Predicting Which Cells Hold the Voltage: Single-Cell Resolution of Mitochondrial Membrane-Potential Heterogeneity Under Metabolic Stress

Drugs that attack mitochondria directly — by welding a fat-soluble, positively charged "delivery address" like triphenylphosphonium (TPP) onto a payload — accumulate electrophoretically only in mitochondria that maintain a strong negative inner-membrane potential ($\Delta\psi_m$) Potential-Driven Molecular Accumulation in Mitochondria. The uptake is exponential in the voltage: a cell whose mitochondria run just 10 mV more negative than its neighbor traps roughly $e^{10/26.7} \approx 1.5$-fold more of the drug at 37 °C, so a 30–40 mV spread across cells produces order-of-magnitude differences in who actually gets poisoned. The problem is that we almost never know which cells those are. The landmark demonstration that TPP-linked molecules selectively kill hyperpolarized Dnmt3a-mutant blood stem cells relied on bulk-sorted populations and bulk Seahorse flux assays Elevated mitochondrial membrane potential is a therapeutic vulnerability in Dnmt3a-mutant clonal hematopoiesis, which average oxygen consumption over hundreds of thousands of cells and therefore cannot say whether every cell is mildly oxidative or a small subpopulation is extremes oxidative. The same averaging problem plausibly explains a standing contradiction in the literature: the metabolic-reprogramming canon says stressed and malignant cells suppress mitochondrial respiration in favor of glycolysis (Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development; Elevated expression of glycolytic genes as a prominent feature of early-onset preeclampsia), while the organelle-targeting literature reports the opposite — hyperpolarized, respiration-heavy malignant cells Ju-Hee Kim et al.; Ferroptosis in cancer: revealing the multifaceted functions of mitochondria. Both cannot be true of the same cell, but both can be true of the same tissue, if averaging hides the mixture.

Core Claim

The proposal has three linked parts. First, build the missing measurement: a paired single-cell benchmark in which a potentiometric, organelle-targeted fluorophore (a membrane-potential-sensitive dye that doubles as an uptake reporter, in the style of the heptamethine cyanine IR-61 Jing Liu et al.) is imaged in the same wells, and a subset of the same physical cells, that yield single-cell RNA sequencing, across a panel of chemical and metabolic stressors. Second, train a conditional generative model that predicts per-cell $\Delta\psi_m$ and, critically, per-cell sensitivity to TPP-linked toxins from the transcriptome alone — so the hyperpolarized subpopulation can be identified in any existing scRNA-seq dataset. Third, close the loop at the bulk scale: apply discrete count deconvolution, specifically the Poisson birth-death generative model of Count Bridges enable Modeling and Deconvolving Transcriptomic Data, to bulk RNA-seq from stressed tissues to recover the fraction of hyperpolarized, drug-targetable cells — the number a bulk Seahorse assay cannot report and that no existing pipeline estimates.

The scientific payoff is a direct test of whether the Warburg-versus-OXPHOS contradiction is a masking artifact, in the sense argued conceptually by Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development: if single-cell voltage distributions within stressed populations are bimodal — a glycolytic, depolarized mode and an oxidative, hyperpolarized mode coexisting and possibly exchanging lactate — then bulk assays reporting an "intermediate" phenotype are reporting a mixture, and TPP-targeted therapies should kill a minority subpopulation, not the average cell.

Research Program

The dataset extends the scGeneScope design rather than replacing it. scGeneScope paired Cell Painting microscopy with scRNA-seq across 28 chemical compounds and DMSO controls in U2-OS cells over five biological replicates — 627,704 single-cell expression profiles and 716,767 single-cell images — but used broad morphological dyes and paired by treatment condition, not per cell, and measured nothing about mitochondrial potential. We keep the compound panel, cell line, and replicate structure (so our numbers are directly comparable to its published baselines) and add two things. One is a potentiometric cationic dye channel — candidates include TPP-conjugated fluorescent reporters and heptamethine cyanines, which are already validated as dual targeting-and-imaging agents in vascular smooth muscle cells Jianwu Wang et al. and whose accumulation reports exactly the electrophoretic sink that TPP therapeutics exploit L. Palanikumar et al.. The other is an index-sorting layer on a subset of wells, so dye intensity and transcriptome are measured on the same individual cells rather than matched sets — the pairing failure that limited PhenoSeq to set-level conditioning. The compound panel should deliberately include TPP-linked payloads (MitoQ-like antioxidants, and a DHODH-directed TPP conjugate in the spirit of compound B2 Ferroptosis in cancer) so the model can be trained to predict uptake and toxicity, not just voltage; the open problem in the targeting literature — that depolarized, stressed mitochondria lose their electrophoretic pull and abolish specificity — becomes a measurable, per-cell event rather than a population caveat.

