Cell-Resolved Correlates of Protection from Bulk Vaccine Trial Archives

When a vaccine efficacy trial measures immune responses, it usually does so on blood samples in which millions of cells of many types are ground together and read out as one averaged gene-expression profile. Studies doing exactly this have found genuinely predictive signals — for example, an antiviral interferon program that associates with reduced HIV acquisition in the RV144 trial — but the averaging hides which cells actually carry that program. Is the protective signal coming from plasmacytoid dendritic cells, from monocytes, or from rare antigen-specific T cells? Nobody can tell, because the measurement is a smoothie. Worse, the archived trial samples that could answer this were sequenced at different centers with different protocols, so non-biological "batch" artifacts — systematic distortions from reagent lots, machines, and handling — sit entangled with the biology in the same smoothed average. And the one computational tool that could un-mix the average, deconvolution, assumes each cell type looks like a fixed reference profile, which vaccination itself violates: immunization switches on interferon-stimulated genes inside many cell types at once, so the reference no longer matches the cells.

I propose to separate these three confounded unknowns — how many of each cell type, what state each type is in, and which lab ran the assay — directly from the bulk trial archives, and to use the result to build what I call cellular immunobridging: the extension of a calibration trick that regulators already trust for antibody titers to high-dimensional, cell-resolved features.

Core Claim

The claim is that a single generative decomposition, fitted to existing bulk RNA data plus an external single-cell reference atlas, can recover cell-type-resolved correlates of protection that bulk analyses find only in blurred form, and that these features transfer across trials and sequencing centers where raw bulk features fail. Concretely, for participant $i$ sequenced at center $s(i)$, model the observed gene-count vector as

$$y_i \;\approx\; \sum_{k=1}^{K} c_{ik}\,\bigl(x_k + \delta_{ik}\bigr) \;+\; \beta_{s(i)},$$

where $c_{ik}$ is the proportion of cell type $k$ in participant $i$'s sample, $x_k$ is that type's expression profile in a single-cell reference atlas, $\delta_{ik}$ is a sparse within-person state shift of type $k$ (the vaccine-induced interferon activation, metabolic reprogramming, and so on), and $\beta_{s(i)}$ is a batch field shared by every sample processed at the same center. The decomposition is identifiable for a structural reason, not a hopeful one: $\beta$ is common to all participants at a center while $c$ and $\delta$ vary person to person within that center, so within-center variation isolates the biology; and placebo-arm samples, which by design carry no vaccine-induced state shift, pin $\delta$ to zero and anchor the batch field. This anchoring is the direct descendant of the idea that let Khoury et al. collapse titers from labs reporting up to 50-fold discrepancies onto one universal protection curve by normalizing each study to a shared convalescent reference panel — except here the "titer" is a vector of latent cellular features and the reference panel is the placebo arm plus, prospectively, a pooled reference RNA aliquot sequenced at every center.

Prior Work and the Gap

Three lines of work motivate and constrain this design. First, systems vaccinology has already found bulk transcriptional correlates but cannot localize them: Fourati et al. showed that an IRF7/Type I interferon program correlated with reduced HIV acquisition risk in RV144 (IFN-γ signaling odds ratio 0.883) while mTORC1 (OR 1.33) and NF-κB (OR 1.24) signaled increased risk, and that adding bulk pathways to antibody markers raised balanced accuracy only from 62.3% to 67.9% — a modest gain with zero attribution to cell types. Andersen-Nissen et al. got one step further in HVTN 097, applying coarse deconvolution and observing a Day-1 monocyte expansion in high responders, but could not resolve monocyte subsets or distinguish composition change from within-cell activation, and used static references blind to batch. Second, deconvolution methodology has advanced without ever touching vaccine endpoints: SCDC handles the case of multiple single-cell references each carrying its own lab artifacts by deconvolving with each independently and ensemble-weighting the results, but its references are static and it has never been run on trial cohorts; POPPY maps bulk tumors into a single-cell latent space with cross-attention attribution, but only in oncology and without any batch-correction mechanism or count-level fidelity; Count Bridges generates integer-valued single-cell count profiles that sum exactly to observed bulk counts, but has no cross-study adaptation. Third, batch-correction methods like BERMUDA and scAdapt are strong but operate only on single-cell inputs, leaving bulk archives untouched. No existing study connects all three directions; the nearest neighbors each miss one leg of the tripod.

