Two research communities have independently run into the same wall. In social-media popularity forecasting, the speed at which a post gathers its first few shares predicts its final audience remarkably well — yet when analysts hold that early speed constant, predictive accuracy collapses from above 85% to below 65%, proving that early speed is a proxy for something else, not the mechanism that sustains growth (Predictability of Popularity: Gaps between Prediction and Understanding). In vaccinology, the cheap and fast biomarker is the serum neutralizing antibody titer rather than the retweet rate, and the same pattern appears: in a 44,325-person vaccine trial, participants whose antibody levels were too low to measure still retained 63.1% (95% CI 40.0%, 77.3%) protection against severe disease, and a formal mediation analysis attributed only 28.6% of protection against severe outcomes to neutralizing antibodies at all (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). The residual is widely believed to be cellular immunity — T cells that kill infected cells once antibodies have failed to block infection — but because T-cell assays require frozen viable cells and labor-intensive protocols, the cellular arm was never measured in these trials and survives only as an unmodeled number left over after antibody effects are subtracted.
The opportunity is that the popularity-forecasting literature has already built, and stress-tested, exactly the machinery this situation calls for: generative cascade models that estimate a latent growth-and-truncation process from only the first sliver of observations (A Bayesian Approach for Predicting the Popularity of Tweets), matched-evaluation designs that expose whether a strong predictor is a proxy or a cause (Predictability of Popularity: Gaps between Prediction and Understanding), and derivative-aware training objectives that force models to match the velocity and acceleration of an early trajectory (SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity). Nobody has pointed this machinery at the immune system. This proposal does.
Core Claim
The claim is that vaccine protection is structurally a two-arm cascade, and that the second arm can be recovered statistically even though it was never measured. An exposure either gets blocked at the door by antibodies (the humoral arm, which we observe through titers), or it penetrates and becomes a within-host replication cascade whose final size — and hence whether the infection is mild or severe — depends on how fast infected cells are cleared by T cells (the cellular arm, which we do not observe). By writing down this two-stage generative process and estimating the per-person cellular clearance rate as a latent variable with hierarchical partial pooling, we can forecast the antibody-independent component of protection from existing trial data, reproduce the published vaccine-efficacy curves without fitting to them, and — importing Shulman's matched-evaluation design — give immunology its first general test for whether an immune correlate is a proxy or a mechanism.
Research Program
The first component is the model itself. For each vaccinated person $i$, we track an antibody trajectory $a_i(t)$ that wanes from its measured peak; we adopt the exposure-proximal reconstruction of titers on the actual day of exposure from 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 normalize titers as fold-differences from a reference geometric mean to absorb the up-to-50-fold inter-laboratory assay variability documented in Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection. Exposure at time $\tau_i$ leads to infection with a probability that rises as $a_i(\tau_i)$ falls, modulated by viral sequence features — the physicochemical Hamming distances and antibody-escape scores that Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features showed carve into efficacy. Given infection, viral load grows as a birth-and-death branching process — the direct analog of the binomial retweet branching in A Bayesian Approach for Predicting the Popularity of Tweets — and the outcome is severe if the cascade crosses a threshold $V^{*}$ before it is extinguished:
$$\Pr(\text{severe}_i) \;=\; p_{\text{inf}}\!\big(a_i(\tau_i)\big)\;\cdot\;\Pr\!\big[V_i(t) > V^{*} \text{ before extinction, with per-person clearance rate } c_i\big]$$
where $c_i$ is the latent cellular clearance rate, drawn from a population distribution whose hyperparameters are shared across the cohort and estimated by Markov Chain Monte Carlo with shrinkage, exactly as A Bayesian Approach for Predicting the Popularity of Tweets pooled hyperpriors across 52 historical cascades to predict new ones within minutes. This partial pooling is not a stylistic choice: ENSEMBLE contains only 42 severe breakthrough cases, and shrinkage across individuals is the standard small-sample remedy, just as it was for 52 cascades. Sieve features enter only the first stage, which mirrors a finding from the cascade literature: structural features are idiosyncratic and their effects reverse across platforms, while kinetic features transfer universally (Predictability of Popularity: Gaps between Prediction and Understanding). Our cross-domain prediction, testable below, is that early viral-growth kinetics will predict severity across vaccine platforms with near-identical coefficients, while antibody-marker effects will not — PREVENT-19's hazard ratio of 0.13 per 10-fold titer increase versus the naive cohort's 0.74 in COVAIL (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; Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost) already hints at exactly this split.
