Every large vaccine trial ends with the same awkward number. In the ENSEMBLE trial of the Ad26.COV2.S vaccine, neutralizing antibodies measured four weeks after vaccination were a genuine correlate of protection — a hazard ratio of 0.35 (95% CI 0.13–0.90) per ten-fold titer increase against severe-critical COVID-19 — yet mediation analysis attributed only 28.6% of the vaccine's protection against severe disease to them, and participants with unquantifiable titers still retained 63.1% efficacy (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial). The same floor appears in PREVENT-19, where vaccine efficacy was 75.7% even at a neutralization titer of just 50 IU50/ml (Immune correlates analysis of the PREVENT-19 COVID-19 vaccine efficacy clinical trial). Most protection is carried by something the univariate marker cannot see.
Cognitive neuroscience spent two decades solving a formally identical problem. When distinct neuronal sub-populations are spatially intermingled, the averaged signal shows no difference between conditions even though the distributed pattern across voxels separates them perfectly — which is how subjective perception of physically identical ambiguous syllables was decoded at 60% accuracy against a 50% chance level from auditory cortex (Auditory Cortex Encodes the Perceptual Interpretation of Ambiguous Sound). Multivariate pattern analysis (MVPA) treats each measurement snapshot as a point in a high-dimensional space and asks whether a linear boundary separates conditions there; time-resolved MVPA goes further and asks whether a boundary trained at one moment still works at another, revealing whether a neural code is stable or dynamically reorganized (How face perception unfolds over time). This proposal applies that second question to the immune system: it treats each person's multi-omic immune profile at each sampling date as a pattern vector, and trains classifiers at early time points to test at later ones, on the months-long timescale of immune memory rather than the millisecond timescale of perception.
Core Claim
The multivariate immune state that predicts durable antibody responses is not one stable configuration but a relay: an early innate and monocyte-driven code, handed off to a later cellular-memory code. This is a direct, testable extrapolation from two strands of evidence. In humans vaccinated against H5N1, gene-expression changes in classical monocytes and CD8+ naive-like T cells persist for at least 100 days and early response patterns predict later antibody titers (Acute and persistent responses after H5N1 vaccination in humans); meanwhile, across SARS-CoV-2 breakthrough infection and Ebola vaccination, humoral titers wane while memory B and T cell responses stay stable for months to years (Longitudinal Analysis of Humoral and Cellular Immune Response up to 6 Months after SARS-CoV-2 BA.5/BF.7/XBB Breakthrough Infection and BA.5/BF.7-XBB Reinfection; Long-term cellular immunity of vaccines for Zaire Ebola Virus Diseases). Whether these are the same underlying representation carried forward, or genuinely different compartments taking over, cannot be answered by any static regression — including the Elastic Net models already run on the H5N1 data. Only cross-temporal decoding can, and it has never been applied to longitudinal immune profiling.
The payoff is twofold. Scientifically, the shape of the temporal generalization matrix tells us whether protective immune information is preserved or re-encoded across the contraction phase. Practically, if some sampling dates are informationally redundant with others, longitudinal tracking protocols — which currently suffer from cost and participant attrition — can be compressed.
Research Program
The primary substrate is the H5N1 cohort of Acute and persistent responses after H5N1 vaccination in humans: 42 volunteers, 42 with 14 blood draws from within 24 hours to Day 100, profiled by PBMC microarrays, 10-color flow cytometry, Luminex cytokines, and single-cell CITE-seq and scATAC-seq. With 42 subjects and thousands of raw features, we work in a reduced feature space: the 14 transcriptomic and 8 cell-population response patterns the authors already extracted, plus a matched antibody-panel vector. The core object we build is the temporal generalization matrix. For each training time $t_{\text{train}}$ and test time $t_{\text{test}}$,
$$A(t_{\text{train}}, t_{\text{test}}) = \text{Balanced accuracy of a classifier trained on profiles from } t_{\text{train}} \text{, evaluated on profiles from } t_{\text{test}} \text{, never mixing one subject across the split.}$$
Classifiers are linear — regularized logistic regression and linear SVMs — both because they survive small samples and because, as the MVPA literature argues, a linear boundary corresponds to a biologically plausible readout: a downstream cell that integrates weighted inputs. The decoding target is immune fate: membership in the top versus bottom tercile of late antibody titer (Day 28 onward), so that early profiles are asked to predict durable outcome, not contemporaneous state. A full matrix with sustained off-diagonal accuracy means a stable code; blocks along the diagonal with sharp falloff mean a relay; asymmetric transfer identifies the handoff window.
