Orthogonal Ex-Vivo Perturbation for Tracking Durable Immunity

When a vaccine is evaluated, the headline number comes from a compressed window: the mRNA-1273 trial, for instance, reported 93.2% efficacy against symptomatic COVID-19 over a median follow-up of 5.3 months (El Sahly et al.). But protection is not a number; it is a set of decaying curves, and the curves for antibodies and for immune memory cells decay at different rates and through different mechanisms. Roughly half of vaccinated participants in one H1N1 cohort lost protective neutralizing titers by Day 180 (Ming Wang et al.), and 12-month tracking shows exponential antibody waning with accelerated erosion against immune-evasive variants (Juan Li et al.), while memory B and T cells stay comparatively stable (Xun Wang et al.). The field's central embarrassment is that it cannot say which mechanism carries protection at any given month: in the ENSEMBLE trial, people with unquantifiable neutralizing antibodies still retained 63.1% efficacy against severe disease, and antibodies mediated only 28.6% of that protection (Carpp et al.); in COVAIL, antibodies explained only 27.3% of variant-booster protection (Zhang et al.). The remaining ~70% is a statistical residual, not a measurement.

This proposal imports a method from cognitive neuroscience to close that gap. In orthogonal task design, researchers hold the physical stimulus identical while alternating the internal state or instruction, so that any change in the measured signal can be attributed to one process rather than to the stimulus (Petit et al.), or cross two factors factorially, such as improvisation versus recall with vocalized versus imagined performance, to isolate a construct from motor artifacts (Vergara et al.). Immunology has never done the equivalent: existing longitudinal studies measure whatever the blood happens to contain and then try to disentangle the components statistically after the fact. We propose instead to design the disentangling into the experiment, by splitting each blood sample into a factorial grid of ex-vivo perturbation conditions, so that decay curves for humoral output, CD4-mediated help, and CD8-mediated cytotoxicity are measured as separately varied dimensions rather than statistically inferred entangled covariates.

Core Claim

The core claim is that pathway-level attribution of durable immunity requires orthogonal perturbation, not more observational measurement. Concretely: at each longitudinal timepoint, each participant's PBMC sample (a standard vial of frozen white blood cells) is split into aliquots crossed along three designed dimensions — stimulation depth (immediate ex-vivo restimulation versus an 8-day in-vitro expansion that reveals low-frequency memory, the assay that rescued the PREVAC trial's cellular readout after standard assays read "undetectable" at 1 year, Wiedemann et al.); antigen panel (a pool of conserved ancestral epitopes versus variant-specific escape-mutant epitopes); and pathway perturbation (no perturbation, class-II MHC blockade to silence CD4 help, class-I MHC blockade to silence CD8 recognition, and B-cell depletion to silence humoral output). Every aliquot receives the same antigen concentrations, culture medium, and readout panel, so differences between aliquots are attributable to the designed factor, exactly as identical stimuli with alternating instructions isolate covert cognition (Petit et al.). A mixed-effects decomposition then assigns each measured response to subject, time, and designed factors:

$$r_{i,t} = \mu + s_i + f(t) + \alpha_d + \beta_a + \gamma_p + \text{interactions} + \epsilon_{i,t}$$

where $r_{i,t}$ is a response readout (cytokine-positive T-cell frequency or antibody output) for subject $i$ at month $t$ after boosting, $s_i$ is a subject random effect, $f(t)$ is the shared decay curve, and $\alpha_d$, $\beta_a$, $\gamma_p$ are the main effects of stimulation depth, antigen panel, and pathway perturbation. The interaction terms are the scientific payload: they test the additivity assumption that both neuroscience factorial designs (Vergara et al.) and immune-correlates models currently make implicitly, and localize — for example — whether the naïve-versus-hybrid immunity gap that COVAIL detected only as a binary covariate (HR 0.74 versus 0.41 per 10-fold titer, interaction p = 0.013, Zhang et al.) mechanistically resides in the humoral arm specifically.

