it still leaks disease signal when stage is strongly imbalanced between cohorts.
Why not just run single-nucleus sequencing? EPIC-unmix found that 57.4% of its cell-type-specific differentially expressed genes in large bulk cohorts were not seen in direct single-cell studies (EPIC-unmix). Cohort size matters, and nuclear isolation itself under-represents dying cells. Single-nucleus data is also unable, on its own, to tell a dead neuron from one lost during isolation.
Proposed Experiments
Experiment 1: loss versus silencing on semi-synthetic mixtures. We build pseudo-bulk samples by pooling atlas nuclei within each donor and region. We then apply one of three interventions to a vulnerable subtype. The first removes a fraction of its nuclei (loss). The second multiplies a set of genes in all of its nuclei by a factor drawn from real pathology-associated log-fold changes estimated in held-out donors (silencing). The third combines both. Finally, we multiply in gene-level batch factors drawn from observed differences between cohorts.
The vulnerable subtypes are L2/3 IT, L5 IT, Sst and Pvalb. As controls we use subtypes the atlas reports as relatively spared. We measure abundance error, state-shift error, the area under the receiver operating curve for classifying loss against silencing, and interval coverage. We also sweep neuronal depletion toward near-total loss, to probe the "depletion trap" in which residual neuronal signal is reassigned to glia. The critical test is held out: $W$ and the anchors are learned on one set of donors, and silencing is injected using changes estimated from different donors. This keeps the simulation from simply replaying the model's own assumptions.
Experiment 2: confounded batch design. We generate two synthetic cohorts whose disease-stage distributions range from balanced to heavily imbalanced. For every method we report how much true depletion is absorbed into the batch term. Our model should hold its abundance error roughly constant as imbalance grows, while the no-covariate baselines should degrade.
Experiment 3: real cohorts. We apply all methods to ROSMAP and MSBB bulk data. Wherever donors also have single-nucleus profiles, we compare inferred abundances with nucleus fractions, noting that nucleus fractions are themselves biased against fragile cells. We then test a biological prediction taken from the atlas: in limbic regions, where hyperphosphorylated tau accumulates early, state shifts in vulnerable subtypes should appear at earlier pathology stages than abundance loss, and this ordering should be weaker in neocortex.
Experiment 4: exploratory pilot in ALS. Using the multiplexed in situ imaging in $SOD1^{G93A}$ spinal cord across the P1–P112 time course (Spinal inhibitory neurons degenerate before motor neurons and excitatory neurons in a mouse model of ALS), we sum transcript counts per section to form small-panel pseudo-bulks. We then ask whether the model places the P63 decline of Renshaw and Ia interneurons in the state term before the abundance term, with the Sp8 clade serving as a resistant control. Because cell identity in this dataset is itself defined by markers, we treat the pilot as a consistency check, not ground truth.
The decisive result is Experiment 1 under held-out silencing. If the model cannot beat the anchors-only baseline and the stage-split reference baseline at telling loss from silencing, the claim fails, whatever happens on the real cohorts.
References
- Spinal inhibitory neurons degenerate before motor neurons and excitatory neurons in a mouse model of ALS
- Count Bridges enable Modeling and Deconvolving Transcriptomic Data
- Cell-type specific inference from bulk RNA-sequencing data by integrating single cell reference profiles via EPIC-unmix
- GTM-decon: guided-topic modeling of single-cell transcriptomes enables sub-cell-type and disease-subtype deconvolution of bulk transcriptomes
- SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
- A likelihood-based deconvolution of bulk gene expression data using single-cell references
- Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
- BERMUDA: a novel deep transfer learning method for single-cell RNA sequencing batch correction reveals hidden high-resolution cellular subtypes
- scAdapt: virtual adversarial domain adaptation network for single cell RNA-seq data classification across platforms and species
- Federated-inspired Single-cell Batch Integration in Latent Space
- An integrated, multiregional, and multimodal cell atlas of Alzheimer’s disease
- Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures