Mitochondria inside a cell constantly split apart (fission) and merge (fusion). A healthy cell holds these two opposing processes in balance. Many diseases tip it: fission runs unchecked, and the organelle network shatters into dysfunctional fragments. Nearly everything known about this imbalance in tissue comes from methods that average over millions of cells, such as Western blots, respiration assays, and electron micrographs of a few fields. Those methods cannot say which cells in a diseased organ have tipped. Single-cell RNA sequencing (scRNA-seq), which reads the transcripts of thousands of individual cells, could in principle answer that. In practice there is no validated way to read a two-sided balance from sparse single-cell counts. There is also a quieter problem: standard quality control throws away cells with many mitochondrial reads, which may be exactly the cells under the most mitochondrial stress.
We propose a signed, per-cell score of fission-versus-fusion balance with two anchors. First, its gene weights are learned from cells whose mitochondrial shape was imaged after the fission and fusion genes were genetically perturbed, not from hand-picked machinery genes. Second, it is paired with a quality-control step that separates ruptured cells from cells under real mitochondrial stress. The central claim is testable: once the score is anchored to imaging, it predicts which way the balance tipped in held-out perturbations and disease tissue better than subtracting two off-the-shelf gene-set scores. The main risk is that fission and fusion are largely regulated at the protein level, so transcripts may carry too little signal. The first experiment is designed to find that out cheaply.
Why this gap is real
The mitochondrial dynamics literature keeps finding imbalance in disease through bulk readouts. Vanadium raises Drp1 and lowers Mfn1/Mfn2 in duck heart (Vanadium Induces Oxidative Stress…). Cigarette smoke lowers MFN2 and OPA1 during lung senescence (Role of mitochondrial fusion proteins MFN2 and OPA1…). Restoring SIRT3 rescues dynamics after liver ischemia-reperfusion (Dexmedetomidine ameliorates hepatic ischemia reperfusion injury…). These studies motivate the proposal: they show the balance moves in disease, but only at the tissue average.
The same literature also constrains the design, because the balance can move in either direction. Mild stress drives protective hyperfusion in presenilin-1 mutant neurons (Mitochondrial dysfunction and Ca2+ dysregulation in human iPSC-derived neurons…). Drp1 also has non-fission roles: it builds tunneling nanotubes that carry damaged mitochondria between cells in septic heart (Cytoskeletal remodeling promotes tunneling nanotube formation…). A score that simply tracks DNM1L transcript level would therefore mix up at least two biologies.
The two papers that already join mitochondrial dynamics with scRNA-seq use single cells only descriptively. The sepsis study clusters cardiac resident cells into Lcn2+ fibroblasts and Metal+ macrophages, then does its dynamics work by microscopy. The pulmonary fibrosis study (YBX1 regulation of alveolar type II epithelial cells…) follows YBX1 along an alveolar type II trajectory and measures mitochondria only through ATP and survival. Neither study scores the balance in each cell.
Single-cell enrichment methods supply machinery we can reuse, but none of them represents opposing forces. DeepGSEA (DeepGSEA) learns a separate neural head for each gene set and tests enrichment against prototypes. UNIFAN (Unsupervised Cell Functional Annotation) constrains an autoencoder decoder with a binary gene–pathway matrix. scETM (Learning interpretable cellular and gene signature embeddings) learns additive, positive-weight topics; it even recovered a mitochondrial topic in Alzheimer's tissue. iDEA (Integrative differential expression and gene set enrichment…) unifies differential expression and enrichment, but only at the cluster level. All four treat a pathway as an unsigned list whose members move together. ANS (ANS: Adjusted Neighborhood Scoring) examines how standard per-cell signature scores are biased by the expression level of their genes and proposes better-matched control genes. We reuse that idea directly, because the fission and fusion gene sets differ in size and expression level.
Core Claim
A per-cell balance score has to satisfy three properties that simpler scores lack.
The first property is that it is signed and removes the shared axis. When mitochondrial content or biogenesis rises in a cell, fission and fusion genes rise together. A naive difference of scores then partly reflects gene-set size and expression level rather than balance. We therefore remove a shared "total mitochondrial program" axis before contrasting the two sides.
