During REM sleep, the brain is electrically awake while the body is paralyzed. The small groups of neurons in the midbrain and brainstem that generate this state — and that actively shut down the motor system to prevent us from acting out dreams — sit in some of the most iron-rich territory in the brain. That matters because ferroptosis, an iron-catalyzed form of cell death in which membrane fats oxidize and rip the membrane apart, is normally held in check by antioxidant enzymes like GPX4, and it is triggered precisely by the combination of high labile iron and weakened lipid defenses. Meanwhile, genetic lineage tracing — the technique of permanently marking an ancestor cell's DNA so that all of its descendants remain identifiable no matter what genes they later turn on or off — has never been applied to these sleep-generating nuclei. The result is that we cannot currently answer a basic question: when sleep-generating neurons die, do they die because of where they are (a wave of lipid peroxidation sweeping indiscriminately through neighbors), or because of who they are (a developmental lineage that inherited weak antioxidant defenses)?
This question is not idle. Sleep loss has been shown to cause ferroptosis in hippocampal neurons, and loss of muscle paralysis during REM sleep is one of the earliest predictors of Parkinson's disease, decades before motor symptoms appear, in patients whose substantia nigra is an established epicenter of iron accumulation. But the only study connecting sleep loss to ferroptosis examined the hippocampus under total sleep deprivation, with no way to distinguish vulnerable from resistant developmental lineages. The sleep-circuit field, in turn, identifies neurons by current marker genes (Gad2, Pvalb, Vglut2) and treats them as static populations, blind to clonal history. This proposal combines all three directions to test a specific, falsifiable claim.
Core Claim
Ferroptotic death within the REM-generating circuits of the substantia nigra pars reticulata (SNr) and the dorsal pons is not spatially indiscriminate but clonally structured: among neurons sharing the same marker identity, specific developmental lineages — those that inherited low baseline expression of the System Xc⁻/glutathione/GPX4 antioxidant axis — act as the initiators of propagating lipid-peroxidation waves, while sibling lineages with stronger inherited defenses act as firebreaks. This claim is currently untestable for two reasons: no one has barcoded these circuits' lineages, and no computational lineage model can even represent a dying branch, since every existing method (Deep Lineage, TreeVAE, CoSpar, LCL) implicitly assumes all sampled cells survive.
The claim deliberately pits two literature findings against each other. In acute kidney injury, PAFAH2 suppresses synchronized ferroptosis to ameliorate acute kidney injury characterizes ferroptosis as a synchronized, non-cell-autonomous wave crossing neighboring cells — under a pure wave model, clonal identity should predict nothing. In contrast, lineage-tracing work in differentiation and cancer (Deep Lineage, CoSpar, Actionable biological programs to enhance EGFR-targeted therapy response) shows cell fates are strongly determined by heritable clonal history — under a clonal-competence model, death should cluster by lineage even among interleaved neighbors. This conflict has never been tested in any tissue, and the brain's sleep nuclei are the ideal place to test it because the functional consequence of the answer — which neurons die, and therefore whether REM paralysis survives — is directly measurable.
Research Program
The program has three linked parts: a wet-lab lineage-and-stress experiment in mouse, a spatial-versus-clonal death analysis, and a new computational model that adds an explicit death state to lineage inference.
The biological experiment begins by permanently barcoding midbrain and dorsal pontine progenitors during a defined developmental window using an inducible recombinase system — a tamoxifen-activated CreER that, at a low pulse dose, excises a stop cassette in a sparse set of ancestor cells, permanently switching on both a fluorescent reporter and a slowly mutating CRISPR/Cas9 barcode cassette that descendants accumulate, exactly the architecture used to track 603 lung cancer cells in TreeVAE and clones in Deep Lineage. Because the recombination mark is permanent and heritable, descendants remain identifiable even if stress causes them to downregulate the very markers (Gad2, Vglut2) normally used to find them — the failure mode that conventional Cre lines cannot escape.
