Cross-Species Priors for Personalizing Vagal Anti-Inflammatory Stimulation

The vagus nerve — the long cranial nerve connecting the brainstem to the heart, gut, and immune organs — carries a built-in brake on inflammation. When stimulated, its efferent fibers release acetylcholine onto alpha-7 nicotinic receptors on macrophages, which suppresses the release of the pro-inflammatory cytokines TNF-α, IL-1β, and IL-6; this circuit is called the cholinergic anti-inflammatory pathway, and it is the reason vagus nerve stimulation is being pursued as a drug-free treatment for inflammatory disease Mechanism and Applications of Vagus Nerve Stimulation, Bioelectronic Medicine: a multidisciplinary roadmap from biophysics to precision therapies. The problem is not whether the pathway exists; it is that nobody can set the stimulator correctly for a given patient. Standard clinical stimulation reduces seizures by about half in only half of patients, and parameter search by trial and error takes months Vagus nerve stimulation: Laying the groundwork for predictive network-based computer models. Meanwhile the data that could guide calibration sits in the wrong species: rodents give us clean cytokine time courses, swine give us realistic nerve anatomy and electrical recordings, and humans give us almost nothing except scattered single-patient reports whose outcomes cannot be distinguished from individual noise Acupuncture for bevacizumab-related paralysis and changes in peripheral lymphocyte cells in advanced lung adenocarcinoma, Vertebral artery dolichoectasia leading to hemifacial spasm.

I propose to build a hierarchical, two-timescale model of the cholinergic anti-inflammatory pathway that is pre-trained across species and personalized to an individual human with a handful of stimulation probes. The core move is to treat cross-species divergence not as an obstacle but as structure: shared molecular readouts (the cytokines themselves) act as anchors that align rodent, swine, and human response curves into one latent space, in direct analogy to how orthologous genes anchor cross-species protein-network alignment in ETNA Joint embedding of biological networks for cross-species functional alignment. A single patient's few observations then become a posterior update on a cross-species prior rather than an ungeneralizable anecdote — the single-case limitation is addressed by design, not by wishing for larger cohorts.

Core Claim

A model that jointly captures fast electrophysiology (milliseconds to minutes: which nerve fibers a stimulus recruits) and slow immune kinetics (hours: how that recruitment suppresses cytokines), pre-trained on rodent and swine data and aligned across species through shared cytokine anchors, can predict an individual human's anti-inflammatory response to vagus stimulation from roughly two to five probe stimuli, with calibrated uncertainty. No existing work tests this. The closest prior, Mao et al.'s meta-learning framework for vagus nerve stimulation Rapidly Inferring Personalized Neurostimulation Parameters with Meta-Learning: A Case Study of Individualized Fiber Recruitment in Vagus Nerve Stimulation, pre-trains and personalizes within a single species (acute swine) and predicts only cardiorespiratory biomarkers; it never models cytokines and never transfers across species. Ingham et al.'s connectome model Vagus nerve stimulation: Laying the groundwork for predictive network-based computer models is explicitly human-only and explicitly excludes the anti-inflammatory axis; the authors state it captures nothing beyond minutes. The gap between these two — few-shot personalization that actually reaches the inflammatory endpoint, and that survives the species jump — is what this proposal fills.

Research Program

The model has two coupled layers. The fast layer $f_\theta$ maps a stimulation waveform $u_i(t)$ (current, frequency, pulse width) to an efferent drive $r_i(t)$, calibrated against evoked compound action potentials and heart-rate modulation as in the swine benchmarks of Mao et al. The slow layer $h_\phi$ maps that drive to cytokine trajectories. Per-subject parameters are drawn from a species-level prior that is a learned reparameterization of one shared prior:

$$c_i(t+\Delta) \;=\; c_i(t) + \Delta \, h_\phi\big(f_{\theta_i}(u_i(t)),\, c_i(t)\big), \qquad \theta_i = A_{s(i)}\,\theta_{\text{prior}} + \epsilon_i$$

