An Algebraic Stability Budget for Correlates of Protection

When a vaccine works, regulators would like to license updated formulations without running a new 30,000-person trial. The accepted substitute is a correlate of protection: an immune measurement, usually an antibody titer, that statistically stands in for clinical protection, so that a lab assay can replace the endpoint Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection. The trouble is that single titers fail exactly where they are most needed — for mutable pathogens like HIV, where RV144 protection tracked non-neutralizing V1V2 antibodies rather than the neutralization titers everyone expected Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial documents the same fragility across assays. The informative signal lives in high-dimensional multi-omic profiles Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine, and the standard way to handle such profiles is an embedding — a map from thousands of measured features down to a few dozen coordinates. But embeddings built by stochastic algorithms jitter: re-run the same method on the same data and 77–90% of the resulting configurations are statistically distinct On the Surprising Behaviour of \texttt{node2vec}. Nobody currently models that jitter when using embeddings to estimate protection, which means the uncertainty quoted on protection estimates is wrong in an unquantified way.

This proposal builds a correlate-of-protection pipeline that treats embedding instability as a first-class, measurable error source, and it borrows the tool for deciding which latent directions survive from an unexpected place: condensed matter theory. There, when conserved quantities fail to commute, the non-commutation pairs up and gaps out low-energy modes, so the number of surviving gapless modes equals the naive count minus the rank of the commutator matrix Counting Nambu-Goldstone Modes of Higher-Form Global Symmetries. We import that counting logic — not the physics — as a data-driven rule for how many coordinates of a stochastic embedding carry reproducible signal versus irreducible run-to-run noise, and we feed the surviving coordinates into a hierarchical Bayesian model that partial-pools protection estimates across patient subgroups.

Core Claim

The claim is that the number of trustworthy directions in a stochastic embedding can be estimated algebraically from an ensemble of runs, via a commutator-rank correction, and that using only those directions inside a hierarchical Bayesian protection model yields calibrated credible intervals on protection — intervals that survive contact with regulatory-grade data — where three current approaches each fail in a different way: univariate titer models ignore the multi-marker geometry, aligned embeddings give point estimates with no uncertainty at all, and the strongest existing embedding-plus-hierarchy method treats the embedding as noise-free.

The Mechanism: A Stability Budget from Commutator Rank

Hidaka, Hirono, and Yokokura show that in systems with generalized symmetries, the ground-state expectation of commutators of conserved charges, written as a matrix $M_{\alpha\beta}$, canonically pairs symmetry generators and gaps half of the paired modes, giving $N_{\text{NG}} = \sum_A (D - 2p_A) - \operatorname{rank} M_{\alpha\beta}$ Counting Nambu-Goldstone Modes of Higher-Form Global Symmetries. The structural content is a rank correction: constraints that cannot be simultaneously satisfied remove degrees of freedom, and the rank of the obstruction matrix counts exactly how many. Hirono and colleagues extend this to charges that fail to commute with spatial translations, where an inverse-Higgs effect eliminates redundant modes altogether A symmetry principle for gauge theories with fractons.

Our translation is as follows. Run an embedding algorithm $R$ times on the same immune-profile matrix. For each run, estimate the drift generator $G_\alpha$ — the linear map that best carries that run's coordinates onto the ensemble consensus, expressed in a learned basis indexed by $\alpha$. Two drift directions are irreconcilable when their generators do not commute: removing drift in direction $\alpha$ necessarily injects drift in direction $\beta$. We estimate the run-averaged commutator matrix $\hat{M}_{\alpha\beta} = \mathbb{E}_{\text{runs}}\,[G_\alpha, G_\beta]$, measured by its action on consensus coordinates, and declare a stability budget

$$N_{\text{stable}} = d - \operatorname{rank} \hat{M},$$

where $d$ is the embedding dimensionality. Directions in the rank-$\hat{M}$ obstruction subspace are "gapped": they are excluded from the downstream model and their variance is attributed to algorithmic noise. This replaces the heuristic alignment weight $\alpha$ in 2-MAP: Aligned Visualizations for Comparison of High-Dimensional Point Sets, which its own authors flag as risking under- or over-constraining, and the hand-tuned temporal penalty in Application of Aligned-UMAP to longitudinal biomedical studies. Crucially, the gapping decision is made from data, per dataset, rather than imposed as a fixed prior.

