Counting Depolarized Cells: Mito-State Deconvolution of Bulk Transcriptomes

Nearly every large clinical cohort in sepsis, stroke, and acute kidney injury is profiled with bulk RNA-seq, which pools the messenger RNA of thousands to millions of cells into one average measurement. That averaging hides which cells actually carry a disease signal, and a family of computational tools answers by deconvolution: given a single-cell atlas as a reference, estimate what mixture of cell types produced a given bulk sample. But there is a second, quieter loss sitting inside those references themselves. Standard single-cell pipelines delete any cell whose sequenced reads are dominated by mitochondrial genes, on the assumption that such cells had ruptured plasma membranes and leaked their cytoplasmic RNA before capture. In diseases whose primary lesion is the mitochondrial membrane itself — septic inflammatory lysis, ischemic injury, PANoptosis (a coordinated execution of apoptosis, necroptosis, and pyroptosis) — the cells actively undergoing that lesion look exactly like the artifact the filter is designed to remove.

The consequence is a survivorship bias that compounds across all three research directions. Quality control deletes depolarizing cells from single-cell atlases; the corrupted atlases become references for deconvolution; and deconvolution then cannot see mitochondrial stress states in bulk clinical data at all — precisely the limitation POPPY names outright: states absent from the reference cannot be unmixed. Meanwhile, the mitochondrial stabilization literature (Chen et al.; Ma et al.) shows these discarded cells are the ones a stabilizer drug could rescue. This proposal builds the missing measurement instrument: a deconvolution that reads mitochondrial membrane failure, and the effect of membrane-stabilizing drugs, directly out of bulk tissue transcriptomes.

Core Claim

Standard single-cell quality control does not merely remove noise; it removes the target population in mitochondrial-stress diseases, and every reference-based deconvolution inherits that blindness. By (i) quantifying the filter's deletion rate per physiological state, (ii) constructing references annotated by both cell type and mitochondrial state, and (iii) extending ensemble deconvolution to jointly estimate cell-type-by-state fractions, we can recover depolarized, hyperpolarized, and organelle-recipient cell fractions from bulk RNA-seq that current methods cannot resolve.

Two relations carry the whole argument. First, the reference bias. Let states $s \in \mathcal{S}$ = {polarized, depolarized, hyperpolarized} occur within cell type $t$ at true proportions $\pi_{ts}$, and let $r_s = \Pr(\text{mito\%} < \tau \mid s)$ be the probability a cell in state $s$ survives the standard mitochondrial-read-percentage threshold $\tau$. A quality-controlled reference then estimates

$$\hat{x}_t \propto \sum_{s \in \mathcal{S}} \pi_{ts}\, r_s\, x_{ts},$$

where $x_{ts}$ is the true expression profile of type $t$ in state $s$ — so depolarized states are downweighted in direct proportion to how aggressively the filter deletes them. Second, the fix. We deconvolve bulk profiles $b$ with the state-resolved mixture model

$$b \approx \sum_{t \in \mathcal{T}} \sum_{s \in \mathcal{S}} w_{ts}\, x_{ts}, \qquad \sum_{t,s} w_{ts} = 1,$$

estimating fractions $w_{ts}$ jointly, where every existing method collapses $s$ to a single value and is therefore structurally incapable of separating "cells changed state" from "cell proportions changed" — the exact confound Adams et al. and Tsoucas et al. identify as the core ambiguity of bulk data.

Research Program

The first work package measures $r_s$ directly rather than assuming it. We use the assay systems already established in the stabilization literature: 6-OHDA-treated differentiated PC12 cells and L-glutamate excitotoxicity (Wang et al.), LPS-challenged RAW264.7 macrophages (Chen et al.), and H₂O₂-stressed B35 neuroblastoma cells (Hu et al.). In each system we sort cells with JC-1, the ratiometric dye these same papers use to read the voltage across the inner mitochondrial membrane ($\Delta\Psi_m$), into polarized and depolarized fractions, sequence both with deep coverage, and measure how much of each fraction the conventional 5–10% mitochondrial-read filter removes. Deep sequencing with diluted bulk RNA controls, the design Dueck et al. used at roughly 5 million uniquely mapped reads per cell to prove that within-type expression variation is conserved biology rather than technical noise (their cross-species correlations reached $p < 10^{-13}$), lets us confirm that depolarization signatures are genuine regulated programs and not lysis artifacts — a prerequisite for treating them as reference states. In parallel, a zero-cost reanalysis of public disease atlases (the multiple sclerosis, Alzheimer's, and traumatic brain injury microglia datasets GSE234700, GSE243243, GSE167459 used by the microglia disease study) rescues the discarded high-mitochondrial barcodes from raw matrices and tests whether they form coherent cGAS-STING and PANoptosis-signature populations.

