A Cross-Species Atlas of Pain-Ganglion Reprogramming under Multimodal Analgesia

After surgery, pain is increasingly managed not with a single opioid but with a layered combination: a local anesthetic that silences the injured nerve, a steroid such as dexamethasone that dampens inflammation, and non-opioid drugs like ketamine or dexmedetomidine that quiet central pain amplification. These multimodal regimens are validated almost entirely in rodents, yet a large fraction of preclinical pain findings fail in humans. One likely reason sits in the genome's control layer rather than its genes: the proteins that make up a nerve cell are nearly identical between mouse and human, but the DNA switches that turn them on — enhancers, promoters, and long non-coding RNAs — evolve fast. In the human brain, only about 30% of candidate regulatory switches are conserved in mouse (A comparative atlas of single-cell chromatin accessibility in the human brain). Separately, there is an untested metabolic story: stressed cells, from cancers to immune cells, switch from efficient mitochondrial energy production to fast aerobic glycolysis (the Warburg effect), and this switch both fuels and is fueled by inflammatory gene programs. Whether injured pain-sensing neurons do the same, and whether analgesic drugs reverse it, is simply unknown.

This proposal builds the missing triple intersection. I will construct a paired single-cell atlas of the dorsal root ganglion (DRG — the cluster of pain-sensing neurons and supporting glia just outside the spinal cord) in mouse, rat, and human, profiled under nerve injury and under clinically realistic multimodal drug regimens, reading out both gene expression and chromatin accessibility (which DNA switches are open) in the same cells. The atlas feeds a computational model that extends recent cross-species cell-integration methods to non-coding RNAs and chromatin, so it can predict which drug-induced changes in mouse DRG will carry over to human DRG. The falsifiable sub-hypothesis: the metabolic reprogramming module (protein-coding enzyme genes such as hexokinase 2 and pyruvate dehydrogenase kinases) is far more translatable across species than the non-coding inflammatory module (lncRNA hubs like Neat1), because one is carried by deeply conserved coding genes and the other by rapidly drifting regulatory elements. If true, it reframes analgesic target selection: aim at conserved metabolic gatekeepers, not at rodent-discovered lncRNA hubs.

Core Claim

The molecular response of sensory ganglia to multimodal analgesia is wired substantially by species-specific regulatory elements, and current cross-species translation methods — which work only on protein-coding genes — systematically overstate how much of that response will transfer to humans. By measuring drug-induced regulatory and metabolic change at single-cell resolution in two rodents and human tissue, and by building the first cross-species integration model that represents non-coding transcripts and open chromatin, we can quantify, per cell type and per gene module, the fraction of analgesic effect that is translatable — and validate the top predicted targets in vivo.

The logic rests on three known facts that no study has yet combined. First, the existing DRG non-coding maps come only from untreated injury: Transcriptome analysis of non-coding RNAs and mRNAs in the dorsal root ganglion of peripheral nerve injury-induced neuropathic pain catalogued 337 differentially expressed lncRNAs and 32 circular RNAs in a mouse nerve-injury model but tested no drugs, and Neat1 lncRNA organizes the inflammatory gene expressions in the dorsal root ganglion in neuropathic pain caused by nerve injury showed in rat that one lncRNA, Neat1, stabilizes pro-inflammatory chemokine transcripts and controls roughly 20% of all injury-upregulated genes — again with no analgesic perturbation and no metabolic measurement. Second, clinical multimodal trials measure only bulk serum markers: Jie Xu et al. tracked TNF-α and CRP, and the randomized dexmedetomidine–ketamine trial in liver resection (Opioid sparing anesthesia in patients with liver cirrhosis undergoing liver resection) reported roughly 40% intraoperative and 55% postoperative opioid reduction — but neither approach sees inside the DRG. Third, metabolic reprogramming is characterized in tumors (Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development; Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development) and hypoxic placenta (Jie He et al., showing hypoxia-driven HK2 upregulation), never in post-mitotic sensory neurons under drug exposure.

Research Program

The program has three parts: a drug-arm atlas, a non-coding-aware translation model, and targeted in vivo validation.