Here $c_i(t)$ is the cytokine vector (TNF-α, IL-1β, IL-6, IL-10) for subject $i$ of species $s(i)$, $A_{s(i)}$ is a small species-alignment matrix, and $\epsilon_i$ is the individual deviation inferred by few-shot Bayesian optimization — the same machinery Mao et al. showed can pin down a subject's fiber-recruitment curve from as few as two stimulation queries. Training proceeds in three stages. First, the slow layer is fit to rodent cytokine time courses: the cecal-ligation-and-puncture sepsis model with its human methylation cohort Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway and the dextran-sulfate colitis model with its dose-dependent TNF-α, IL-6, IL-17 suppression and IL-10 boost Paricalcitol Ameliorated Dextran-Sulfate-Sodium-Induced Colitis in Mice through MKP-1/P38 MAPK Signaling Pathway — these matter because they constrain the shape and dose-response of the immune layer, including the narrow therapeutic window (Decitabine at 1 mg/kg improved survival with $p = 0.0081$ while 1.5 mg/kg added no anti-inflammatory benefit). Second, the fast layer and the coupling are fit to the acute swine vagus-stimulation recordings with eCAP and cardiorespiratory readouts Rapidly Inferring Personalized Neurostimulation Parameters with Meta-Learning: A Case Study of Individualized Fiber Recruitment in Vagus Nerve Stimulation. Third, the alignment matrices $A_s$ are trained so that rodent, swine, and — where data exists — human cytokine response curves occupy a shared latent space, using the cytokines as anchors the way ETNA uses orthologs; ETNA's own result that topology-based alignment beats sequence homology (AUPRC 0.852 vs 0.775 for MUNK on human–yeast networks) motivates aligning on response-curve geometry rather than on raw anatomical homology, which is exactly where cross-species translation fails Joint embedding of biological networks for cross-species functional alignment. The anatomical scaling problem — sub-millimeter murine vagal trunks versus multi-fascicular human nerves Novel adhesive sponge for vascular closure, Evaluating rapid hemostatic hydrogels in porcine puncture wounds — enters through the fast layer's species-dependent input geometry, learned from the swine recordings rather than assumed transferable.

Proposed Experiments

Experiment 1 tests cross-species transfer on the fast layer, using the real VNS benchmark (CBI-C) from Mao et al. plus their synthetic multivariate benchmarks. Pre-train the population prior on rodent-shaped synthetic data and swine recordings, then personalize with 2 and 5 shots on held-out swine subjects, and report relative RMSE for predicting held-out stimulation responses. The decisive comparison is against the same architecture trained on swine alone at matched compute and tuning budget: if cross-species pre-training does not reduce the number of shots needed or the error at fixed shots, the alignment hypothesis fails. Experiment 2 tests the slow layer by retrodiction: can the model, given only the stimulation or dosing protocol, reproduce the reported cytokine trajectories and dose-response ceilings in the sepsis and colitis datasets? Success is calibrated prediction of cytokine levels over 8 hours to 8 days, with 90% credible intervals achieving roughly 90% empirical coverage on held-out animals. Experiment 3 tests the human leap on the only human inflammatory data in reach: the ESNV-SPA crossover trial of non-invasive auricular vagus stimulation in biologic-resistant spondyloarthritis Randomized Cross Over Study Assessing the Efficacy of Non-invasive Stimulation of the Vagus Nerve in Patients With Axial Spondyloarthritis Resistant to Biotherapies: The ESNV-SPA Study Protocol, supplemented by retrospective single-case series with longitudinal lymphocyte and cytokine measurements such as the acupuncture chemotherapy-toxicity case Acupuncture for bevacizumab-related paralysis and changes in peripheral lymphocyte cells in advanced lung adenocarcinoma. I must be honest: no reported result found on Lacuna for a human vagus-stimulation cytokine dose–response dataset, so Experiment 3 is a pre-registered prediction exercise — the model predicts individual responders' cytokine trajectories before unblinding, and its interval coverage is scored afterward. Experiment 4 is a simulation check of the fast-slow coupling: driving the fast layer into Ingham et al.'s 22-population neural mass model, where NTS input above −0.52 shifts the network into a monostable, seizure-free state, and confirming that our slow layer's predicted cytokine response to that same drive is consistent with the rodent retrodictions — a first, crude bridge across the millisecond-versus-hours timescale tension that Ingham et al. explicitly leave open.

Baselines

The naive reference point is the unpersonalized population model: a single response curve fit to all training subjects and applied unchanged to a new one, which is the statistical content of "one-size-fits-all" stimulation protocols. Its failure is documented — standard clinical stimulation leaves half of patients unhelped Vagus nerve stimulation: Laying the groundwork for predictive network-based computer models, and on the CBI-C benchmark with 2 shots the non-adapted-style baselines in Mao et al.'s evaluation, such as MAML at RMSE 47.97±3.15, sit far above the adapted methods. The strongest existing competitor is Mao et al.'s metaRG-gu — a Gaussian-process population prior with Bayesian-optimization few-shot rescaling, which achieved RMSE 31.20±2.85 on CBI-C with 2 shots Rapidly Inferring Personalized Neurostimulation Parameters with Meta-Learning: A Case Study of Individualized Fiber Recruitment in Vagus Nerve Stimulation. It will be run on identical data, compute, and tuning budget, with its shot counts (2 and 5) matched exactly; beating it means lower relative RMSE at equal shots, or equal RMSE with fewer shots, on held-out subjects of the target species — and, critically, on the cytokine endpoint it was never designed to predict. The key ablation is the same architecture with the cross-species prior replaced by a within-species prior (swine-only pre-training for swine targets): this isolates whether the alignment machinery, rather than merely extra parameters, produces the gain. A second ablation removes the slow layer and instead maps fast-layer outputs to cytokines through a separately fit static function, isolating whether the coupled two-timescale dynamics are necessary or a simple readout map would do.

Why not just…?

