Two unrelated fields have stumbled into the same failure. In Earth-system modeling, soil carbon is simulated with first-order decay kinetics driven only by pool size, temperature, and moisture, while the actual controllers of decomposition — microbial biomass, mineral stabilization, nutrient coupling — go unmeasured and unmodeled; the result is that three structurally distinct land models all predict England and Wales gained soil carbon between 1980 and 1995 (+1.7 to +8.1 g C m⁻² yr⁻¹) when the National Soil Inventory measured a loss of −31.4 g C m⁻² yr⁻¹ (Can we model observed soil carbon changes from a dense inventory?). In vaccine immunology, clinical protection is assessed almost entirely through one measurable quantity — the serum neutralizing antibody (nAb) titer, the lab readout of the "humoral" arm of immunity that blocks infection at the mucosal surface — while the "cellular" arm (memory T cells that recognize conserved viral fragments and kill infected cells) is rarely measured because its assays are slow and cell-hungry. The consequence is structurally identical: in the 44,325-participant ENSEMBLE trial, participants whose nAb titers were entirely unquantifiable nonetheless retained 63.1% vaccine efficacy against severe-critical COVID-19, and mediation analysis attributed only 28.6% of total protection to antibodies (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). Roughly seventy percent of the protection signal is, today, an unexplained residual.
This proposal treats those two situations as the same estimation problem: a two-component dynamical system in which one component is observed and fast-decaying, the other is latent and slow-decaying, and standard practice is to fit a single-marker model that silently absorbs the second component into error. The plan is to build one Bayesian two-compartment estimator, prove it first on the soil twin — where laboratory incubations give partial ground truth for the latent microbial component — and then deploy it on vaccine trial data to produce the first identified, individual-level estimates of cellular protection from trials that never measured T cells.
Core Claim
A coupled stochastic state-space model, fitted by the multi-stream Bayesian assimilation machinery that was proven for inferring unobserved soil carbon pools, can jointly explain signature results that single-marker models cannot, on both sides of the isomorphism. The key testable predictions are that such a model will (a) recover field priming intensity in soils that the standard cheap approximation underestimates by roughly half, and (b) simultaneously reproduce the ENSEMBLE finding that severe-disease protection persists at zero measurable antibody, the COVAIL finding that antibodies predict much less in infection-naïve than in previously infected participants, and the ENSEMBLE sieve finding that severe-disease protection tracks sequence divergence in conserved viral regions rather than in the antibody-facing spike regions.
The Shared Structure
The parallel is precise enough to be exploited rather than admired. On the soil side, the fast arm is fresh labile carbon entering litter, which microbes use to accelerate the breakdown of old soil carbon — the "priming" effect, where added carbon causes net carbon loss; the slow arm is mineral-associated carbon stabilized on particle surfaces. The quasi-equilibrium trick used in the ORCHIDEE PRIM module replaces an explicit microbial-biomass differential equation with a single coupling parameter, which is cheap but demonstrably lossy: against five litter-manipulation field experiments, the PRIM parameterization achieved R² = 0.54 but a regression slope of 0.55 against observed priming, underestimating real priming intensity by about 45% (Towards a representation of priming on soil carbon decomposition in the global land biosphere model ORCHIDEE (version 1.9.5.2)). On the immunity side, the fast arm is circulating antibody, which wanes over months and is escaped by single mutations in the spike's antibody-facing S1/RBD regions; the slow arm is T-cell memory, which decays over years and targets conserved internal regions, which is exactly why the ENSEMBLE sieve analysis found vaccine efficacy against severe disease held up as viral sequences diverged in the antibody-facing regions while dropping with divergence there (Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features). The methodological bridge is CQUESST, which showed that unobservable multi-pool carbon states can be inferred under rigorous uncertainty from aggregate observations when multiple distinct measurement streams (total, particulate, and resistant organic carbon) are assimilated simultaneously in a dynamic stochastic model (CQUESST: A dynamical stochastic framework for predicting soil-carbon sequestration). Vaccine trials have an equivalent multi-stream setup that nobody is using jointly: longitudinal antibody titers, infection event times stratified by severity, and breakthrough-sequence divergence.