The second component is a matched-evaluation diagnostic imported wholesale from Predictability of Popularity: Gaps between Prediction and Understanding. Their temporally matched design held early adoption speed constant and watched accuracy collapse, exposing speed as a proxy; the immunological analog is a titer-matched correlates analysis: among breakthrough cases, match individuals on exposure-proximal antibody titer and ask whether any measured covariate still predicts severity. If predictive power collapses the way it did on Last.fm, Twitter, Flickr, and Goodreads, the residual protection really is latent and the two-arm model is justified; if observables retain predictive power, the latent arm is unnecessary and we will say so.
The third component applies the early-velocity forecasting result to infection itself. Just as the first ten comments raise predictive $R^2$ from near zero to 0.47 in Cannot Predict Comment Volume of a News Article before (a few) Users Read It, and Day-1 views cut Day-30 mean absolute error by more than 50% in SMTPD, the slope and acceleration of viral load over the first serial PCR samples should forecast final infection duration and severity. The Oxford healthcare-worker cohort — 11,475 workers with 245 PCR-confirmed infections and serial sampling, from Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections — is the one dataset in this corpus with the temporal resolution to test this; no reported result found on Lacuna for viral-load-trajectory-based severity prediction in that cohort, which is precisely the gap.
Proposed Experiments
Experiment one fits the two-arm model on the ENSEMBLE case-cohort (NCT04505722; 44,325 participants, 42 severe-critical cases, Day 29 markers, 1,345 sieve-sequenced Spike sequences) using only moderate-disease endpoints and non-severe breakthroughs, then predicts the held-out severe-disease efficacy curve — the 63.1% floor at unquantifiable titers, 85.2% at 5.2 IU/ml, and 95.1% at the 90th percentile reported by Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial — plus the 28.6% mediation fraction. Success is mean absolute error under roughly ten VE percentage points across titer bins with nominal credible-interval coverage. Experiment two tests cross-platform transfer: fit on PREVENT-19 (287 Delta cases, 446 noncases) and predict only the direction and rough magnitude of the COVAIL naive-versus-hybrid divergence (HR 0.74 versus 0.41, interaction p = 0.013, mediation 27.3%, per Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost); prior infection should shift the inferred clearance distribution upward, and the model is never shown the COVAIL interaction when fitting. Experiment three runs the titer-matched collapse test on pooled breakthrough data. Experiment four runs the early-velocity severity forecast on the Oxford serial-PCR sub-cohort, trained with a derivative-aware objective in the spirit of the Composite Gradient Loss, and evaluated as median absolute percentage error on time-to-clearance — with the caveat that Zaman's MAPE below 40% from observing the first 10% of a retweet cascade is an analogy benchmark from the other domain, not a like-for-like number. Experiment five tests kinetic universality directly: refit the severity model per platform and compare coefficient stability for kinetic versus marker features.
Baselines
The naive reference point is the do-nothing assumption that the antibody titer is the whole story, operationalized as the logistic titer-protection curve of Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection, in which 50% protection sits at roughly 20% of the convalescent mean titer and a vaccine at 0.44-fold of the Moderna reference geometric mean exceeds 80% efficacy against ancestral strains. We implement this curve on the same normalized titers and fit it on infection endpoints; it is the strongest competitor for the claim that nothing beyond antibodies is needed, and it is already falsified on the key datum — it predicts near-zero severe-disease efficacy at unquantifiable titers, where the observed value is 63.1%. Beating it means predicting that floor within our credible interval, which no single-arm curve can.
The strongest existing method is the weighted Cox correlates-plus-mediation analysis of Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial, which reports a hazard ratio of 0.35 (95% CI 0.13, 0.90) per 10-fold increase in Day 29 nAb-ID50 for severe-critical disease (0.31, CI 0.11–0.89, in the multivariable model) and a 28.6% mediation fraction. We will run it ourselves on the same case-cohort data, same markers, same covariates, and same compute budget — Cox fitting is cheap, so we also cap our MCMC at comparable wall-clock time and give our model no extra data — and compare on the held-out severe-VE-curve prediction from Experiment one, where both methods must output a titer-binned efficacy curve with uncertainty. Cox regression yields associations conditional on the fitted data; it cannot forecast an individual breakthrough's severity, and it cannot transfer across platforms, so those are secondary comparisons we expect to win by construction, and we will report them as such rather than as fair fights.
