Two laboratories can measure the same blood serum's virus-neutralizing antibodies and report answers that differ by a factor of three to five. When Huang et al. ran identical serum panels through the Duke and Monogram pseudovirus neutralization assays, Monogram readouts ran roughly 2.8-fold higher on ID50 and 5.2-fold higher on ID80, and this kind of divergence is what blocks "immunobridging" — the regulatory practice of inferring a new vaccine's efficacy by comparing its measured immune response against a benchmark calibrated in a previous trial, rather than running a fresh placebo-controlled trial. Meanwhile, a separate literature in fair machine learning has spent years solving what is essentially the same mathematical problem from the other side: when you adjust scores across demographic groups to satisfy a group-level fairness criterion, how do you guarantee you never scramble the ordering of individuals within a group? The tools built there — monotone quantile matching, Wasserstein barycenters, constrained score transformation — are exactly the machinery immunobridging lacks, and no one has connected them (Chzhen et al. frame everything in terms of recidivism, credit, and crime rates; Huang et al. calibrate with parametric regression and never test whether individual orderings survive).
The core move is a change of variables: treat the assay platform as the "group" attribute, and treat the requirement that a calibration never reorder patients within a laboratory cohort as the intra-group rank preservation constraint. One caveat that keeps this honest: any monotone calibration, including a simple linear rescaling, preserves within-assay rank, so monotonicity alone is not the contribution. The contribution is time-conditioned, censoring-aware quantile alignment: the calibration map must depend on time since vaccination (because antibody affinity maturation makes assays drift apart — Kroidl et al. showed WHO-standard conversion holds for only about 80 days post-infection, after which correction factors of up to 10 are needed), and it must handle readouts censored at assay detection limits, which neither the fairness nor the calibration literature currently does.
Core Claim
Every (assay, time-point) pair defines its own arbitrary measurement scale. The claim is that the right harmonized quantity is not a converted titer but a percentile on a shared reference distribution, computed by a monotone map
$$\hat{y}^{\,a,t}_{\text{ref}} \;=\; F_{\text{ref}}^{-1}\!\big(F_{a,t}(\hat{y})\big),$$
where $F_{a,t}$ is the cumulative distribution of assay $a$'s readouts at time $t$ after vaccination and $F_{\text{ref}}$ is a reference distribution, taken either as the Wasserstein barycenter (the distribution whose average transport cost to all assay-conditional distributions is minimal) across platforms or anchored to the WHO International Standard. This is a direct transplant of the fairness construction of Chzhen et al., who proved that mapping group-conditional score distributions to a barycenter enforces distributional parity while preserving within-group percentiles exactly. The transplant requires three extensions that are open problems even inside the fairness literature: conditioning the quantile map on a continuous variable (time), estimating the conditional CDF under left/right censoring at detection limits (using their "jittering trick" for ties extended to censored empirical quantiles), and choosing the reference distribution robustly — since Cui et al. proved the barycenter is strictly suboptimal for any loss other than squared error, and titer distributions are skewed and heavy-tailed, the target should be learned with monotone I-splines under a robust loss rather than assumed to be the barycenter.
A second component extends the same constraint to multi-marker aggregation. In the COVE mRNA-1273 trial, unconstrained ensemble models combining binding antibodies, pseudovirus neutralization, and live-virus microneutralization failed to improve risk prediction over the single best marker, PsV-nAb ID50 (Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial). I propose testing whether a composite protective index built under a monotonicity constraint in each harmonized marker — the mechanism Wei et al. developed as FairScoreTransformer, which solves a constrained dual optimization with a proven monotonicity guarantee per group — can beat the best single marker because the constraint removes the assay-scale noise and ordering scrambling that plagued the unconstrained ensemble. This is a genuine hypothesis that may fail; the experiments are designed to settle it either way.
Research Program
Three things get built. First, the time-conditioned quantile alignment estimator: per assay and per visit window, estimate $F_{a,t}$ from pooled cohort data with censoring-aware empirical CDFs, jittering tied and truncated readouts as in Chzhen et al., then learn the reference quantile function with monotone I-splines under Huber or LAD loss following Cui et al., whose spline construction guarantees strict rank preservation by design. Second, an error model: both fields report substantial measurement noise — Huang et al. estimate about 30% laboratory measurement error — and raw empirical quantiles conflate biological spread with assay noise, so the estimator will deconvolve a reported noise level before matching quantiles, connecting to the error-in-variables logic of their bivariate normal model. Third, the constrained composite index: a dual-space optimization in the style of Wei et al. that combines the harmonized markers into one protective score, monotone in each input, so no patient's composite standing can flip because one assay's scale was noisy.