Research Program
The object to be built is a two-compartment stochastic survival model, implemented in Stan with the same MapReduce parallelization pattern CQUESST used. For participant (or soil plot) $i$, the instantaneous hazard of the adverse event (severe infection, or a unit of old-carbon loss) is
$$h_i(t) = h_0(t)\,\exp\!\big(\beta_A \log A_i(t) + \beta_C \log C_i(t)\big),$$
where $A_i(t)$ is the observed fast arm (measured nAb titer, or measured fresh-carbon input) and $C_i(t)$ is the latent slow arm (cellular immunity, or microbial-mineral stabilized state). The fast arm is directly observed with assay noise; the latent arm evolves as
$$C_i(t+\Delta) = C_i(t)\,e^{-\lambda_C \Delta} + \gamma\,x_i + \eta_{it}, \qquad \eta_{it}\sim\mathcal{N}(0,\sigma_C^2),$$
with decay rate $\lambda_C$ constrained to be much smaller than the antibody decay rate, a single "priming-style" coupling parameter $\gamma$ scaling an initialization covariate $x_i$ (vaccination versus placebo; hybrid versus naïve infection history; litter addition versus control), and process noise $\eta_{it}$. Three design decisions come directly from the soil literature. First, the identifiability strategy follows CQUESST: no latent state is trusted unless at least three distinct observation streams pin it down — here, titers, severity-stratified event times, and sieve distances, playing the roles that total, particulate, and resistant carbon played there. Second, the failure of the quasi-equilibrium closure in PRIM (slope 0.55) motivates treating dynamic latent estimation versus cheap closure as the central ablation rather than a footnote. Third, the danger of adding complexity without fidelity — the NSI study showed that adding priming or carbon-nitrogen coupling to ORCHIDEE still left every variant predicting the wrong sign of change (Can we model observed soil carbon changes from a dense inventory?) — is answered by a twin-system protocol: every estimator choice, prior, and identifiability diagnostic is first validated on the soil twin where partial ground truth exists, and only the configuration that survives there is applied to trial data. The immunology side also supplies a diagnostic warning: assay thresholds themselves hide the fast arm in 10–30% of mild infections, creating a "high-negative" serology zone that behaves like prior infection (Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections), so titer observation models must be threshold-aware rather than Gaussian.
Baselines
The naive reference point is the antibody-only world, which is not a strawman but current practice: a single-marker Cox proportional hazards regression on nAb titer, exactly as fitted in the ENSEMBLE correlate analysis, which reported a multivariable hazard ratio of 0.31 (95% CI 0.11, 0.89) per 10-fold titer increase and, critically, leaves the 63.1% VE at unquantifiable titers as an unexplained stratum effect (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). Beating this baseline means the model must reproduce that stratum value within its confidence interval (40.0%, 77.3%) from fitted mechanism rather than as an artifact, and must improve out-of-sample discrimination (time-dependent concordance) on severe endpoints under cross-validation. The strongest existing competitor is the variant-adjusted threshold-of-protection Cox model, which scales serum concentration by in vitro IC₅₀ against circulating variants weighted by genomic surveillance; calibrated on the PROVENT trial, it externally predicted sipavibart efficacy in the SUPERNOVA trial with Lin's concordance of 0.86 at 3 months (0.75 at 6 months) and predicted average overall efficacy within 5.0% of observed clinical data (A SARS-CoV-2 variant-adjusted threshold of protection model for monoclonal antibody pre-exposure prophylaxis against COVID-19). That model will be refit on the same ENSEMBLE and COVAIL data, with the same covariates and the same tuning budget, so the comparison is a pure test of structure: it is the strongest competitor precisely because it is the only antibody-only model with demonstrated external predictive skill, but it is a one-arm model by construction and should fail wherever the latent arm carries the protection. A third baseline is a gradient-boosted survival model on all observed markers and covariates, the clinical analog of the Random Forest that reached R² ≈ 0.6 on 37,691 WoSIS soil profiles while process models MIMICS and MES-C managed only R² ≈ 0.13–0.16 (Using explainable AI to diagnose the representation of environmental drivers in process-based soil organic carbon models); it will get the same features and hyperparameter search. Finally, the key ablation replaces the dynamically estimated latent state $C_i(t)$ with the quasi-equilibrium closure used by PRIM — a single coupling parameter tying the slow arm to the fast arm's input — which isolates exactly the ingredient the soil evidence says is decisive. On the soil twin specifically, the PRIM parameterization's own reported numbers (R² = 0.54, slope 0.55 on field priming) are the baseline to beat.