The model is a two-head conditional generator. The transcriptome (or a fit-to-data latent embedding in the scVI family, per the scGeneScope finding that task-specific fitting beats zero-shot foundation models) conditions a discrete generative head over integer-valued expression counts — reusing the machinery of Count Bridges, whose birth-death process on integers was built precisely because continuous diffusion ignores that 10 counts is adjacent to 11 but far from 100 — and a regression head over per-cell $\Delta\psi_m$ and dye uptake. The deconvolution application then follows directly: pseudo-bulk the paired data at every mixture ratio, show the model recovers the hyperpolarized fraction from bulk counts alone, and validate the integer-preserving prior against continuous-flow baselines on exactly the perturbation regime where Count Bridges has never been tested (it was validated only on resting developmental tissue and standard slices).

A second validation system grounds the claim in the motivating biology: the competitive Dnmt3a-mutant versus wild-type transplantation setting of Elevated mitochondrial membrane potential.... We would deconvolve bulk RNA from mixed grafts and predict the hyperpolarized clone fraction over the 28-week outgrowth, testing against FACS-sorted ground truth — a prediction that paper's bulk Seahorse design could not make (no numerical OCR value is reported; no reported result found on Lacuna for single-cell $\Delta\psi_m$ distributions in that system).

Proposed Experiments

The primary metrics follow the scGeneScope protocol: balanced accuracy for per-cell and per-treatment classification under within-experiment and held-out-experiment splits, and — new here — calibration of predicted per-cell voltage against measured dye intensity (rank correlation, and absolute error in mV for the index-sorted subset), plus predicted-versus-measured TPP-toxicity at the single-cell level. The decisive checks are four. Does adding the potentiometric channel beat the published morphology-only ceiling on stressors that fail morphologically — Simvastatin being the canonical total breakdown in PhenoSeq? Does the single-cell voltage distribution under metabolic stressors come out bimodal where the bulk phenotype looks intermediate — the mixture-artifact test of the Warburg contradiction? Does transcriptome-only prediction of the hyperpolarized fraction transfer to held-out experiments, or does it degrade the 5–7% balanced-accuracy loss that foundation models showed on held-out splits in scGeneScope? And does bulk deconvolution recover the hyperpolarized fraction more accurately than reference-free topic modeling, on mixtures constructed outside any reference atlas — the regime where an atlas-dependent method like POPPY, which improved cross-cohort integration by up to 73% silhouette score but requires comprehensive reference atlases Cross-modal transfer learning for mapping bulk transcriptomes at cellular level, is expected to misclassify metabolically rewired intermediate states?

Baselines

The naive reference point is the do-nothing measurement that actually produced the motivating result: bulk Seahorse oxygen-consumption plus bulk RNA-seq on sorted populations, exactly as run in Elevated mitochondrial membrane potential.... This baseline is the strongest competitor by influence because it is the field's current standard and it did successfully identify the therapeutic window in clonal hematopoiesis; beating it means demonstrating a prediction it cannot make at any cost — namely, the identity and fraction of hyperpolarized cells within an unsorted mixture, validated against index-sorted single-cell dye measurements. It will be run on aliquots of the very same wells used for single-cell assays, so the comparison is between information content, not biological material.

The strongest existing methodological competitor is the PhenoSeq pipeline of Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion, which is the closest published system: microscopy-conditioned generative prediction of single-cell transcriptomic embeddings on 28 chemical perturbations, reporting 0.293 versus 0.270 balanced accuracy for synthetic-RNA embeddings over raw imaging, recovery of roughly 29% of the imaging-to-sequencing performance gap, and 0.271 versus 0.157 imaging-only accuracy when fusing 32 cells per sample. We will run PhenoSeq unchanged on our images as the incumbent generative baseline, with the same data, embedding space, and tuning budget, and additionally give it our potentiometric channel as a fourth input — because if raw morphology plus voltage closes the gap, that is the honest result. Our method counts as beating it only if per-cell voltage prediction and stressor classification on held-out experiments exceed PhenoSeq's numbers and survive the Simvastatin-style perturbations where its accuracy collapses. For the deconvolution arm, the incumbent is Count Bridges itself applied off-the-shelf Count Bridges, which reported Jensen-Shannon Divergence 0.231 and RMSE 0.110 versus STDeconvolve's 0.288 and 0.177 on simulated spots; our contribution there is not a new deconvolver but the first validation under organellar-targeted and metabolic-stress perturbations, so the fair run is the same model, same compute, retrained and zero-shot on our stress mixtures.

The key ablation isolates the proposal's central ingredient: the identical generative model, training schedule, and data, with the potentiometric organelle-targeted dye channel removed, leaving standard Cell Painting morphology. If the ablated model matches the full model on per-cell voltage prediction and on held-out-experiment stress classification, then the dye is decorative and the claim reduces to "morphology already predicts metabolism" — which the PhenoSeq Simvastatin failure makes unlikely but which must be tested rather than assumed. A second ablation swaps the discrete count prior for a continuous diffusion head, isolating whether integer-valued modeling matters specifically in the stressed, perturbation regime.

Why not just…?