Why not just use Mao's framework on human data directly? Because the human training cohort it requires does not exist; that framework needs many historical subjects of the same species to build its population prior, and the whole premise of the single-case problem is that human observations arrive one at a time A RESECTED CASE OF ADVANCED ESOPHAGEAL CARCINOMA SHOWING A PROMINENT RESPONSE TO NEOADJUVANT CHEMOTHERAPY WITH 5-FU AND LEUCOVORIN. If abundant human data appeared, this alternative would be a genuine threat, and the within-species ablation is designed to settle it: at matched data budget, cross-species pre-training must beat within-species pre-training or the proposal's central claim is wrong.

Why not just sweep parameters in each patient until something works? Because in-vivo sweeps are slow and risky — off-target recruitment causes bradycardia, cough, and pain — and the quantitative evidence is against them: single-task fitting from few or biased queries was consistently the worst performer in Mao et al.'s benchmarks, while biophysically constrained calibration achieved 90.6% prediction accuracy from minutes of recording rather than hours of probing Learning Biophysical Models of Large-Scale Multineuronal Data To Enable Precise Neurostimulation, supported by the released fitting code amrith1/rgc-ei-fit-hh. Our shot-count metric directly prices this trade-off.

Why not just use the meta-analysis prescription? Zhang et al.'s mining of 42 trials found that stimulating the ST36 and ST37 acupoints activates the vagal-adrenal axis and speeds postoperative gastrointestinal recovery — time to first defecation reduced by 20.93 hours on average, ST36 appearing in 93% of effective prescriptions The lower He-sea points playing a significant role in postoperative ileus in colorectal cancer treated with acupuncture: based on machine-learning. But the same paper reports heterogeneity above $I^2 = 90\%$, which means the population average describes almost no individual; a fixed prescription inherits exactly the over-generalization this proposal is built to escape, and Experiment 3's per-subject coverage scores will show whether individualized prediction beats the group rule.

Why not just use drugs? Systemic immunosuppressants carry off-target toxicities Bioelectronic Medicine: a multidisciplinary roadmap from biophysics to precision therapies, and the target population — patients who have already failed biologics, as in the ESNV-SPA trial Randomized Cross Over Study Assessing the Efficacy of Non-invasive Stimulation of the Vagus Nerve in Patients With Axial Spondyloarthritis Resistant to Biotherapies: The ESNV-SPA Study Protocol — has exhausted that option by construction.

Decisive Checks

Three results would make or break the claim. First, the within-species ablation: if cross-species pre-training does not reduce shots or error relative to swine-only pre-training, the anchoring idea adds nothing and the proposal reduces to re-running Mao et al. with a cytokine head — still useful, but a smaller paper. Second, calibration on held-out species: if the 90% credible intervals on human cytokine trajectories in Experiment 3 systematically miss, the alignment is absorbing species differences into false confidence, which is worse than no model. Third, the coupling ablation: if the static readout map matches the full two-timescale dynamics, then Ingham et al.'s timescale tension is a non-problem for prediction purposes and the slow layer is unnecessary machinery. Each of these is a pre-registered failure mode, and any of them, if triggered, tells us exactly which component to discard.

References

  1. Mechanism and Applications of Vagus Nerve Stimulation
  2. Bioelectronic Medicine: a multidisciplinary roadmap from biophysics to precision therapies
  3. Vagus nerve stimulation: Laying the groundwork for predictive network-based computer models
  4. Acupuncture for bevacizumab-related paralysis and changes in peripheral lymphocyte cells in advanced lung adenocarcinoma
  5. Vertebral artery dolichoectasia leading to hemifacial spasm
  6. Joint embedding of biological networks for cross-species functional alignment
  7. Rapidly Inferring Personalized Neurostimulation Parameters with Meta-Learning: A Case Study of Individualized Fiber Recruitment in Vagus Nerve Stimulation
  8. Inhibiting DNA Methylation Improves Survival in Severe Sepsis by Regulating NF-κB Pathway
  9. Paricalcitol Ameliorated Dextran-Sulfate-Sodium-Induced Colitis in Mice through MKP-1/P38 MAPK Signaling Pathway
  10. A Single‐Component Adhesive Sponge Based on Blood‐Triggered and Autopenetrative Adhesion for Robust Vascular Closure
  11. Lyophilized apoptotic vesicle-encapsulated adhesive hydrogel sponge as a rapid hemostat for traumatic hemorrhage in coagulopathy
  12. Randomized Cross Over Study Assessing the Efficacy of Non-invasive Stimulation of the Vagus Nerve in Patients With Axial Spondyloarthritis Resistant to Biotherapies: The ESNV-SPA Study Protocol
  13. A RESECTED CASE OF ADVANCED ESOPHAGEAL CARCINOMA SHOWING A PROMINENT RESPONSE TO NEOADJUVANT CHEMOTHERAPY WITH 5-FU AND LEUCOVORIN
  14. Learning Biophysical Models of Large-Scale Multineuronal Data To Enable Precise Neurostimulation
  15. The lower He-sea points playing a significant role in postoperative ileus in colorectal cancer treated with acupuncture: based on machine-learning