The animal atlas uses the two injury models the field already trusts, with drug arms transplanted directly from the clinical literature. Mouse bilateral chronic constriction injury (following Wanxia Xiong et al.) and rat spinal nerve ligation (following Motoyo Maruyama et al.) receive one of four perioperative arms: vehicle; perineural local anesthetic; a single dexamethasone dose (the regimen validated by Zhenhong Zou et al. and Can Wang et al.); the dexmedetomidine-plus-ketamine combination from the liver-resection trial ( Opioid sparing anesthesia in patients with liver cirrhosis undergoing liver resection: a controlled randomized double-blind study); and the full multimodal cocktail. DRG are harvested at days 3, 7, and 14 (matching the Neat1 kinetics Maruyama reported), with single-cell paired expression-plus-accessibility profiling per condition. Sample sizes follow the design guideline established by scPerturb: Harmonized Single-Cell Perturbation Data: at least 300 cells per perturbation and roughly 1,000 UMI counts per cell, below which perturbation signals drown in technical noise.

The human arm uses DRG and dorsal nerve-root tissue from organ donors and surgical nerve-root resections, profiled with the same paired multiome assay, plus an ex vivo inflammatory challenge with and without drug exposure (dexamethasone, ketamine, lidocaine). This is the weakest link — ex vivo human ganglia cannot fully model perioperative physiology — and the proposal treats it accordingly: human data serve as the prediction target, not the discovery substrate.

The computational model extends the "macrogene" idea. SATURN (Towards Universal Cell Embeddings) and Unify (Unify: Learning Cellular Evolution with Universal Multimodal Embeddings) overcome the failure of one-to-one gene matching by clustering genes into cross-species functional groups using protein embeddings; both are structurally blind to lncRNAs, circular RNAs, and ATAC peaks, which have no protein to embed — and both papers acknowledge this. Our extension adds two embedding channels: non-coding transcripts embedded by sequence and secondary-structure features plus cross-species co-expression with coding genes, and open chromatin peaks embedded by DNA k-mer sequence features scored with a mouse-trained classifier, the same conservation-scoring strategy used in the human brain atlas ( A comparative atlas of single-cell chromatin accessibility in the human brain). The model is trained on mouse and rat drug arms and asked to predict human responses; the metabolic module score (glycolysis, OXPHOS, hexosamine pathway genes, following the module definitions of Elevated glucose metabolism via the hexosamine biosynthesis pathway) and the non-coding inflammatory module score are tracked separately per cell type.

A central descriptive statistic is the conserved fraction of injury-opened human DRG enhancers:

$$\tau \;=\; \frac{\lvert \{\, e \in E^{\,\text{human}}_{\text{injury}} \;:\; f_{\text{mouse}}(e) > 0 \,\} \rvert}{\lvert E^{\,\text{human}}_{\text{injury}} \rvert}$$

where $E^{\text{human}}_{\text{injury}}$ is the set of chromatin peaks opened by injury in human DRG neurons and $f_{\text{mouse}}$ is a mouse-trained sequence-activity score, mirroring the gkmSVM setup of the brain atlas, which found $\tau \approx 0.30$ for brain. The form is simply an empirical fraction; the brain number is the prior against which the DRG value is interpreted. Given that neuronal subtypes show the most intense species-specific enhancer specialization ( A comparative atlas of single-cell chromatin accessibility in the human brain), we expect $\tau$ in nociceptors to be low — but measuring it is the point.

Validation closes the loop in rat, reusing the AAV6 sensory-neuron-targeted knockdown toolkit that Maruyama et al. proved out for Neat1. The two predicted-conserved regulators with the largest predicted effect on the metabolic module (candidates: the rodent orthologs of PDK isoforms and HK2, whose roles as metabolic gatekeepers were established in glioblastoma by Claire Larrieu et al. and in hypoxic tissue by Jie He et al.) are knocked down in the injury model, with mechanical withdrawal thresholds as the behavioral readout. If the metabolic-lock hypothesis is right, knocking down the glycolytic dam should relieve allodynia and collapse the non-coding inflammatory module downstream — the same logic by which Maruyama's Neat1 knockdown returned thresholds toward baseline.

Baselines

The naive reference point is the implicit practice of the field: assume the rodent response transfers unchanged to human by mapping one-to-one orthologs and copying the differential-expression result. Concretely this means taking the untreated-injury catalogs of Xiong et al. and Maruyama et al. as the expected human response. It is the strongest competitor in the sense that it costs nothing and is what preclinical decisions currently rest on; beating it means showing, on the human tissue, that a measurable fraction of the rodent program (non-orthologous lncRNAs, unconserved enhancers) is absent or inverted, and that the transfer model's predictions are closer to observed human responses than the copied rodent signature, measured per-module correlation on held-out human drug arms.