The theory deliverables mirror the fairness results: a proof that the time-conditioned map preserves within-cohort order at every fixed time point (immediate from monotonicity, but worth stating with the censoring and jittering extensions); a characterization of when the barycenter reference is suboptimal for titer-like skewed distributions, extending the suboptimality result of Cui et al.; and a feedback contribution to fairness itself, since conditioning quantile alignment on a continuous variable (time here, any continuous covariate there) is listed as an open problem in the fair regression line.
Baselines
The "do nothing" reference point is raw, unharmonized titers compared across platforms. Its documented failure is the motivation for the field: Monogram reads ~2.8-fold high on ID50 and ~5.2-fold on ID80 relative to Duke on identical sera (Huang et al.), and BAU-standardized readouts from two commercial assays diverge by factors up to 10 past day 80 post-infection (Kroidl et al.). Any proposed method must at minimum collapse this disagreement on held-out samples.
The strongest existing competitor is the bivariate normal calibration model of Huang et al., which fits a joint log-normal distribution to paired assay readouts while explicitly estimating the ~30% measurement error in each laboratory; it achieved a Concordance Correlation Coefficient of 0.87 for ID50 between Duke and Monogram, the best of the three approaches they tested. It will be run on identical train/test splits of the same paired panels, with the same noise assumptions, and given the same tuning budget (an equal grid over spline knots and loss choices for my method, error-model parameters for theirs). Beating it means a higher held-out CCC on the Duke–Monogram panels with non-overlapping bootstrap intervals — but, honestly, on cross-sectional log-normal-ish data it may well match a quantile map, so the decisive comparison is the longitudinal and censored regimes where its static, parametric form has documented failure modes. Two further baselines anchor the regulatory-status-quo end: multiplicative WHO-standard scaling (Approach 1 of Huang et al., which performed nearly as well as their bivariate model cross-sectionally but fails over time per Kroidl et al.), and convalescent-serum ratio centering, which divides each study's titers by the geometric mean of a convalescent panel assayed in the same laboratory (Khoury et al.; Evidence for antibody as a protective correlate for COVID-19 vaccines) and which reconciled seven platforms at the group level ($\rho = 0.79$ for neutralization, $\rho = 0.93$ for binding IgG against efficacy) while leaving individual-level scale differences untouched. Finally, two ablations isolate the proposal's key ingredients: a static quantile map versus the time-conditioned one (isolating the temporal component), and an unconstrained Super Learner versus the monotonicity-constrained composite (isolating the aggregation constraint).
Why not just…?
Why not just use the WHO standard and multiply? Because the file's own evidence shows this works only in a narrow regime. Huang et al. found scalar conversion nearly as good as structural calibration for cross-sectional Duke–Monogram panels, but Kroidl et al. tracked the same WHO-standardized assays for 18 months and found equivalence only within roughly 80 days of infection, with required corrections growing to 10-fold as avidity maturation made one assay decay while the other stayed flat. A scalar cannot represent a time-varying, non-linear offset; the longitudinal experiment below is designed precisely to quantify this failure against the time-conditioned map.
Why not just fit a regression or bivariate normal model? This is the real threat, and I will not argue around it. The bivariate normal model already achieves CCC = 0.87 on the cross-sectional panels (Huang et al.), and if titer pairs are close to bivariate log-normal there, a quantile map may not beat it there. The proposal instead stakes its case on three regimes where the parametric form is documented to break or is untested: longitudinal drift (Kroidl et al.), censoring at detection limits (values pile up below limits like the < 2.612 IU50/mL floor in the AZD1222 analysis, and naive regressions are biased by this clustering), and heavy tails, where Cui et al. showed $L_2$-style methods are fragile and monotone spline transforms reduced unfairness from 0.3856 to about 0.01 on a heavy-tailed benchmark while linear moment-matching could not. If the bivariate model matches the quantile map in all three regimes, the proposal's contribution collapses to guarantees rather than accuracy — and the experiments are constructed so that this outcome would be visible and reportable.
Why not just divide by the convalescent-panel mean? Ratio centering is cheap and remarkably effective at the group level — it is what unified seven platforms into one sigmoidal protection curve, with 50% protection at roughly 20% of the convalescent mean titer (Khoury et al.). But it aligns only the first moment: two assays with the same mean and different variance, saturation, or censoring behavior remain misaligned at every other percentile, and convalescent panels themselves differ across laboratories in severity and sampling. The cross-platform concordance experiment measures exactly what moment-matching leaves on the table, percentile by percentile.