Why not just…?
Why not just use the variant-adjusted Cox threshold model, since it predicts real trial outcomes within 5%? Because its demonstrated success came in the one setting where the single arm is the whole story: monoclonal-antibody prophylaxis in immunocompromised patients who precisely lack robust endogenous cellular responses, so a one-arm model is nearly misspecified-free there. In vaccine-induced polyclonal immunity the same structure demonstrably breaks down: unquantifiable titers with 63.1% residual VE in ENSEMBLE (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial), and only 27.3% of relative booster protection mediated by antibodies in COVAIL (Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost). The decisive experiment is already built into the program: refit the threshold model on ENSEMBLE and test whether it can reproduce the VE-versus-titer curve at the unquantifiable stratum; the prediction, stated in advance, is that it extrapolates toward zero VE there and misses 63.1% by construction.
Why not just train a black-box machine-learning model on all available markers? This is a real threat, and the soil precedent is genuinely on its side — the data-driven Random Forest crushed the process models on global soil profiles (R² ≈ 0.6 versus 0.13–0.16) (Using explainable AI to diagnose the representation of environmental drivers in process-based soil organic carbon models). We will run it with an equal budget, and if it matches our model on every decisive check below, the mechanistic claim should be abandoned. But we expect it to fail on the one axis that matters for a protection model: extrapolation in time and across variants, because the cellular arm's signature is precisely its slow decay and conserved-target resilience, which a model free to overfit fast-decaying antibody features has no reason to encode. The out-of-variant extrapolation experiment (train on pre-Omicron waves, predict later-wave severe VE) is designed to settle this rather than argue around it. There is also a threshold-detection caveat: the same explainable-AI logic shows that what looks like predictive failure can be an artifact of assay cutoffs hiding 10–30% of true seropositivity (Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections), so both models will be evaluated with threshold-aware observation models.
Why not just measure T cells directly instead of estimating them? Honestly, this is the best answer if it were available — the Swiss gold-nanoparticle CD8 vaccine trial shows the community is already building the tools for T-cell-centric protection (A gold nanoparticle/peptide vaccine designed to induce SARS-CoV-2-specific CD8 T cells), though no reported efficacy result was found on Lacuna for that trial. The obstacle is that the correlate cohorts of the major efficacy trials, including ENSEMBLE and COVAIL, were simply not assayed for T-cell responses, so the question of what those trials can tell us about cellular protection is answerable only by latent-state estimation. Moreover, the estimator makes direct measurement cheaper rather than redundant: once the posterior variance of $C_i(t)$ is computed across participants, it identifies exactly which individuals a small, targeted ELISpot substudy should sample to shrink uncertainty fastest — an optimal-design product that a purely empirical analysis cannot provide.
Why won't the added latent compartment just be unidentifiable? This is the objection the soil twin exists to answer before it can be raised. The evidence in the file cuts both ways: adding compartments to ORCHIDEE did not fix its sign error on the NSI inventory, which warns that complexity without the right drivers is worthless (Can we model observed soil carbon changes from a dense inventory?); but CQUESST demonstrated that latent multi-pool states become identifiable when multiple observation streams are assimilated jointly in a fully stochastic setup (CQUESST). Our protocol therefore requires a pre-registered posterior-identifiability check (profile posteriors over $\lambda_C$ and $\gamma$ on simulated data with trial-matched sample sizes) to pass on the soil twin before any clinical claim is made, and the small number of severe endpoints — 42 severe-critical breakthrough cases in ENSEMBLE — is a real constraint we will report honestly: if the latent arm is only weakly identified at that sample size, the deliverable becomes calibrated uncertainty plus the substudy design above, not a point claim.