The second property is that its weights are calibrated against organelle shape, not only machinery membership. The core machinery is about ten genes: DNM1L, FIS1, MFF and MIEF1/2 on the fission side, and MFN1, MFN2 and OPA1 on the fusion side. Many of these drop out, meaning they read as zero in most cells despite being expressed. Their activity is also set largely by post-translational events. We therefore learn a wider signed signature: the genes whose expression shifts when imaged morphology moves toward fragmentation or elongation after perturbation.
The third property is that it is computed on cells kept by a stress-aware quality-control step rather than a fixed mitochondrial-read cutoff.
Concretely, let $x_c$ be the normalized, control-gene-adjusted expression vector of cell $c$. Let $u$ be the shared mitochondrial axis, estimated as the first principal component of mitochondria-annotated nuclear genes. Let $w$ be a signed weight vector over genes, positive for fragmentation-associated genes and negative for elongation-associated ones. The balance score is
$$b_c = w^\top \left(I - \frac{u u^\top}{u^\top u}\right) x_c ,$$
which is the signed signature applied after projecting out the shared axis. The weights $w$ are fit by sparse regression of an imaging fragmentation index on expression in the calibration data described below. They are constrained so the core machinery genes carry their known signs.
Research Program
The program has three components: a calibration dataset, a stress-aware quality-control step, and a signed enrichment test.
Calibration dataset. This is the one new piece of data collection, and the project depends on it. In a tractable cell line, we deliver pooled CRISPR knockdowns of DNM1L, MFF, FIS1, MFN1, MFN2 and OPA1, together with non-targeting controls. We also apply two stressors known to push the balance in opposite directions: a mild stressor that induces hyperfusion and a toxic one that induces fragmentation. Each arm is imaged with a mitochondrial dye, giving a per-well fragmentation index (network branch length and fragment count). Each arm is also profiled by scRNA-seq with guide capture. Imaging and sequencing are matched at the arm level, not per cell, so $w$ is fit on arm-level pseudobulk labels and then applied per cell. Held-out arms are used for testing.
Stress-aware quality control. A ruptured cell shows high mitochondrial-DNA-encoded read fraction together with low total counts and depleted nuclear-encoded transcripts. A cell under genuine mitochondrial stress shows high mitochondrial reads with normal library size and a coordinated nuclear response. We fit a two-component mixture model on these features in place of the usual fixed cutoff. We then report how many high-mitochondrial-read cells are rescued, and whether those rescued cells score as fragmentation-leaning.
Signed enrichment test. Within each cell type, we test whether conditions differ in $b_c$ using a permutation null that reshuffles condition labels within sample. This keeps the unit of replication at the donor or animal, which avoids the inflated false-positive rates that come from treating individual cells as independent replicates.
Baselines
The naive reference point is DNM1L-to-MFN2 expression ratio per cell, with standard quality control that discards cells above a fixed mitochondrial-read fraction. This is what a practitioner would do with no new method. It should fail badly under dropout.
The strongest simple competitor is the difference of two standard module scores: a Seurat/Scanpy-style score of the GO "mitochondrial fission" set minus a score of the "mitochondrial fusion" set. We also run it with ANS control-gene adjustment, since ANS is the method in the file explicitly designed to fix biases in single-cell signature scores. This is the baseline to beat. It is cheap, it is signed, and ANS already handles the gene-set-level bias that motivates part of our design. No reported result found on Lacuna for any method scoring fission–fusion balance on these datasets, so the comparison numbers will be ours.
The strongest learned enrichment competitor is DeepGSEA, run with separate heads for the fission and fusion GO sets, and the balance read as the difference in prototype similarity. DeepGSEA was benchmarked against ssGSEA, GSVA, AUCell, Vision, SCPA and scGSEA on Splatter simulations and glioblastoma, influenza and Alzheimer's data. However, the file excerpt does not preserve its numeric scores: no reported result found on Lacuna for DeepGSEA's sensitivity or specificity values. We also include p-scETM with the two sets as fixed topics, to test whether a pathway-informed topic model recovers the contrast unaided.
For fairness, all methods receive the same count matrices, the same cells after each quality-control variant, and the same held-out arms. Learned methods get the same hyperparameter search budget, about 20 configurations each, selected on the same validation arms.