Research Program

Three components. First, a computational pilot on HR-VILAGE-3K3M, the public benchmark harmonizing 66 longitudinal respiratory-vaccination transcriptomic studies (3,178 subjects, 14,136 profiles). The pilot tests the measurement-invariance half of the design: the benchmark's authors document that ComBat batch harmonization removes technical artifacts but also suppresses the genuine biological variance separating vaccine responders from non-responders. We will replace cross-study mean removal with normalization anchored to an invariant within-batch reference condition — the wet-lab analog of holding the stimulus fixed — and test whether responder/non-responder separation survives correction. Second, a prospective cohort of roughly 150 adults receiving a variant-adapted mRNA booster, sampled at Months 0, 1, 3, 6, and 12, with each visit's sample split across the 2 × 2 × 4 perturbation grid described above, plus serum neutralizing titers at every visit. Third, breakthrough-infection surveillance over 12 months, modeled on the COVAIL protocol (Zhang et al.), so that mechanistic attribution can be linked to actual risk. We are deliberately honest about scale: 150 subjects cannot re-estimate vaccine efficacy, so the primary endpoints are within-subject mechanistic contrasts (highly efficient because each subject is their own control across perturbation conditions), with breakthrough prediction as a secondary, individually-powered endpoint. A second honesty note: the B-cell-depletion and MHC-blockade conditions are exactly the kind of labor-intensive cellular work the field avoids; that is the point, but it also caps throughput, which is why the third component matters — any portable signature the perturbation grid extracts (a gene-expression or cytokine fingerprint that predicts which pathway carries an individual's memory) can then be tested for transportability back onto the HR-VILAGE meta-cohort, where no perturbation labels exist.

The design also answers the assay-sensitivity problem directly. PREVAC showed that standard ex-vivo intracellular cytokine staining falsely reports cellular immunity as vanished, while 8-day expansion shows it intact at 60 months (Wiedemann et al.). Crossing depth with pathway blockade goes one step further: when the ex-vivo readout hits undetectable at Month 6, the expanded-plus-blocked conditions tell us not just that memory exists but which pathway holds it, and whether the load-bearing pathway shifts over time (say, from CD4 help toward CD8 recognition as titers fall) — a transition no current durability study can resolve.

Proposed Experiments

The pilot's benchmark numbers come from the file: on HR-VILAGE-3K3M, a PCA-based logistic regression predicted vaccine responsiveness with F1 = 0.822 and AUC = 0.748 (Sun et al.), and it outperformed several deep architectures. Our pilot succeeds if reference-anchored normalization retains or improves that AUC while preserving responder/non-responder variance that ComBat demonstrably erases; it fails informatively if the invariance anchor cannot beat cross-study correction, which would argue the batch problem needs experimental, not analytical, solutions in future collections. The prospective cohort's primary metrics are, first, the mechanistic attribution share: the fraction of between-subject protection-relevant variance assigned to named pathways, benchmarked against the 27.3% that antibodies explain in COVAIL (Zhang et al.); second, pathway-localized hybrid-immunity structure: whether the naïve/hybrid divergence localizes to the B-cell-depleted arm, which would give COVAIL's p = 0.013 interaction a mechanism; third, additivity tests on the interaction terms of the decomposition above, with a pre-registered interpretation that large interactions are themselves the finding — a measured failure of the decomposition assumptions both fields rely on.

Baselines

The naive reference point is serum-only tracking: measure neutralizing titers longitudinally, fit the biphasic decay model, and predict risk from predicted-at-exposure titer, exactly as COVAIL does (Zhang et al.). This is the strongest cheap competitor because it is what the field actually runs: the cited paper reports a hazard ratio of 0.74 (95% CI 0.59–0.94) per 10-fold titer increase in infection-naïve participants and 0.41 (95% CI 0.23–0.64) in the non-naïve, with antibodies mediating 27.3% of booster protection. We will run it on our own cohort's serum, with the same surveillance definition and the same Cox machinery, so the comparison is purely added-information, not added-data. Beating it means two things: held-out breakthrough prediction with a confidence interval on the AUC improvement excluding the titer-only model, and a mechanistically explained variance share materially above the 27.3% benchmark, with named pathways rather than a residual. If the titer-only model matches our panel's predictive performance, the attribution is still informative, but we will report the predictive null honestly.

The strongest existing method is the PREVAC dual-depth cellular pipeline: standard ex-vivo intracellular cytokine staining paired with the 8-day EpiMax expansion, tracked over 60 months, which is the best cellular durability measurement in the file (Wiedemann et al.). We will run it as a nested baseline on the identical aliquots, same batches, same gates, same technician time, so the only difference is the absence of the perturbation dimension. Beating it means showing attribution it cannot provide: specifically, that two participants with identical PREVAC-style readouts (both "memory present, stable") diverge in predicted risk depending on their blockade profile — that is, the perturbation grid carries risk information the depth-only assay misses. The relevant reported comparison is qualitative: PREVAC demonstrated no significant decline in expanded T-cell frequencies between 24 and 60 months, but could not say which pathway sustained them.

