When we ask a model "is this person protected from severe disease?", the easiest measurable signal is the concentration of neutralizing antibodies in their blood. High-throughput serology makes this cheap, and on average antibody titer tracks protection well — so predictive models, and public-health decisions built on them, lean on it heavily. But the lean is misleading. In the ENSEMBLE trial of the single-dose Ad26.COV2.S vaccine, participants whose neutralizing antibody levels were too low to quantify at Day 29 still retained 63.1% (95% CI: 40.0%–77.3%) vaccine efficacy against severe-critical COVID-19, and mediation analysis attributed only about 28.6% of the vaccine's protection against severe disease to neutralizing antibodies (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). The missing protection is carried largely by cellular immunity — T cells recognizing conserved viral peptide fragments — which current assays cannot measure at trial scale because it requires live frozen blood samples.
This is, structurally, a shortcut-learning problem that the machine-learning fairness literature has already solved in other domains. A model trained on aggregate loss exploits features that work for the majority and fails silently on a minority whose behavior is governed by a different mechanism. A growing line of work shows that penalizing unevenness of loss across individuals or subgroups blocks exactly these shortcuts: minimizing per-individual loss variance upper-bounds the worst discrepancy against unlabelled latent subgroups (Maximum Weighted Loss Discrepancy); forcing a shared model to perform acceptably on every group restricts the hypothesis space so that small groups borrow statistical strength from large ones instead of being overfit (To Pool or Not To Pool: Analyzing the Regularizing Effects of Group-Fair Training on Shared Models); and in deepfake detection, removing a demographic shortcut via fairness training improved out-of-distribution accuracy by an average of 2.35 percentage points (Fair Deepfake Detectors Can Generalize). The humoral–cellular gap is the same phenomenon wearing a biological costume: antibody titer is the majority shortcut, and the "latent subgroup" whose protection the shortcut misses is precisely the cellular-immunity phenotype — unmeasured, unlabelled, and impossible to annotate at scale.
Core Claim
Standard immune-correlate analyses fit weighted Cox survival models that maximize aggregate partial likelihood; like any empirical risk minimization, this makes them "crowd-pleasing" in the sense of Maximum Weighted Loss Discrepancy — they calibrate risk to the antibody signal that dominates the bulk population and systematically miscalibrate the seronegative-protected minority, with downstream bias across infection histories (in the COVAIL booster trial, a 10-fold titer increase reduces hazard by 59% in previously infected participants but only 26% in infection-naive ones, interaction p = 0.013; Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost). We claim that adding a fairness regularizer — a loss-variance or minimax penalty over per-individual survival losses and over discovered immune-phenotype clusters — to a discrete-time survival model of protection (a) materially improves risk calibration and severity discrimination within the unquantifiable-antibody stratum, where titer carries no signal, and (b) makes protection predictions stable under variant escape, because the regularizer forces the model off the Spike-surface antibody feature and toward invariant signals, mirroring how fairness training stabilized deepfake detectors across datasets (Fair Deepfake Detectors Can Generalize) and stabilized tail classes under label noise (Fairness Improves Learning from Noisily Labeled Long-Tailed Data).
What Exists and Where This Sits
No paper in the Lacuna index bridges these two directions, and the closest works each miss the proposal from a different side. The correlates analyses themselves (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial, Immune correlates analysis of the ENSEMBLE single Ad26.COV2.S dose vaccine efficacy clinical trial, 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) treat unmeasured cellular immunity as residual confounding in a weighted Cox model, with no subgroup-fair objective and no mechanism for the 42 severe breakthrough cases among 44,325 participants to borrow strength from the rest of the cohort. The method papers supply machinery but have never touched censored clinical data: Maximum Weighted Loss Discrepancy is evaluated only on tabular socio-economic benchmarks, and To Pool or Not To Pool only on 15-dimensional synthetic linear models. On the design side, Using epitomes to model genetic diversity: Rational design of HIV vaccines optimizes cellular epitope coverage on a population average with no fairness across host HLA alleles, and EVE-Vax (Computationally designed proteins mimic antibody immune evasion in viral evolution) designs antibody-evasion-mimicking antigens while ignoring whether the designs destroy conserved T-cell epitopes; both are immunogen-design problems, not protection-prediction problems, and neither uses any fairness machinery. Our proposal differs from all of these in objective, data type, and claim: we bring unlabelled-subgroup fairness regularization into censored survival modeling of vaccine protection, and we test whether the regularizer recovers the cellular component of protection that titer-based models provably miss.