The stressor is applied in adulthood, after circuits mature. Two arms run in parallel: a chronic REM-fragmentation arm (validated by continuous EEG/EMG with automated behavioral state classification, following the four-state framework of A common hub for sleep and motor control in the substantia nigra), and a pharmacological arm that directly blocks cystine import through the System Xc⁻ antiporter, which starves cells of the glutathione precursor needed by GPX4 and is the canonical ferroptosis trigger described in Ferroptosis: mechanism, immunotherapy and role in ovarian cancer and Multifaceted mechanisms mediating cystine starvation-induced ferroptosis. Rescue arms add lipid-radical-trapping inhibitors and iron chelation, following the cytoprotective logic of Silibinin Inhibits Cell Ferroptosis and Ferroptosis-Related Tissue Injuries.
The death readout is per-cell lipid peroxidation (C11-BODIPY) plus GPX4 and ferritin immunoreactivity in barcoded SNr and dorsal pontine neurons, with dying cells captured before membrane lysis destroys their transcriptome. The single decisive statistic is clonal overdispersion: if the fraction of dead cells within each clone, controlling for cell type, varies far beyond binomial expectation, death is lineage-structured; if death fractions are uniform across clones but spatially clustered among physically adjacent neurons, the wave model wins. Spatial autocorrelation of death versus clonal autocorrelation of death, computed on the same tissue sections, cleanly separates the two. The behavioral endpoint is whether REM atonia degrades as susceptible clones die — the mouse analogue of the REM-sleep-without-atonia measured clinically in Detection of REM Sleep Behaviour Disorder by Automated Polysomnography Analysis, which achieved 92% diagnostic accuracy with 91% sensitivity and 93% specificity on a 106-participant cohort.
The computational part closes the methodological gap the field itself has flagged. We will extend the autoencoder-over-phylogeny architecture of TreeVAE with a per-branch death hazard, so that the model learns the latent pre-death states of branches whose terminal cells never appear in the sequenced pool. The model is trained jointly on our barcoded sleep-circuit data and validated on the mouse hematopoiesis benchmark (clonal differentiation of blood progenitors over six days), where Deep Lineage reports 77% early fate-prediction accuracy versus 65% for CoSpar and WOT, and where removing clonal information collapses accuracy to 52% — the clearest published evidence that clonal context, not snapshot transcriptomics, carries fate information.
Proposed Experiments
Experiment 1 maps the baseline ferroptosis-defense landscape of the sleep nuclei before any stress: single-nucleus RNA sequencing of SNr and dorsal pons, testing whether GAD2 sleep-promoting neurons, PV motor-arrest neurons (the functional split established in A common hub for sleep and motor control in the substantia nigra), and REM-active glutamatergic versus wake-active GABAergic dorsal pontine neurons (the segregation found in Calcium imaging of sleep–wake related neuronal activity in the dorsal pons) differ systematically in GPX4, SLC7A11, ferritin, and PAFAH2 expression. This establishes whether marker class already predicts defense capacity, and provides the prior against which clonal deviations are measured. No reported result found on Lacuna for antioxidant-axis expression levels in these specific nuclei — this measurement itself is new.
Experiment 2 is the core stress-and-death experiment described above, with the clonal overdispersion statistic as the primary endpoint and the spatial-autocorrelation contrast as the competing hypothesis test. Secondary endpoints include the REM atonia index under the same three-channel reduced montage logic used clinically (where automated staging achieved an overall Cohen's kappa of 0.62, the floor we would need to match or exceed in mice), and dose-response of cell loss to the cystine-import blockade.
Experiment 3 trains and benchmarks the death-augmented lineage model. On hematopoiesis, the target is to match or exceed the 77% accuracy reported for Deep Lineage while additionally predicting which clones lose cells to an induced death stimulus — a task no current tool addresses. On our sleep-circuit dataset, the target is beatable performance: predicting clone-level death fractions from pre-stress transcriptomes plus barcodes. For ancestral state reconstruction, the reference point is TreeVAE's reported latent mean squared error of 1.89 versus 2.28 for a standard VAE and its cross-entropy improvement from 21,348 to 3,051 on the lung cancer benchmark; our model should not degrade these figures while gaining death-branch inference.
Experiment 4 is the causal check: whether blocking ferroptosis pharmacologically (inhibitors in the class reviewed in Ferroptosis inhibitors: past, present and future) spares the vulnerable clones and preserves REM atonia, and whether experimentally raising GPX4 output in the predicted-vulnerable lineage specifically (by targeting the axis characterized in MAPK14/SLC7A11/GPX4 axis dysregulation drives podocyte ferroptosis) converts initiators into firebreaks.