Why not just use Cell Painting morphology without any new dyes? Because the morphology ceiling is documented on exactly this benchmark family: when a perturbation leaves gross cell shape unchanged, as Simvastatin does in U2-OS cells, image-based prediction fails entirely in PhenoSeq, and even where it works, imaging alone reached only 0.270 balanced accuracy versus 0.293 for predicted transcriptomic embeddings. Metabolic shifts are the canonical invisible phenotype — a cell can swap glycolysis for oxidative phosphorylation with little morphological signature — so the potentiometric channel is aimed precisely at the failure mode. That said, this alternative is a real threat to the dataset contribution if morphology turns out to be sufficient, and the ablation above is designed to settle it rather than argue around it.

Why not just use zero-shot single-cell foundation models (scGPT, Geneformer) to predict the stressed states? Because on the very benchmark we extend, the rigorous comparison in scGeneScope showed zero-shot foundation models were consistently beaten by classical fit-to-data methods like scVI, and lost a further 5–7% balanced accuracy on held-out experiments. Foundation models are trained on atlas-scale data that under-represents acute organellar stress, so we adopt fit-to-data latents as the conditioning representation. We will still run one foundation model zero-shot as a reference, but the evidence already says it will not be the winner.

Why not just run Seahorse and bulk RNA-seq, as the Dnmt3a study did? That study is the proposal's motivation, not its competitor to dismiss: it proved the therapeutic window exists, and its own limitation section concedes the single-cell structure was unresolved. Bulk assays cannot distinguish a uniform intermediate metabolic state from two coexisting subpopulations trading lactate — the distinction this proposal exists to measure — and bulk sequencing demonstrably cannot separate cell-autonomous stress cascades from compensatory bystander responses, as shown directly by the Efnb2 knockout analysis in RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons and by the loss of rare resistance-driving stem-like cells in population averages Palmer et al.. If our single-cell voltage distributions turn out unimodal and tightly clustered, the bulk view is vindicated and the mixture-artifact explanation of the Warburg contradiction fails — that is the falsifiable outcome.

Why not just use existing deconvolution (Count Bridges or STDeconvolve) on bulk data as-is? Because both were validated only on resting or unperturbed tissue. The open limitation stated for Count Bridges itself is that it has not been applied to metabolically reprogrammed cells under drug intervention, and the limitation for atlas-based transfer like POPPY is that metabolically rewired intermediate states fall outside any pre-existing reference atlas. Our contribution is a deconvolution validation regime — organellar-targeted, metabolically stressed mixtures with single-cell ground truth — plus the first estimate of a quantity no deconvolution paper has attempted: the hyperpolarized, drug-sensitive cell fraction. If off-the-shelf Count Bridges already recovers that fraction accurately under stress, our deconvolution claim narrows to the benchmark contribution, and we would report that plainly.

Decisive Checks

The proposal stands or falls on two numbers. First, on the held-out-experiment split of the stress panel, the full model (transcriptome plus potentiometric channel) must beat both the published PhenoSeq imaging-conditioned numbers (0.293 single-profile, 0.271 at 32 cells per sample) and its own no-dye ablation, with the margin concentrated on the perturbations that are morphologically silent. Second, on pseudo-bulk and on the Dnmt3a graft system, the deconvolved hyperpolarized fraction must track the sorted ground truth with error small enough to rank candidate tissues by targetable-cell fraction — the ranking a clinician would actually need before giving a TPP-linked drug. If both hold, the masking artifact behind the glycolysis-versus-OXPHOS contradiction gains its first direct single-cell evidence, and mitochondrial targeting gains the patient-selection instrument it currently lacks.

References

  1. Emerging magic bullet: subcellular organelle-targeted cancer therapy
  2. Elevated mitochondrial membrane potential is a therapeutic vulnerability in Dnmt3a-mutant clonal hematopoiesis
  3. Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
  4. Elevated expression of glycolytic genes as a prominent feature of early-onset preeclampsia: insights from integrative transcriptomic analysis
  5. Intra-mitochondrial reaction for cancer cell imaging and anti-cancer therapy by aggregation-induced emission
  6. Ferroptosis in cancer: revealing the multifaceted functions of mitochondria
  7. A mitochondria-targeting heptamethine cyanine-chlorambucil formulated polymeric nanoparticle to potentiate native tumor chemotherapeutic efficacy
  8. Count Bridges enable Modeling and Deconvolving Transcriptomic Data
  9. Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development
  10. scGeneScope: A Treatment-Matched Single Cell Imaging and Transcriptomics Dataset and Benchmark for Treatment Response Modeling
  11. IR-61 Improves Voiding Function via Mitochondrial Protection in Diabetic Rats
  12. Cell Painting Generates Single-Cell Transcriptomics via Conditional Diffusion
  13. Cross-modal transfer learning for mapping bulk transcriptomes at cellular level
  14. RNA-seq co-expression network analysis reveals anxiolytic behavior of mice with Efnb2 knockout in parvalbumin+ neurons
  15. A gene expression profile of stem cell pluripotentiality and differentiation is conserved across diverse solid and hematopoietic cancers