Research Program

Three components. First, the estimator: the drift-generator ensemble, the commutator matrix, and the rank selection, validated on synthetic data where the signal subspace is known. Second, the downstream model: a hierarchical Bayesian protection analysis in which patient-level protection parameters are partially pooled — noisy individuals shrunk toward population means, data-rich individuals allowed to deviate — across subgroups defined by position along the stable coordinates only, with the gapped subspace's variance entering as an explicit measurement-noise term. This directly extends Correcting Prompt Dependence in LLM Benchmarks: A Bayesian Hierarchical Model with Embedding-Space Clustering, which builds a hierarchy over embedding clusters but, as its authors concede, treats those clusters as fixed ground truth and assumes within-cluster homogeneity. We differ on both counts: instability is propagated, not frozen, and pooling operates on continuous stable coordinates rather than discrete clusters, which matters because immune protection is a continuous gradient, not a set of discrete boxes Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection. The mechanistic-population tradition of Modeles mathematiques dynamiques pour la medecine personnalisee supplies the variance-partitioning template — separating within-subject assay noise from between-subject heterogeneity — which we apply to the embedding layer that their explicit ODE systems never had. Third, the honesty check: an ablation that replaces commutator-rank gapping with simple variance-threshold truncation (drop the highest-variance principal components). If the two match, the algebraic ingredient is decorative and we will say so.

Baselines

The naive reference point is the univariate normalized-titer Cox model: normalize each assay against a reference cohort mean as in Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection and fit a proportional-hazards model on the single titer. This is the de facto regulatory workhorse, and it is the strongest competitor whenever one marker carries most of the signal: on PREVENT-19, Fong et al. report a hazard ratio of 0.13 (95% CI 0.03, 0.57) per 10-fold increase in Day 35 Delta nAb-ID50 for severe COVID-19 Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine. It will be run on exactly our data splits with the same event definitions. The strongest existing method at the embedding-plus-hierarchy intersection is BHM-ESC Correcting Prompt Dependence in LLM Benchmarks: A Bayesian Hierarchical Model with Embedding-Space Clustering, which reports 4–73% mean-absolute-error reduction and expected-log-posterior-density improvements of 20 to 450 units over frequentist baselines on LLM benchmarks (MMLU, GSM8K, and others); no reported result found on Lacuna for BHM-ESC on any immune or clinical dataset, so we re-implement it on our immune embeddings with the same sampler, the same compute budget, and the same number of posterior draws. Beating it means higher expected log posterior density on held-out patients and, more importantly, 95% credible intervals on hazard ratios that achieve nominal coverage under bootstrap resampling of embedding runs, which BHM-ESC by construction cannot do because it conditions on one embedding. The key ablation is our own model with the drift ensemble removed — a single-run embedding, no commutator matrix, no propagated variance — which isolates whether the calibration gains come from modeling instability rather than from the hierarchy alone.

Why not just…?

Why not just align the runs with Procrustes rotation? Because rigid alignment provably does not reach the problem. Hacker and Rieck show that after alignment considerations, 77–90% of pairwise embedding comparisons across hyperparameter settings remain statistically distinct On the Surprising Behaviour of \texttt{node2vec}, and the TRACE diagnostic suite documents local neighborhood tearing that no global rotation can repair Pattern or Artifact? Interactively Exploring Embedding Quality with TRACE. Our synthetic experiment includes a Procrustes-aligned baseline precisely to quantify the residual, and we expect it to fail the coverage test.

Why not just normalize titers to a reference pool and read off a protection curve? For ancestral SARS-CoV-2 this genuinely works — Khoury et al. show a candidate at 0.44-fold of the Moderna geometric mean titer achieves over 80% efficacy Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection — and we treat this as a real threat, not a strawman. But the same literature shows the approach collapses where markers diverge: in RV144, neutralization titers showed no correlation with protection while V1V2-binding antibodies carried the signal Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine, and multi-omic signatures resist any single-axis normalization Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial. The decisive experiment is head-to-head on the PREVENT-19 severe endpoint with variant-challenged markers: if the univariate curve matches our model's predictive density there, our claim should be rejected.