Research Program
Three workstreams. First, the estimator. We use a discrete-time survival model (a person-period logistic hazard model), because its loss decomposes into independent per-person, per-interval Bernoulli terms — a necessary property, since the Cox partial likelihood couples individuals through shared risk sets and has no well-defined per-individual loss to regularize. The training objective adds the loss-variance penalty of Maximum Weighted Loss Discrepancy's style to the average survival loss:
$$\min_\theta \;\; \frac{1}{n}\sum_{i=1}^{n} \ell_i(\theta) \;+\; \lambda \sqrt{\mathrm{Var}_i\!\left[\ell_i(\theta)\right]},$$
where $\ell_i(\theta)$ is person-interval $i$'s binary cross-entropy loss under model parameters $\theta$, and $\lambda$ controls the fairness pressure. The square-root-of-variance form is chosen because it upper-bounds the worst weighted loss discrepancy against any latent subgroup, which is exactly our setting: cellular immune status is an unlabelled attribute that partitions the cohort and cannot be annotated at serology throughput. A second variant replaces the variance penalty with a minimax objective over immune-phenotype clusters discovered by unsupervised KNN or confidence-based splitting of baseline features, following Fairness Improves Learning from Noisily Labeled Long-Tailed Data; this handles the case where the latent subgroup structure is discoverable rather than merely bounded. Second, theory. We extend the variance bound to censored discrete-time losses (censoring truncates which $\ell_i$ are observed, so the variance must be estimated over the observed-risk person-periods), and we port the restricted-hypothesis-space argument of To Pool or Not To Pool to this setting to show why the 42 severe cases can borrow strength from the full cohort only under a group-fair objective. Third, data. Individual-level ENSEMBLE, PREVENT-19, and COVAIL correlates data require data-use agreements; until and unless those land, we develop and power-tune the method on a simulator calibrated to every published quantity cited below (hazard ratios, VE by titer stratum, case counts, assay sensitivities), which is also how we run honest power analyses for strata with single-digit case counts.
Baselines
The do-nothing reference point is a threshold-of-protection rule: declare a person protected above a fixed antibody cutoff, in the spirit of the variant-adjusted threshold model of A SARS-CoV-2 variant-adjusted threshold of protection model for monoclonal antibody pre-exposure prophylaxis against COVID-19. This baseline is not a strawman — it is roughly how immunobridging decisions are made — and it fails in a well-documented way: at unquantifiable titer it predicts near-zero protection, while the true vaccine efficacy there is 63.1% against severe disease (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). That gap is the error we aim to recover.
The strongest existing method is the multivariable weighted Cox model of the ENSEMBLE correlates analysis, which reports a hazard ratio of 0.31 (95% CI: 0.11–0.89) per 10-fold increase in neutralizing antibody titer and 1.22 (0.49–3.02) for anti-Spike binding antibody concentration (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial); the earlier ENSEMBLE correlates analysis reported a Day-29 hazard ratio of 0.49 per 10-fold ID50 increase, P = 0.006 (Immune correlates analysis of the ENSEMBLE single Ad26.COV2.S dose vaccine efficacy clinical trial). We will re-fit this Cox specification ourselves on the same features, folds, and case-control weights, with an identical tuning budget, so the comparison isolates the regularizer rather than plumbing. We count ourselves as beating it only if, at no worse cohort-level discrimination, the regularized model achieves materially better calibration and severity discrimination inside the unquantifiable-titer stratum and more uniform predictive performance across the COVAIL naive/non-naive subgroups. No reported result found on Lacuna for stratum-specific concordance or calibration in any of these trials, which is itself part of the gap.
Two further competitors bracket the proposal. A case-control-weighted Cox model stratified by titer tertile and infection history represents the "just re-weight or stratify" alternative, run with the same features and compute. And an ablation with the identical survival architecture but $\lambda = 0$ (plain empirical risk minimization), plus a minimax-over-observed-strata variant that drops the unlabelled-subgroup machinery, isolates whether the key ingredient — the variance penalty over individual losses that bounds discrepancy against unobserved subgroups — is doing the work, since observed-stratum minimax cannot see the cellular-immunity phenotype at all.
Why not just…?