The strongest existing method is scGen (Generative modeling and latent space arithmetics predict single-cell perturbation response across cell types, studies and species), which predicts a perturbation's effect by adding a learned "difference vector" in a variational autoencoder's latent space and which reported R² values up to 0.97 between predicted and actual perturbed expression profiles, including cross-species phagocyte responses to LPS. It will be trained on the mouse drug arms and evaluated on exactly the same held-out human conditions, with the same cell numbers, the same tuning budget (matched GPU-hours and hyperparameter search), and the same evaluation code as our model. Beating scGen means two things: higher per-gene and per-module R² on multi-drug (cocktail) conditions specifically, and producing predictions for non-coding and chromatin features at all, where scGen by construction outputs nothing for transcripts without human–mouse orthologs.

The key ablation is our own architecture with the non-coding and ATAC channels removed, reducing it to a protein-coding macrogene model in the SATURN/Unify family. SATURN reported 85.8% median label-transfer accuracy across species, a 119% gain over the prior leading method; our ablation should match that level of cell-type alignment while failing to predict the non-coding module — that contrast is what isolates the contribution of the new channels. Unify reports outperforming scGen, SATURN, and Seurat v4 at phylogenetic distances up to 700 million years, but no reported integration metric values were found on Lacuna for Unify, so it is treated as a method to reimplement and match, not as a numeric benchmark.

Why not just…?

Why not just filter everything to one-to-one orthologs and use standard pipelines? This discards the very regulatory layer where translation fails. The Mouse ENCODE consortium showed roughly half the human genome cannot be aligned to mouse, concentrated in regulatory regions (Yue et al.), and several human disease mechanisms run entirely through lncRNAs with no rodent counterpart: NEXN-AS1 in atherosclerosis (Hu et al.), CYP1B1-AS1 in host defense (Arunima et al.), and MIR205HG acting through primate-specific Alu repeats (Takashima et al.). An ortholog-filtered analysis of our atlas would answer a different, easier question — and the p53/miR-138 case, where regulatory polarity inverts outright between rodents and humans (Li et al.), shows it can even give the wrong sign.

Why not just use bulk RNA-seq and serum biomarkers, as the clinical trials already do? Bulk averaging is precisely the failure mode the proposal attacks. Neat1 is expressed in both sensory neurons and satellite glia but acts differently in each, resolvable only by cell-type-targeted tools (Maruyama et al.); bulk DRG sequencing would blend them with macrophages and endothelium. Serum TNF-α and CRP (Jie Xu et al.) report systemic inflammation, not the intracellular metabolic or chromatin state of the ganglion. And the brain atlas showed that the most species-divergent regulatory specialization lives in neuronal subtypes, which bulk profiling averages away ( A comparative atlas of single-cell chromatin accessibility in the human brain). Single-cell multi-omics costs more, but it is the only readout that can see the object of study.

Why not just use scGen's latent arithmetic for the drug-response prediction? This is the most serious alternative, and for single-agent perturbations it may well win — its R² of up to 0.97 on cross-species immune responses is formidable ( Generative modeling and latent space arithmetics predict single-cell perturbation response across cell types, studies and species). The proposal does not argue around this; it tests it directly. The comparison is run separately on single-drug arms and cocktail arms, because two results in the file predict linearity breaks exactly where multimodal analgesia lives: regulatory topologies can invert between species (Li et al.), which a single difference vector cannot represent, and cis-regulatory grammar is combinatorial and non-linear (Nvwa). If scGen matches our model on cocktail conditions, that is a genuine finding against the proposal, and it will be reported as such.

Why not just target one downstream cytokine, say CCL2, with an inhibitor? Redundancy and upstream control. Injury simultaneously activates hundreds of pro-inflammatory transcripts, and Maruyama's Neat1 work shows one upstream lncRNA stabilizes roughly a fifth of them, including Ccl2 and Cxcl9 (Neat1 lncRNA organizes the inflammatory gene expressions in the dorsal root ganglion in neuropathic pain caused by nerve injury Neat1 lncRNA organizes the inflammatory gene expressions in the dorsal root ganglion in neuropathic pain caused by nerve injury) — picking off one target leaves the rest. More fundamentally, the metabolic hypothesis says the bioenergetic state sustains the firing and glial activation; blocking a cytokine does not reverse a glycolytic lock imposed by gatekeepers like PDHK (Larrieu et al.). Our validation experiment — knocking down the metabolic dam and watching both behavior and the inflammatory module — is designed to test whether the upstream story is the right one.