Why not just use a deterministic embedding algorithm? SURREAL replaces random walks with network conductance Deterministic Graph Embeddings and Tree Preserving Embedding imposes topological tree constraints Tree Preserving Embedding. These remove run-to-run jitter but impose rigid geometric priors — a tree topology misrepresents looping or multi-pathway immune trajectories — and, more fundamentally, they confound algorithmic reproducibility with measurement noise: identical coordinates do not mean correct ones. Our experiment compares deterministic baselines on the synthetic ground-truth task, where we can check whether their imposed topology distorts the known signal subspace.

Why not just fit a Cox model on all measured markers directly? Because the interesting endpoints are sparse: PREVENT-19 had only 8 severe cases among 287 breakthrough infections Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine, and frequentist estimation without partial pooling is exactly the regime where Correcting Prompt Dependence in LLM Benchmarks: A Bayesian Hierarchical Model with Embedding-Space Clustering showed naive analyses inflate effective sample size by 1.3x to 5.6x. A bootstrap simulation on the COVE data will measure how badly the direct Cox model's confidence intervals undercover in the sparse regime.

Proposed Experiments

The program runs four studies. First, a synthetic stress test on Stochastic Block Model graphs, where naive node2vec link reconstruction collapses to AUPRC near 0.09 against a chance level around 0.09 On the Surprising Behaviour of \texttt{node2vec}: we plant a known signal subspace, inject controlled drift, and test whether the commutator rank recovers it and whether gapped coordinates improve downstream link prediction over Procrustes-aligned and ungapped baselines. Second, the COVE trial data with its multi-assay panel (binding antibody, pseudovirus neutralization cID50, live-virus LV-MN50), where Gilbert et al. report a Day 57 cID50 hazard ratio of 0.42 per 10-fold increase and 68.5% of vaccine efficacy mediated by Day 29 cID50 Immune Correlates Analysis of the mRNA-1273 COVID-19 Vaccine Efficacy Trial: our target is to reproduce that point estimate within credible intervals while adding what no univariate analysis provides — a multi-marker stable-coordinate model whose intervals cover under run-resampling. Third, the PREVENT-19 severe endpoint, the sparse-event regime where partial pooling must earn its keep, benchmarked against Fong et al.'s reported hazard ratio of 0.13 (95% CI 0.03, 0.57) per 10-fold Delta nAb-ID50 increase and their cumulative-incidence drop from 0.0338 at the 2.5th titer percentile to 0.0002 at the 95th Analysis of Antibody Markers as Immune Correlates of Risk of Severe COVID-19 in the PREVENT-19 Efficacy Trial of the NVX-CoV2373 Recombinant Protein Vaccine. Fourth, the decisive calibration check: simulate cohorts with known protection signatures and known drift structure, and ask whether 95% credible intervals on the hazard ratio achieve nominal coverage — the single number that determines whether any of this is usable for immunobridging. The stochastic-interventional machinery of Stochastic Interventional Vaccine Efficacy and Principal Surrogate Analyses of Antibody Markers as Correlates of Protection against Symptomatic COVID-19 in the COVE mRNA-1273 Trial, which reports vaccine efficacy ranging from 84.2% to 97.6% under a one-log shift in Day 57 Spike IgG, provides the endpoint language in which our interval-calibrated results would be reported.

The honest risk register has two entries. If variance-threshold truncation matches commutator-rank gapping on the synthetic task, the algebraic contribution reduces to a reordering of a simpler idea and the paper becomes a study of instability propagation in hierarchical protection models — still worthwhile, smaller. And if the univariate Khoury curve matches us on every real dataset, the multi-marker geometry is not yet buying protection-relevant signal, and the proposal should stop at the calibration result. Both outcomes are publishable; neither is hidden.