Why not just re-weight or stratify the Cox model? This is the most serious objection, and we run the stratified weighted Cox as a full baseline rather than arguing around it. The available evidence says re-weighting alone is insufficient in analogous settings: robust re-weighting losses such as focal loss and label smoothing degrade tail subpopulations to near-zero accuracy under label noise, whereas the fairness regularizer lifted noisy imbalanced CIFAR-10 accuracy from 60.03% to 65.12% (Fairness Improves Learning from Noisily Labeled Long-Tailed Data), and inverse propensity weighting alone left spurious demographic correlation intact until paired with latent alignment (Fair Deepfake Detectors Can Generalize). More fundamentally, stratification can only balance strata we observe; cellular immune status is unobserved, so a stratified Cox conditions on the wrong partition. Experiment 1 below settles this empirically: if the stratified Cox matches our stratum calibration, the proposal's core claim fails.
Why not just measure T cells directly? If a high-throughput cellular assay existed at serology scale, this proposal would be unnecessary, and we say so plainly. It does not: cellular readouts require live peripheral blood mononuclear cells, which is why measurements are sparse even in Phase 3 trials (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). Methods that require subgroup labels are therefore the wrong tool class here; the loss-variance approach was built precisely for subgroups that cannot be labelled (Maximum Weighted Loss Discrepancy).
Why not just use variant-specific titers or adjusted assay thresholds? Both have been tried and quantifiably fail. Even with Delta-specific neutralization measurement, 37.5% of severe breakthrough cases in PREVENT-19 had unquantifiable titers, and severe cases clustered in the lowest tertiles with geometric mean 9.6 AU/mL versus 209.5 AU/mL in noncases (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). Threshold adjustment trades one error for another: stringent IgG cutoffs misclassify 10–30% of mild PCR-confirmed infections, dropping Oxford ELISA sensitivity from 99.1% to 89.8% and Abbott from 92.7% to 79.3% (Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections). No monotone function of antibody titer can recover 63.1% efficacy in people whose titer is unmeasurable — the information is simply not in that feature.
Proposed Experiments and Decisive Checks
Experiment 1 targets the core claim on ENSEMBLE. Within the unquantifiable-titer stratum, true efficacy is 63.1% against severe but only 32.5% against moderate disease (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial); a model relying on titer has no signal to separate these, so we measure stratum-specific discrimination and calibration for the plain, stratified, and variance-regularized models under nested cross-validation, with a power analysis driven by the 42 severe-critical breakthrough cases. Experiment 2 tests cross-population stability on COVAIL: a pooled titer model has hazard ratios of 0.41 versus 0.74 across non-naive and naive participants (interaction p = 0.013; Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost), and we test whether the regularized model's subgroup-wise predictive performance is uniform where the pooled model's is not. Experiment 3 tests variant robustness using the sieve cohort of 1,345 Spike sequences from 1,224 participants (Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features): moderate-disease efficacy decays with physicochemical-weighted Hamming distance (FWER p < 0.001) and neutralization resistance (p = 0.011), while severe-disease efficacy stays stable across variants, so a titer-heavy model should degrade with escape score while a regularized model that has moved off the antibody feature should not — this differential is the decisive signature of shortcut removal. A secondary testbed, run only if the first three succeed, applies the cluster-discovery variant to the Oxford healthcare-worker cohort of 11,475 individuals with 245 prior PCR-confirmed infections (Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections), where the 15.5% versus 3.7% anosmia gap across seronegative zones gives an observable symptom-feature proxy for the mislabelled cellular-immune subgroup. Finally, a placebo check keeps us honest: in the high-titer regime against moderate disease, where antibodies genuinely dominate the mechanism, the regularizer should change little, and if the cellular signal is absent from the available features entirely, no regularizer can rescue the unquantifiable stratum — in which case the deliverable is a quantified negative result bounding what titer-only correlate models can ever achieve.
References
- Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial
- Maximum Weighted Loss Discrepancy.
- To Pool or Not To Pool: Analyzing the Regularizing Effects of Group-Fair Training on Shared Models
- Fair Deepfake Detectors Can Generalize
- Neutralizing antibody immune correlates in COVAIL trial recipients of an mRNA second COVID-19 vaccine boost
- Robust Learning from Noisily Labeled Long-Tailed Data via Fairness Regularizer
- Immune correlates analysis of the ENSEMBLE single Ad26.COV2.S dose vaccine efficacy clinical trial
- 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
- Using epitomes to model genetic diversity: Rational design of HIV vaccines
- Computationally designed proteins mimic antibody immune evasion in viral evolution
- A SARS-CoV-2 variant‑adjusted threshold of protection model for monoclonal antibody pre-exposure prophylaxis against COVID-19
- Stringent thresholds in SARS-CoV-2 IgG assays lead to under-detection of mild infections
- Quantifying how single dose Ad26.COV2.S vaccine efficacy depends on Spike sequence features