Baselines
The first baseline is the do-nothing reference: uniform vulnerability within each marker class. Concretely, this is the assumption, implicit in all current sleep-circuit work (A common hub for sleep and motor control in the substantia nigra, Calcium imaging of sleep–wake related neuronal activity in the dorsal pons), that every GAD2 or glutamatergic neuron in a nucleus carries the same death risk, so that per-clone death fractions are binomial draws around the class mean. It is the strongest competitor in the sense that if it holds, the entire clonal apparatus is unnecessary. It will be tested on exactly the same animals and sections as the clonal analysis, and it is beaten if clone-level death fractions show statistically significant overdispersion beyond binomial expectation after correcting for cell type and tissue position. This costs nothing to run and must be reported first.
The second and strongest existing-method baseline is Deep Lineage — a deep-learning model that couples CRISPR barcodes to recurrent networks over clonal sister-cell sequences, reporting 77% fate-prediction accuracy on hematopoiesis, 94% day-28 outcome prediction in the fibroblast reprogramming dataset, and gene-imputation correlation of R = 0.81. It is the strongest competitor because it is the best published predictor of cell fate from barcode-plus-transcriptome, exactly our prediction target modulo death. We will run it unmodified on our dataset with the same cell counts, barcode depth, and tuning budget as our death-augmented model, and additionally on hematopoiesis as reported. Our model beats it if it predicts clone-level death fractions significantly better on our data while staying within a point or two of 77% on hematopoiesis — showing the death-hazard addition costs nothing on the standard benchmark. CoSpar, which recovered 73% of known reprogramming marker genes even with 90% of clonal data discarded, serves as the transition-matrix alternative run under the same regime.
The third baseline is the key ablation: our own model with the clonal barcode input removed, so it sees only transcriptomes. The precedent is Deep Lineage's own ablation, where dropping clone information collapsed accuracy from 77% to 52%. If our death-hazard model shows a comparable collapse in predicting clone death fractions without barcodes, the clonal ingredient is doing real work; if it does not, the proposal's central premise is wrong and we would report that.
Why not just…?
Why not just use standard Cre driver lines plus single-cell sequencing or RNA velocity? This is the simplest alternative and the most tempting. The literature gives three concrete reasons it falls short. First, the ablation in Deep Lineage shows that removing clonal information drops fate prediction from 77% to 52% — snapshot transcriptomics alone carries barely half the predictive signal. Second, as LCL: Contrastive Learning for Lineage Barcoded scRNA-seq Data demonstrates, unsupervised transcriptomic models cluster cells by transient confounders like cell-cycle phase rather than heritable fate programs, and those confounders are exactly what metabolic stress amplifies. Third, and decisive for this system: stressed cells downregulate the marker genes used to define them — the phenomenon documented in cardiac fibroblast tracing where cells extinguish their baseline markers during conversion — so a Gad2-driven readout literally cannot tell a dead GAD2 neuron from a stressed one that went silent. Only a permanent recombination mark survives that ambiguity. If, despite this, a marker-based analysis in Experiment 2 predicts death as well as the clonal one, that result would itself settle the question in favor of the simpler approach, and we would report it as such.
Why not just treat this as apoptosis and run standard caspase or TUNEL assays? Because the one study that has directly connected sleep loss to neuronal death found the ferroptotic, not apoptotic, mechanism: Mechanistic study on electroacupuncture-regulated circadian autophagy for inhibiting ferroptosis in hippocampal neurons shows sleep deprivation produces iron overload and lipid peroxidation in neurons, not caspase execution. The anatomical context reinforces this — the substantia nigra is a documented in vivo epicenter of iron accumulation in Parkinson's disease (Stability of AI-Enabled Diagnosis of Parkinson's Disease: A Study Targeting Substantia Nigra in Quantitative Susceptibility Mapping Imaging), and iron is the fuel for the Fenton chemistry that drives ferroptosis (Ferroptosis: mechanism, immunotherapy and role in ovarian cancer). Apoptosis assays would also miss the propagating-wave question entirely, since apoptosis is strictly cell-autonomous. That said, the alternative is a real possibility for some neurons in these nuclei, so Experiment 2 includes caspase-3 staining in parallel; if death loads on the apoptotic rather than peroxidative marker, the ferroptotic premise is falsified for those populations.