Proposed Experiments and Decisive Checks
The program has three stages, each with a pre-stated success criterion drawn from numbers reported in the file. The first stage is the soil twin, using the Moyano incubation database (20 calibration experiments, 164 independent evaluation incubations, and the five field litter-manipulation experiments used to evaluate PRIM). We fit both the quasi-equilibrium closure and the full dynamic latent-state model; success for the dynamic model is a regression slope against observed field priming whose confidence interval covers 1.0, decisively beating PRIM's reported slope of 0.55, while matching or exceeding its R² of 0.54 (Towards a representation of priming on soil carbon decomposition in the global land biosphere model ORCHIDEE (version 1.9.5.2)). A second soil check reuses the Millennium Tillage Trial stream, where success is reproducing the finding that permanent pasture is the most reliable carbon sink with a 0.94 probability of negative carbon flux (CQUESST), while yielding sharper posterior treatment modifiers than the six-pool original because the microbial latent state is now explicitly coupled rather than folded into noise.
The second stage is the joint ENSEMBLE fit, and the decisive check is whether one fitted parameter set simultaneously reproduces four reported numbers: the multivariable nAb hazard ratio of 0.31 (95% CI 0.11, 0.89); the 63.1% VE (95% CI 40.0%, 77.3%) at unquantifiable titers against severe disease; the divergence between moderate and severe protection at the same just-quantifiable titer (32.5% versus 85.2% VE — the clearest empirical fingerprint of two arms with different decay rates); and a model-implied antibody-mediated fraction consistent with the reported 28.6% (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). The sieve check then fits the hazard to sequence divergence using the 1,345 breakthrough Spike genomes, predicting that the latent arm's contribution loads on divergence in conserved regions (where the trial found protection against severe disease preserved) while the antibody arm's contribution loads on RBD and NTD divergence and deep-mutational-scanning escape scores, matching the reported pattern that efficacy fell with Hamming distance in S1, RBD, and NTD (FWER p < 0.001) while the S2 subunit stayed conserved (Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features).
The third stage is out-of-sample. On COVAIL's 985 boosted participants, a single mechanism — the latent arm initialized higher by prior infection — must reproduce the reported naïve/non-naïve interaction (HR 0.74, 95% CI 0.59–0.94 versus HR 0.41, 95% CI 0.23–0.64 per 10-fold nAb increase; interaction p = 0.013) without arm-specific refitting (Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost). Then comes the extrapolation duel against the variant-adjusted threshold model and the gradient-boosted baseline: fit all three on early-variant data, evaluate on later-variant severe endpoints, with the concordance standard set by the threshold model's own best result (Lin's concordance 0.86 at 3 months on SUPERNOVA; predicted efficacy within 5.0% of observed) (A SARS-CoV-2 variant-adjusted threshold of protection model). The PREVENT-19 antibody-marker analysis of the NVX-CoV2373 vaccine is available as a further external set (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), though no reported numerical metric was found on Lacuna for it, so it will be used for directional consistency rather than a quantitative bar. A final check imports the explainable-AI diagnostic logic in reverse: permutation importance and partial dependence applied to our fitted model must show non-monotone antibody-titer dependence of the kind the soil study found empirically for NPP, rather than the monotone scaling that single-marker Cox and threshold models impose by construction (Using explainable AI to diagnose the representation of environmental drivers in process-based soil organic carbon models).
The honest framing of what would falsify this program: if the dynamic latent state fails to beat the quasi-equilibrium closure on the soil twin, the whole transfer is premature and we should say so; if it succeeds there but the clinical posteriors on 42 severe events are too diffuse to reproduce the ENSEMBLE stratum value, the contribution is an identifiability map and a substudy design, not a protective-threshold claim. Either outcome is informative, because both fields currently carry the same unexplained residual and neither has a principled way to size it.
References
- Can we model observed soil carbon changes from a dense inventory? A case study over England and Wales using three versions of the ORCHIDEE ecosystem model (AR5, AR5-PRIM and O-CN)
- Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial
- Towards a representation of priming on soil carbon decomposition in the global land biosphere model ORCHIDEE (version 1.9.5.2)
- Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features
- CQUESST: A dynamical stochastic framework for predicting soil-carbon sequestration
- Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections
- A SARS-CoV-2 variant‑adjusted threshold of protection model for monoclonal antibody pre-exposure prophylaxis against COVID-19
- Using explainable AI to diagnose the representation of environmental drivers in process-based soil organic carbon models
- Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost
- A gold nanoparticle/peptide vaccine designed to induce SARS-CoV-2-specific CD8 T cells: a double-blind, randomized, phase 1 study in Switzerland
- 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