Two ablations isolate the key ingredients. The first uses our score with only the ~10 machinery genes and no imaging calibration. The second removes the shared-axis projection. Our method counts as better only if it beats the ANS-adjusted module-score difference on held-out perturbation direction by a margin exceeding the across-seed spread, and if both ablations lose ground.
Why not just…?
Why not just subtract a fission module score from a fusion module score? This is the real threat. If the ANS-adjusted difference matches our score on held-out arms, the imaging calibration adds nothing, and we will say so. The reason to expect a gap is structural. The machinery sets are tiny and dropout-prone, and their transcripts need not track morphology when Drp1 is regulated by phosphorylation and recruitment. The machinery-only ablation measures exactly this. If that ablation equals the full model, the proposal collapses to "ANS plus better quality control," and that would be the honest finding.
Why not just run DeepGSEA or UNIFAN on the GO fission and fusion terms? Both treat each set as unsigned and independent. Neither separates a cell where both sides rise with biogenesis from a cell whose balance has shifted. Simulations built with Splatter will test this directly. One condition plants a co-directional increase in both sets; the other plants an antagonistic shift. Unsigned methods should flag both conditions, while a balance score should flag only the second.
Why not just loosen the mitochondrial-read cutoff? Loosening the cutoff admits ruptured cells along with stressed ones, and ruptured cells will look artifactually "mitochondrial." The mixture model is justified only if the cells it rescues differ from cells rescued by a looser threshold. The test is whether the mixture-rescued cells show intact nuclear libraries and fragmentation-leaning scores that agree with imaging in the calibration data, while the extra cells admitted by a looser threshold do not.
Proposed Experiments
The decisive first check is cheap and could end the project. On the calibration data, we ask whether any transcriptome-based score predicts the sign of the imaged morphology shift across held-out knockdown and stressor arms. The metric is sign accuracy, plus rank correlation with the fragmentation index. If even the calibrated score is near chance, transcripts do not carry the balance, and that negative result is itself worth reporting to a field that routinely infers dynamics from DNM1L and MFN2 transcripts.
If the check passes, we transfer to disease data without retraining. The first transfer uses the human IPF and bleomycin mouse scRNA-seq sets used in the YBX1 study (GSE128033, GSE240134). The prediction is that alveolar type II cells tilt toward fragmentation along the injury trajectory, and that this tilt is stronger where YBX1 is low; that paper's in vitro work links YBX1 loss to mitochondrial collapse. The second transfer uses the septic cardiac data from the tunneling-nanotube study, if its data are available. There, cardiomyocyte Drp1 knockout provides a genetic contrast: resident fibroblasts and macrophages in knockout hearts should show a smaller fragmentation tilt than in wild-type septic hearts, because they receive fewer damaged mitochondria.
As a specificity check, we score the Alzheimer's and PBMC datasets used by DeepGSEA, scETM and iDEA. The balance score should vary across cell types without flagging cells that differ only in overall mitochondrial content. Across all experiments, we report how many cells the stress-aware quality control keeps that the fixed cutoff discards, and where those cells fall on the balance axis.
References
- Vanadium Induces Oxidative Stress and Mitochondrial Quality Control Disorder in the Heart of Ducks
- Role of mitochondrial fusion proteins MFN2 and OPA1 on lung cellular senescence in chronic obstructive pulmonary disease
- Dexmedetomidine ameliorates hepatic ischemia reperfusion injury via modulating SIRT3 mediated mitochondrial quality control
- Mitochondrial dysfunction and Ca2+ dysregulation in human iPSC-derived neurons carrying presenilin-1 mutation arise under stress via an MCU-1-independent mechanism
- Cytoskeletal remodeling promotes tunneling nanotube formation and drives cardiac resident cell mitochondrial transfer in sepsis
- YBX1 regulation of alveolar type II epithelial cells in idiopathic pulmonary fibrosis: mechanistic insights and small-molecule drug screening
- DeepGSEA: explainable deep gene set enrichment analysis for single-cell transcriptomic data
- Unsupervised Cell Functional Annotation for Single-Cell RNA-Seq
- Learning interpretable cellular and gene signature embeddings from single-cell transcriptomic data
- Integrative differential expression and gene set enrichment analysis using summary statistics for scRNA-seq studies
- ANS: Adjusted Neighborhood Scoring to improve gene signature-based cell annotation in single-cell RNA-seq data