Decisive Checks

Four results decide the proposal. First, the value of $\tau$ for injury-opened human DRG enhancers, benchmarked against the 30% brain figure: a value near 0.30 confirms deep regulatory divergence in sensory ganglia; a value near 0.80 would falsify the premise. Second, the held-out prediction comparison: our model must beat scGen (R² up to 0.97 on its home turf, single-agent immune perturbations) on cocktail arms while scGen retains the crown on single-drug arms if linearity holds there — any other pattern is informative. Third, the module split: the metabolic module should transfer across species with substantially higher fidelity than the non-coding inflammatory module; if they transfer equally, the "target metabolic gatekeepers, not lncRNA hubs" recommendation collapses. Fourth, the rat knockdown of the predicted conserved metabolic regulators must shift both mechanical withdrawal thresholds and the downstream non-coding module in the predicted direction, with effect sizes comparable to the Neat1 knockdown that returned thresholds toward baseline. Pairwise fields — rodent DRG non-coding catalogs, clinical biomarker trials, cross-species protein-only embeddings — each supply one leg of this design; none of them, alone or in pairs, can deliver the translatable-fraction number that analgesic target selection actually needs.

References

  1. A comparative atlas of single-cell chromatin accessibility in the human brain
  2. Transcriptome Analysis of Non-coding RNAs and mRNAs in the Dorsal Root Ganglion of Peripheral Nerve Injury-Induced Neuropathic Pain
  3. Neat1 lncRNA organizes the inflammatory gene expressions in the dorsal root ganglion in neuropathic pain caused by nerve injury
  4. Analgesic Effect of Combined Spinal-Epidural Anesthesia and its Effect on TNF-α and CRP Levels in Elderly Patients with Hip Fracture During Surgical Treatment
  5. Opioid sparing anesthesia in patients with liver cirrhosis undergoing liver resection: a controlled randomized double-blind study
  6. Refining the Role of Pyruvate Dehydrogenase Kinases in Glioblastoma Development
  7. Absolute dynamic and relative static: the relationship of glycolysis and OXPHOS in cancer development
  8. Elevated expression of glycolytic genes as a prominent feature of early-onset preeclampsia: insights from integrative transcriptomic analysis
  9. The Impact of Prophylactic Dexamethasone on Nausea and Vomiting after Thyroidectomy: A Systematic Review and Meta-Analysis
  10. A systematic review and meta-analysis on how different dexamethasone administration regimes impact total joint arthroplasty outcomes
  11. scPerturb: Harmonized Single-Cell Perturbation Data
  12. Towards Universal Cell Embeddings: Integrating Single-cell RNA-seq Datasets across Species with SATURN
  13. Unify: Learning Cellular Evolution with Universal Multimodal Embeddings
  14. Elevated glucose metabolism via the hexosamine biosynthesis pathway: A metabolic signature of high-fluorodeoxyglucose-uptake lung adenocarcinoma
  15. Generative modeling and latent space arithmetics predict single-cell perturbation response across cell types, studies and species
  16. A comparative encyclopedia of DNA elements in the mouse genome
  17. Long noncoding RNA NEXN-AS1 mitigates atherosclerosis by regulating the actin-binding protein NEXN
  18. CYP1B1-AS1 regulates CYP1B1 to promote Coxiella burnetii pathogenesis by inhibiting ROS and host cell death
  19. Involvement of lncRNA MIR205HG in idiopathic pulmonary fibrosis and IL-33 regulation via Alu elements
  20. Species-specific mutual regulation of p53 and miR-138 between human, rat and mouse
  21. Generative modeling and latent space arithmetics predict single-cell perturbation response across cell types, studies and species
  22. Inferring predictive genetic models and regulatory elements by deep learning of cross-species single-cell gene expression landscapes
  23. Neat1 lncRNA organizes the inflammatory gene expressions in the dorsal root ganglion in neuropathic pain caused by nerve injury