The third baseline is the observational multivariable approach: measure everything we measure (titers, ex-vivo ICS, expansion readouts) but analyze it with multivariable Cox models and ensemble learners on the unperturbed aliquots only, so no orthogonal information enters. The dengue correlates work shows this approach's ceiling in a related task: Super-Learner modeling of microneutralization titers reached a cross-validated AUC of up to 0.84, and the assays met Prentice surrogate criteria only for DENV-1 among four serotypes (Carpp et al.). The ENSEMBLE multivariable model also shows its pathology: entering correlated humoral markers together inverted the anti-Spike binding antibody hazard ratio to 1.22 (95% CI 0.49–3.02) while neutralizing antibody held at 0.31 (Carpp et al.) — collinearity among entangled covariates distorting effect estimates. We give this baseline the same feature set and tuning budget; if it attributes pathways as cleanly as the perturbation grid, the experimental design is not paying its way.

Finally, the key-ingredient ablation removes only the perturbation dimension while keeping depth, antigen, and all measurement — that is the third baseline above — and a second ablation removes only the expansion depth, keeping blockade. If either ablation alone recovers the full panel's attribution, the corresponding design dimension is the load-bearing one and the factorial is over-engineered.

Why not just…?

Why not just measure more biomarkers and adjust statistically? Because the file's evidence says statistical adjustment breaks exactly where we need it. The ENSEMBLE multivariable model produced an inverted hazard ratio (1.22) for a binding antibody that is biologically protective, a signature of collinearity among entangled markers (Carpp et al.), and COVAIL's sophisticated biphasic-decay-plus-stratification machinery still left 72.7% of protection unexplained (Zhang et al.). That said, this is a real threat for population-level questions: if our head-to-head experiment shows the multivariable baseline predicts breakthrough risk as well as the perturbation grid, the honest conclusion is that designed perturbation matters for mechanism but not for prediction, and we will write that.

Why not just add the EpiMax expansion assay, as PREVAC already did? This is the closest simple alternative and we treat it seriously: depth alone already rescued the false "immunity vanished" readout in PREVAC's rVSV arm. What it cannot do is attribute: PREVAC can say memory T cells persist for 60 months but not which pathway sustains them or why the same persistence protects some regimens more than others (Wiedemann et al.). The nested ablation — identical samples with blockade removed — settles this directly: if depth-only readouts predict risk and attribute pathways as well as the crossed grid, the blockade dimension is unnecessary and the proposal collapses to a PREVAC replication with more antigens.

Why not just run ComBat or deep learning on the existing transcriptomic repositories? The benchmark evidence is unambiguous: on HR-VILAGE-3K3M, ComBat harmonization suppressed exactly the responder/non-responder biological variance we care about (Sun et al.), and a simple PCA logistic regression (F1 0.822, AUC 0.748) beat several deep architectures on the raw data, indicating the entangled signal resists unsupervised separation. Our computational pilot is precisely the test of whether analytical anchoring can do better; if reference-anchored normalization cannot, that strengthens the case for fixing this at collection time.

Why not just rely on serology, since microneutralization correlates at Spearman 0.83–0.97 with the gold-standard assay? Those correlations coexist with surrogate failure: in the dengue trials, neutralization titers met Prentice criteria only for DENV-1 and were only partially valid for the other three serotypes (Carpp et al.); stringent IgG thresholds systematically miss mild infections cleared without seroconversion (Eyre et al.); and ENSEMBLE retained 63.1% severe-disease efficacy at unquantifiable titers (Carpp et al.). Serology is the right cheap baseline, and we run it as such, but it provably misclassifies protected people as vulnerable.

Decisive Checks

The proposal stands or falls on three outcomes: whether invariant-reference normalization preserves responder variance that ComBat erases on HR-VILAGE-3K3M; whether the perturbation grid raises the mechanistically explained share of protection above the 27.3% antibody-only benchmark from COVAIL; and whether the blockade arms localize the naïve/hybrid interaction (p = 0.013 in Zhang et al.) to a named pathway. Each check has a pre-registered failure mode, and any of the three failures would itself tell the field something useful about where the limits of decomposition — experimental or statistical — actually lie.