Why not just z-score within each assay? Standardizing by group mean and standard deviation preserves rank within an assay and is trivially cheap — Rosenblatt & Witter showed exactly this kind of simple normalization beat an elaborate causal model on the Law School data (root mean squared error 0.922 versus about 0.936) while avoiding the causal model's chaotic reordering. But it assumes the only differences between scales are location and spread. Chzhen et al. showed full distribution matching reduces distributional disparity (Kolmogorov–Smirnov distance) by 50–80% beyond what moment alignment achieves, and titer distributions with censoring and saturation are precisely the skewed case where two moments are not enough. The z-score will be included as an ablation to make this comparison concrete rather than rhetorical.
Proposed Experiments
The first experiment uses the paired-panel setting of Huang et al.: the WHO International Standard, 248 convalescent sera, and 90 mRNA-1273 Phase 1 vaccinee sera, each measured on both Duke and Monogram platforms. Where individual-level data cannot be shared, paired samples will be simulated to match every reported statistic (the 2.8-fold and 5.2-fold offsets, the 30% error, the 0.87 CCC target) and validated against the published aggregates — this limitation is stated plainly. Metric: held-out CCC on ID50 and ID80, plus a new metric the field lacks, cross-platform rank agreement (Kendall's $\tau$ between the two platforms' harmonized orderings of the same patients). The bar: match or exceed CCC = 0.87, with the rank-agreement metric reported for all baselines since none of them has ever been evaluated on it.
The second experiment is the decisive one: longitudinal harmonization on the Kroidl et al. cohort structure — 99 convalescent individuals followed up to 18 months and 45 vaccinees, measured on both the Euroimmun total-binding assay and the Roche avidity-sensitive assay. Static methods (WHO scalar, bivariate normal, static barycenter) are fit on the ~80-day equivalence window and then evaluated out-of-window; the time-conditioned quantile map is fit on the same window and evaluated identically. The claim to beat is qualitative but sharply documented: static corrections require up to 10-fold adjustments past day 80. Success means held-out cross-assay CCC that stays roughly flat across time bins where every static baseline degrades, plus within-patient coherence: an individual's percentile trajectory should not depend on which assay is used to compute it.
The third experiment tests the composite-index hypothesis on trial correlate data. Using the case-cohort structures of the AZD1222 trial (Immune correlates analysis of a phase 3 trial of the AZD1222 (ChAdOx1 nCoV-19) vaccine: hazard ratio 0.28 per 10-fold ID50 increase, VE rising from 64.9% at 10 IU50/mL to 94.2% at 270 IU50/mL), ENSEMBLE (Fong et al.: hazard ratio 0.49 per 10-fold ID50), and COVE (Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial: hazard ratio 0.39 per 10-fold LV-MN50 at Day 29), the question is whether harmonizing markers onto the shared percentile scale before aggregation — under the monotonicity constraint — yields a composite that predicts breakthrough risk better than the best single marker, which unconstrained combination failed to do in COVE. The decisive check is a head-to-head against the reported finding that combining markers did not improve over PsV-nAb ID50; if the constrained composite also fails to improve, that is a publishable negative result that still validates the harmonization layer, since the single-marker comparisons themselves become cross-platform comparable for the first time.
A closing sanity experiment runs the machinery back on its home turf: the CRIME regression benchmark from Cui et al., with synthetic censoring and a continuous conditioning variable added, to confirm the time-conditioned, censoring-aware extensions do not degrade the static case (their reported unfairness reduction from 0.3856 to about 0.01 at the $\lambda = 0.52$ Pareto knee is the anchor). The through-line is simple to state to a colleague: assays are groups, percentiles are the only currency every platform accepts, and the fairness literature already proved how to exchange that currency without ever reordering the people inside it.
References
- Calibration of two validated SARS-CoV-2 pseudovirus neutralization assays for COVID-19 vaccine evaluation
- Fair regression with Wasserstein barycenters
- Studying temporal titre evolution of commercial SARS-CoV-2 assays reveals significant shortcomings of using BAU standardization for comparison
- A General Framework for Fair and Robust Regression
- Comparing antibody assays as correlates of protection against COVID-19 in the COVE mRNA-1273 vaccine efficacy trial
- Optimized Score Transformation for Consistent Fair Classification
- Correlates of Protection, Thresholds of Protection, and Immunobridging among Persons with SARS-CoV-2 Infection
- Evidence for antibody as a protective correlate for COVID-19 vaccines
- Counterfactual Fairness Is Basically Demographic Parity
- Immune correlates analysis of a phase 3 trial of the AZD1222 (ChAdOx1 nCoV-19) vaccine
- Immune correlates analysis of the ENSEMBLE single Ad26.COV2.S dose vaccine efficacy clinical trial