Why not just reconstruct lineage trees computationally from ordinary single-cell data, using metric learning? CellTreeQM does exactly this — a Transformer with a phylogenetic four-point loss that achieved 75.7% relative Robinson-Foulds improvement on C. elegans data. But its own authors report that the feature-gating component that powers that number has much milder effect on real biological data than in simulations, because biological noise is heavily correlated with signal, and that reconstruction degrades when latent distances become non-additive — precisely the regime a dying, stressed population creates. More fundamentally, any reconstruction from surviving cells can only infer the lineages of cells that lived; the dead are structurally invisible to it. Prospective barcoding remains the only way to know which branches vanished. We will still run CellTreeQM on our data as a cheap sanity check: if it recovers the same vulnerable clades our prospective barcodes identify, that would be a bonus finding, not a replacement.
Why not just use TreeVAE or a standard VAE and skip building a new model? TreeVAE already outperforms standard VAEs (latent MSE 1.89 versus 2.28; cross-entropy 21,348 to 3,051 with its phylogenetic message passing), and we reuse its architecture wholesale. What it cannot do — by its authors' own construction, since it infers ancestral states of observed leaves — is represent an absorbing death state for branches with no observed leaves. The minimal change we make is adding a per-branch hazard term; if that term fits no better than a constant on our data, we will conclude the death process carries no heritable signal and the simpler tools were sufficient. The honest position is that this is an incremental model change over TreeVAE, justified only by the death-augmented data that no existing benchmark provides.
Decisive Checks
Three outcomes would each be decisive. If clone-level death fractions overdisperse beyond binomial expectation, and clonal autocorrelation of death exceeds spatial autocorrelation, the clonal-competence claim is supported and the field inherits a new variable — developmental lineage — to add to marker identity when predicting neurodegeneration. If instead death is binomial across clones and spatially clustered, the wave model of synchronized ferroptosis extends from kidney epithelium (PAFAH2 suppresses synchronized ferroptosis) to the mammalian brain, which would reorient therapy toward non-cell-autonomous wave breakers rather than cell-intrinsic GPX4 rescue. And if pharmacological ferroptosis inhibition spares the vulnerable clones and preserves REM atonia in Experiment 4, the proposal closes the loop from mechanism back to the clinical phenotype — loss of REM paralysis as the earliest visible cost of iron-dependent death in the sleep-generating brain.
References
- Deep Lineage: Single-Cell Lineage Tracing and Fate Inference Using Deep Learning
- Reconstructing unobserved cellular states from paired single-cell lineage tracing and transcriptomics data
- Learning dynamics by computational integration of single cell genomic and lineage information
- LCL: Contrastive Learning for Lineage Barcoded scRNA-seq Data
- Author Correction: PAFAH2 suppresses synchronized ferroptosis to ameliorate acute kidney injury
- Actionable biological programs to enhance EGFR-targeted therapy response unveiled by single-cell lineage tracing in clinically relevant lung cancer models
- A common hub for sleep and motor control in the substantia nigra
- Ferroptosis: mechanism, immunotherapy and role in ovarian cancer
- Multifaceted mechanisms mediating cystine starvation-induced ferroptosis
- Silibinin Inhibits Cell Ferroptosis and Ferroptosis-Related Tissue Injuries
- Detection of REM Sleep Behaviour Disorder by Automated Polysomnography Analysis
- Calcium imaging of sleep–wake related neuronal activity in the dorsal pons
- Ferroptosis inhibitors: past, present and future
- MAPK14/SLC7A11/GPX4 axis dysregulation drives podocyte ferroptosis via mediating glycerophospholipid metabolism
- Mechanistic study on electroacupuncture-regulated circadian autophagy for inhibiting ferroptosis in hippocampal neurons and alleviating depression-like behaviors in adulthood induced by early chronic sleep deprivation
- Stability of AI-Enabled Diagnosis of Parkinson’s Disease: A Study Targeting Substantia Nigra in Quantitative Susceptibility Mapping Imaging
- Reconstructing Cell Lineage Trees from Phenotypic Features with Metric Learning