The key ablation is our own model with the latent clearance arm removed: a single-arm version in which cascade extinction depends only on measured antibody levels. It shares the code, priors, data, and compute of the full model, so any difference is attributable to the latent cellular parameter alone. The ablation passes if, and only if, it fails to reproduce the zero-titer efficacy floor that the two-arm model recovers — that contrast is the entire content of the proposal, and if the single-arm model somehow reproduces the floor through antibody waning dynamics alone, the proposal is wrong and the experiments will show it.
Why not just…?
Why not just measure T cells directly in the trials? This is the real long-term fix, and trials are moving that way — the Swiss phase 1 gold-nanoparticle/peptide vaccine deliberately built to induce CD8+ T cells (A gold nanoparticle/peptide vaccine designed to induce SARS-CoV-2-specific CD8 T cells: a double-blind, randomized, phase 1 study in Switzerland) will eventually produce exactly the cellular readouts the big Phase 3 trials lack. But viable-cell logistics mean that essentially all existing Phase 3 case-cohort data — the data where severe outcomes actually occurred — will never have cellular measurements, and our model's whole point is to extract quantitative, falsifiable predictions from that data now. The honest concession is that direct measurement is superior where it exists, so the decisive check is correlational: wherever any cellular readout exists in a sub-cohort, the inferred clearance rate $c_i$ must correlate with measured CD8+ responses, and if it does not, the latent we recovered is not cellular immunity but something else, and we will report that failure.
Why not just extend the Cox and mediation analyses with more antibody markers? This has been tried and it runs into a ceiling: the multivariable ENSEMBLE model already found that binding IgG contributes nothing after neutralizing activity is adjusted for (hazard ratio 1.22, 95% CI 0.49, 3.02) (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). Adding more humoral markers cannot identify protection that is not antibody-mediated, because the confounding quantity was never measured; mediation analysis quantifies the residual but neither names nor forecasts it. We keep the one genuine improvement from this line — Fong's exposure-proximal titers (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) — as an input to our model, so the comparison isolates the latent-variable machinery, not better feature engineering.
Why not just train a flexible machine-learning classifier on the case-cohort data? The data cannot support it. Cannot Predict Comment Volume of a News Article before (a few) Users Read It needed 19,433 articles before a Random Forest reached $R^2 = 0.560$, and ENSEMBLE has 42 severe cases; a black box on that sample size will memorize noise. The hierarchical generative model with partial pooling is the principled small-sample tool, and it is the same tool that made 52 cascades sufficient in A Bayesian Approach for Predicting the Popularity of Tweets. There is also a deeper objection from the cascade literature itself: Shulman's predictive paradox shows that high observational accuracy can coexist with total ignorance of mechanism, so even a successful black box would not tell us whether it had recovered cellular protection or merely re-learned the titer. Our titer-matched collapse test would catch that, and we will run it on any classifier baseline as readily as on our own model.
Decisive Checks
Four outcomes would falsify the core claim, and each is built into the program. If the single-arm ablation reproduces the 63.1% zero-titer efficacy floor as well as the two-arm model, the latent cellular arm is redundant. If the titer-matched test shows measured covariates retaining most of their predictive power among breakthroughs, the residual is not latent and the generative model is over-engineering. If Experiment two's cross-platform transfer fails — if kinetic features do not transfer while marker features do — the analogy to universal temporal features in cascades is broken and the cross-domain mapping collapses. And if inferred clearance rates fail to track any available cellular readout, the latent variable is capturing something other than T-cell immunity. Conversely, a two-arm model that predicts the held-out severe-efficacy floor, recovers the mediation fraction out of sample, and transfers across platforms would give vaccinology its first quantitative forecast of the protection it never measured — and would give the popularity-forecasting community the strongest evidence yet that their machinery generalizes far beyond social feeds.
References
- Predictability of Popularity: Gaps between Prediction and Understanding
- Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial
- A Bayesian approach for predicting the popularity of tweets
- SMTPD: A New Benchmark for Temporal Prediction of Social Media Popularity
- 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
- Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection
- Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features
- Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost
- Cannot Predict Comment Volume of a News Article before (a few) Users Read It
- Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections
- A gold nanoparticle/peptide vaccine designed to induce SARS-CoV-2-specific CD8 T cells: a double-blind, randomized, phase 1 study in Switzerland