Two methodological hazards are imported directly from the neuroimaging literature, because they have exact immune analogues. First, run-to-run covariance changes create false positive decoding in fMRI unless cross-validation respects run partitions, and removing scanner-effects from the covariance improves MVPA (Removal of Scanner Effects in Covariance Improves Multivariate Pattern Analysis in Neuroimaging Data); the immune analogue is assay batch and plate effects, so all splits hold out entire processing batches, and we apply the same covariance normalization across sampling dates before asking whether a Day-7 classifier works on Day 100. Second, multivariate decoders are known to degrade sharply across time when trained in one session and tested in another, which is precisely what the CrossPT-EEG benchmark quantifies for EEG sessions days apart (CrossPT-EEG: A Benchmark for Cross-Participant and Cross-Time Generalization of EEG-based Visual Decoding); we treat that degradation as the null structure to be distinguished from genuine transfer, and report it under matched split designs.
A secondary analysis uses representational similarity analysis. We build representational dissimilarity matrices over conditions (time point × adjuvant status) separately for each modality — flow cytometry, transcriptomics, cytokines — and compare them with rank correlations, using the machinery of the existing RSA toolboxes (A Toolbox for Representational Similarity Analysis; A Python Toolbox for Representational Similarity Analysis). Because RDMs abstract away measurement units, this is the one existing method for asking whether the flow-cytometry geometry and the transcriptomic geometry of vaccine response converge or diverge over the 100 days — a cross-platform question the correlates literature raises but never answers geometrically.
Proposed Experiments
The first experiment runs on the H5N1 cohort. A positive control first: decoding adjuvanted versus non-adjuvanted recipients, which the original study shows is separable, must succeed or the pipeline is broken. Then the durability decoders: diagonal accuracy at each of the 14 time points, followed by the full generalization matrix, with subject-level leave-one-out cross-validation, batch-held-out splits, and permutation tests on labels. The pre-registered prediction is relay structure — Day 1–7 profiles dominated by innate and monocyte features fail to transfer to Day 100, while Day 14–28 profiles transfer forward — and its rejection (full transfer) would itself be an informative result about immune coding.
The second experiment tests sampling compression: greedy selection of the minimal subset of time points whose diagonal decoders recover the full panel's late-titer prediction. This directly attacks the attrition and cost bottleneck that the longitudinal-tracking literature names as its main logistical limitation.
The third experiment is the titer-matched test of the protection floor. In the CoVPN 3008 cohort of people with HIV, which measured neutralizing and binding antibodies in hybrid-immune versus vaccine-only groups (Neutralizing and binding antibodies are a correlate of risk of COVID-19 in the CoVPN 3008 study in people with HIV; no numerical hazard ratio is reported in the Lacuna file for this study), we match participants on univariate titer and ask whether the multivariate antibody pattern still separates the two groups. This is the exact logical structure of the ambiguous-sound study: physically identical inputs, different outcomes, information visible only in the pattern. Success would show that pattern-level measurement recovers information that the single marker provably discards.
Baselines
The do-nothing reference is predicting the cohort mean late titer, or the majority durability class, for every subject; it fixes the floor and calibrates every accuracy claim. The strongest existing method is the one already run on this data: Elastic Net regression from early response patterns plus sex to post-vaccination antibody titers, as in Acute and persistent responses after H5N1 vaccination in humans — no reported result found on Lacuna for its prediction accuracy in that paper, so we reimplement it under an identical feature budget, split structure, and tuning via nested cross-validation, and re-establish its performance on the H5N1 data ourselves. Beating it means higher cross-validated balanced accuracy on the durability tercile task with non-overlapping bootstrap confidence intervals, or — the likelier and still publishable outcome — matched prediction accuracy plus the generalization structure that regression cannot produce. The second competitor is the best single marker, peak early antibody titer, whose strength is documented by the COVE correlates analysis, where a discrete Super Learner ensemble over antibody assays added nothing over the single best marker and a ten-fold increase in Day-29 live-virus microneutralization titer carried a hazard ratio of 0.39 (Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial); for any risk-prediction readout, the number to beat is the Day-29 univariate Cox model, anchored by the ENSEMBLE hazard ratio of 0.35 per ten-fold neutralizing-titer increase reported by Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial. Finally, the key ablation isolates the proposal's ingredient: a within-time-only decoder with no cross-time evaluation, and a second ablation that removes the batch-covariance normalization before cross-time testing. If off-diagonal transfer vanishes without the normalization, the "stable code" was a batch artifact, and the paper must say so.