The second package builds the state-annotated references and the deconvolver. We implement the model above in two ways to avoid betting on one inference machinery: an ensemble-weighted extension of the SCDC approach, which already solves the reference-batch problem by weighting multiple single-cell references against bulk reconstruction error (Tsoucas et al.), and an integer-count extension of the Count Bridges expectation-maximization machinery, which models transcript counts as discrete birth-death processes rather than continuous fractions (Count Bridges). The one genuinely new modeling ingredient is state-dependent capture: a depolarized cell retains mitochondrial transcripts while losing cytoplasmic ones, so its mitochondrial-read fraction is a feature of the state, and we parameterize a per-state capture profile in the likelihood rather than treating all genes under identical priors as Count Bridges does.

The third package applies the tool where the biology says it matters. In the cecal-ligation-and-puncture sepsis model, Chen et al. showed by single-cell sequencing that cardiomyocytes export damaged, depolarized mitochondria through tunneling nanotubes, reprogramming recipient endothelial cells, fibroblasts, and macrophages into distinct pathological subpopulations — but whole-tissue assays of such hearts average donor and recipient signals together and no bulk method can recover them. We deconvolve bulk cardiac RNA from septic mice with and without the membrane stabilizer 4-OI, which Ma et al. showed prevents mitochondrial DNA leakage and reduces lung wet-to-dry weight ratio in the same model, and test whether treatment reduces the depolarized and organelle-recipient state fractions within cell types rather than shifting cell-type proportions — the distinction no current bulk analysis can make. A second application targets the therapeutic paradox: Young et al. showed that hyperpolarized, hyper-respiring Dnmt3a-mutant blood stem cells dominate aged marrow over 28 weeks of competitive transplantation yet are selectively killed by the membrane-potential-driven accumulation of lipophilic cations like MitoQ. We deconvolve longitudinal bulk marrow from that experiment to ask whether clonal purge and clonal expansion are visible as hyperpolarized-state-fraction changes — the safety readout needed before any indiscriminate membrane stabilizer is given to a tissue containing both collapsing healthy cells and pre-leukemic hyperpolarized clones.

Baselines

The naive reference point is what a clinical analyst runs today: a standard quality-controlled public atlas fed to off-the-shelf CIBERSORTx or MuSiC. This baseline matters because it is the default, and because Tsoucas et al. showed such tools can detect lineage shifts reasonably in clean mixtures but "failed to find a significant association or produced results that varied wildly" on real tissue (human islets), and the simulated bulk benchmark study showed they degrade badly when states share lineage memory. It will be run on the identical bulk mixtures and atlases with its standard protocol.

The strongest existing competitor is SCDC ensemble deconvolution (Tsoucas et al.), the method with the best reported ground-truth number in the corpus: Pearson $r = 0.991$ against known 6:3:1 mixtures of MDA-MB-468, MCF-7, and fibroblasts, and the only method to recover the HbA1c–beta-cell association at $p = 0.0019$ across 77 islet donors. We run SCDC on our bulk data with the same compute and tuning budget, with both quality-controlled and unfiltered references, and additionally run BayesPrism and DWLS, which the simulated bulk benchmark study identified as most stable across mixture purity and lowest in combined false positives and false negatives on granular lineages; for state-resolved accuracy there is no reported result found on Lacuna for any of these methods, so we set the bar ourselves. Beating the baselines means recovering known state fractions at lower error than every one of them on the mixture benchmarks below, and detecting ground-truth state shifts that the quality-controlled-reference runs miss entirely.

The key ablation isolates the proposal's central ingredient. We run the full state-resolved machinery with the state index collapsed to a single value per cell type, and separately with an unfiltered reference but no state labels and unsupervised clustering. The first ablation tests whether the gain comes from the state parameterization or merely from keeping high-mitochondrial cells in the reference. The second is the honest threat: if unsupervised clusters in unfiltered data recover the JC-1-sorted ground truth as well as explicit state annotation does, the contribution narrows to "turn the filter off," and we will say so.

Why not just…?

Why not just use off-the-shelf deconvolution with public references? This is a real threat for lineage-level questions, and we concede it: CIBERSORTx, MuSiC, and SCDC can detect cell-type proportion shifts. The problem is that public references are already quality-controlled, so the depolarized states are absent before any algorithm runs — the missing-state limitation POPPY states explicitly — and the simulated bulk benchmark study shows these tools confuse states that share lineage memory. Our known-proportion mixture experiments, where ground truth state fractions are fixed by mixing, settle whether any existing method recovers them.

Why not just regress out mitochondrial percentage or use a deep latent model like scVI? These tools treat the mitochondrial-read fraction as nuisance to be removed. SiFT showed that regress_out and scVI distort genuine biological variation when removing strong nuisance signals; here the situation is worse, because the mitochondrial fraction is the target signal, and regressing it out deletes the measurement this proposal exists to make.

Why not just run single-nucleus RNA-seq to avoid the lysis artifact entirely? Sequencing nuclei removes cytoplasm and with it every mitochondrial transcript, so organellar retention, cytoplasmic leakage, and organelle transfer through tunneling nanotubes become unmeasurable — and the single-cell versus single-nucleus comparison study found single-cell sequencing outperforms single-nucleus sequencing for metabolic gene detection anyway. Nuclei also do not solve the bulk problem: the clinical cohorts we want to analyze are already frozen bulk RNA.