Why not just…?
Why not just run Elastic Net or Super Learner on early features and predict late titers? For pure point prediction, this is a real threat, and we run it head-to-head rather than argue around it. But two things limit it. The COVE analysis already showed that an ensemble over a small, collinear assay panel gains nothing over the best single marker (Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial), so the burden is on any multivariate method to show where the extra information lives; and a static regression, whatever its accuracy, is silent on whether the Day-7 and Day-100 states are the same representation or different ones — the question the generalization matrix answers, and the one that determines which blood draws can be dropped. If Elastic Net matches our prediction accuracy, the proposal's claim shifts to the temporal structure, which is untouched by that baseline.
Why not just use the best univariate correlate, as the trial literature already does? Because the trial literature itself has quantified that marker's ceiling: neutralizing antibodies mediate only 28.6% of severe-disease protection in ENSEMBLE, with 63.1% efficacy persisting at unquantifiable titers (Neutralizing antibody correlate of protection against severe-critical COVID-19 in the ENSEMBLE single-dose Ad26.COV2.S vaccine efficacy trial), and PREVENT-19's titer-based efficacy curve rests on only 12 breakthrough cases (Immune correlates analysis of the PREVENT-19 COVID-19 vaccine efficacy clinical trial). The titer-matched CoVPN 3008 experiment is designed precisely to settle this: if pattern information survives matching on the univariate marker, the single-marker approach is demonstrably lossy; if it does not, the protection floor lives in compartments none of these datasets measure, and we will report that honestly.
Why not just model the trajectories statistically, with dynamical correlation or functional data analysis? Dynamical Correlation for Multivariate Longitudinal Data estimates correlations between pairs of longitudinal curves without collapsing them to slopes, and it is the right tool for a question of the form "do these two trajectories move together." But it is pairwise, produces no decision boundary, no decoding accuracy, and no train-here-test-there structure, so it cannot distinguish a stable code from a relay, and it has never been applied to high-dimensional multi-omic vaccine data. We include it as a baseline for any trajectory-similarity claim we make, and expect it to complement rather than substitute for the generalization matrix.
Decisive Checks
The proposal stands or falls on three outcomes. The generalization matrix must show structure that survives batch-held-out splits and permutation testing — relay, stability, or a clean null, each interpretable. The sampling-compression analysis must identify a strictly smaller time-point set with no loss of late-state prediction, converting a methodological claim into a protocol. And the titer-matched decoding must either demonstrate pattern information beyond the univariate marker or localize the protection floor outside the measured feature space — both of which would change how the next correlates-of-protection study is designed.
References
- 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 PREVENT-19 COVID-19 vaccine efficacy clinical trial
- Auditory Cortex Encodes the Perceptual Interpretation of Ambiguous Sound
- How face perception unfolds over time
- Acute and persistent responses after H5N1 vaccination in humans
- Longitudinal Analysis of Humoral and Cellular Immune Response up to 6 Months after SARS-CoV-2 BA.5/BF.7/XBB Breakthrough Infection and BA.5/BF.7-XBB Reinfection
- Long-term cellular immunity of vaccines for Zaire Ebola Virus Diseases
- Removal of Scanner Effects in Covariance Improves Multivariate Pattern Analysis in Neuroimaging Data
- EEG-ImageNet: A Benchmark for Pre-training and Cross-Time Generalization of EEG-based Visual Decoding
- A Toolbox for Representational Similarity Analysis
- A Python Toolbox for Representational Similarity Analysis
- Neutralizing and binding antibodies are a correlate of risk of COVID-19 in the CoVPN 3008 study in people with HIV
- Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial
- Dynamical Correlation for Multivariate Longitudinal Data