Why not just give patients antioxidants or anti-cytokine biologics instead of building a measurement tool for membrane states? Because the stabilization literature indicates upstream membrane closure and downstream scavenging are not interchangeable: keeping the permeability-transition pore closed protects far more robustly than antioxidant scavenging in excitotoxicity (Kumari et al.) and in doxorubicin cardiotoxicity (Zhang et al.), and once mitochondrial DNA escapes it triggers PANoptosis, which executes three parallel death programs that bypass single-cytokine blockade (Chen et al.). Distinguishing which mechanism a therapy actually engaged requires exactly the state-resolved readout we propose — and the Young et al. paradox means indiscriminate stabilization could rescue hyperpolarized pre-leukemic clones, a failure mode only a state-resolved bulk readout can detect in longitudinal cohorts.

Decisive Checks

The primary benchmark mirrors the design that gave SCDC its credibility: experimental mixtures with known composition. Instead of three distinct cell lines, we mix drug-stressed and healthy cells of the same lineage — 6-OHDA-treated and untreated differentiated PC12 cells (Wang et al. report viability dropping to approximately 40% under this insult, so a 40%-stressed mixture is directly grounded) and LPS-treated versus resting RAW264.7 cells — at fixed ratios of 3:7, 5:5, and 7:3, sequenced in bulk. Success is recovering the stressed-state fraction with root-mean-square error below 0.05 and beating every baseline on the same mixtures; SCDC's $r = 0.991$ stands as the accuracy its lineage machinery achieves, with no reported result found on Lacuna for any method on state-resolved mixtures of a single lineage.

The second check is the filter audit: if the JC-1-depolarized fraction is deleted at a substantially higher rate than the polarized fraction under the standard threshold, and the rescued high-mitochondrial barcodes from the public microglia atlases form coherent inflammatory-death programs rather than lysis noise, the survivorship bias is real and consequential. The third check is the in vivo application: in the sepsis model, 4-OI treatment should reduce depolarized and organelle-recipient state fractions within endothelial, fibroblast, and macrophage compartments — the subpopulations Chen et al. defined — in parallel with the physiological rescue Ma et al. measured by lung wet-to-dry ratio and Evans Blue leakage, while leaving overall cell-type proportions comparatively unchanged. The fourth is the paradox arm: in the Dnmt3a-mutant competitive transplantation of Young et al., MitoQ's selective purge should appear in bulk marrow as a collapse of the hyperpolarized stem-cell state fraction over the 28-week course, a signature no lineage-only deconvolution can produce. Finally, the ablation above decides the honest question of whether explicit state annotation is necessary or whether simply keeping the filtered-out cells suffices; either answer is a publishable finding, but only the mixture experiment can tell us which one we have.

References

  1. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
  2. KAE ameliorates LPS-mediated acute lung injury by inhibiting PANoptosis through the intracellular DNA-cGAS-STING axis
  3. 4-Octyl itaconate alleviates endothelial cell inflammation and barrier dysfunction in LPS-induced sepsis via modulating TLR4/MAPK/NF-κB signaling
  4. As we come to the end of 2011, several members of the Genome Biology Editorial Board give their views on the state of play in genomics
  5. SCDC: Bulk Gene Expression Deconvolution by Multiple Single-Cell RNA Sequencing References
  6. Investigation of the neuroprotective effects of a novel synthetic compound via the mitochondrial pathway
  7. 20-Hydroxyecdysone Protects against Oxidative Stress-Induced Neuronal Injury by Scavenging Free Radicals and Modulating NF-κB and JNK Pathways
  8. Deep sequencing reveals cell-type-specific patterns of single-cell transcriptome variation
  9. Analyzing single-cell transcriptome sequencing data reveals immunological mechanisms of microglia-associated diseases
  10. Count Bridges enable Modeling and Deconvolving Transcriptomic Data
  11. Cytoskeletal remodeling promotes tunneling nanotube formation and drives cardiac resident cell mitochondrial transfer in sepsis
  12. Elevated mitochondrial membrane potential is a therapeutic vulnerability in Dnmt3a-mutant clonal hematopoiesis
  13. Performance of tumour microenvironment deconvolution methods in breast cancer using single-cell simulated bulk mixtures
  14. SiFT: uncovering hidden biological processes by probabilistic filtering of single-cell data
  15. Single-Cell RNA Sequencing Outperforms Single-Nucleus RNA Sequencing in Analyzing Pancreatic Cell Diversity and Gene Expression in Goats
  16. Ubisol-Q10 Prevents Glutamate-Induced Cell Death by Blocking Mitochondrial Fragmentation and Permeability Transition Pore Opening
  17. Sirtuin-3 Regulates the Mechanism of Doxorubicin-induced Cardiotoxicity