Research Program

The plan has three parts. The first is the model itself. I will reuse the discrete generative core of Count Bridges — a stochastic bridge process on integer counts driven by Poisson birth–death dynamics, with an expectation-maximization step forcing reconstructed cells to sum to the observed bulk counts — because sequencing reads are integers and pretending otherwise distorts the count geometry. On top of that core I add the $\delta$ state-shift term (sparsity enforced at the pathway level, so a person's monocytes can be "interferon-on" without the model inventing thousands of free parameters) and the $\beta$ batch field, trained with a domain-adversarial objective in the spirit of scAdapt: a gradient-reversal module tries to predict the sequencing center from deconvolved features and the encoder tries to prevent it, while the placebo-anchor constraint keeps the adversary from erasing genuine biology. BERMUDA's cluster-level caution — align only matched populations so that a genuinely novel activated state is not merged away — becomes here a constraint that $\beta$ may not alter reference profiles $x_k$, only the observation of them. Alongside the model I will work out the identifiability conditions in the usual asymptotic sense: how large a placebo arm, and how much marker-gene overlap between reference and bulk panels, are needed for the three factors to be separable.

The second part is a ground-truth benchmark suite, because the failure modes I care about have no published numbers. The third part is the clinical reanalysis of RV144 (223 vaccinees, 40 placebo recipients, infection endpoints) and HVTN 097 (30 participants, longitudinal Days 0–7 and Months 6.5–12), culminating in a cross-trial transfer experiment: HVTN 097 used the same ALVAC-HIV regimen as RV144 in a different country and different sequencing context, which makes it the cleanest existing test of whether cell-resolved correlates immunobridge.

Baselines

The "do nothing" reference point is the analysis actually published: an antibody-only risk model and a bulk-pathway model on RV144, where Fourati et al. report balanced accuracy of 62.3% with antibodies alone and 67.9% after adding bulk transcriptomic pathways. I will rerun both under identical cross-validation folds so the comparison is fold-matched, and the bar for the proposal is to exceed 67.9% using cell-resolved features on the same data.

The strongest existing competitor is SCDC ensemble deconvolution (SCDC), which is the only published method explicitly designed for the reference-batch problem: it reports a Pearson correlation of r = 0.991 with ground-truth proportions on an in vitro mixture of three cell lines mixed 6:3:1, outperforming CIBERSORTx, MuSiC, and Bisque, and it was the only tool to recover the known beta-cell/HbA1c association in 77 pancreatic islet bulk samples (p = 0.0019). It is the strongest competitor precisely because it attacks the same reference-batch confound, and because its accuracy ceiling on matched data is high enough that beating it requires the new ingredients rather than just tuning. I will run SCDC on the same trial data with the same single-cell references, the same compute, and the same tuning budget, then feed its proportions into the identical downstream risk model, so the only difference between pipelines is the unmixing step. Beating it means three things: higher proportion accuracy when the reference and target deliberately come from different platforms, higher balanced accuracy on the RV144 case-control folds, and correct attribution on the state-shift benchmark — where SCDC, having no $\delta$ term at all, must express a vaccine-induced within-cell activation as a spurious proportion change, a predicted failure I will measure rather than assert. I will additionally run POPPY retrained on immune references and Count Bridges as secondary baselines, and three ablations of my own model: one with $\delta$ removed (static references, isolating the state-shift ingredient), one with $\beta$ removed (isolating batch calibration), and one with the placebo anchor removed (testing whether the anchoring, not the architecture, does the work).

Why not just…?

Why not just run ComBat and then CIBERSORTx or MuSiC? This is the pipeline a reviewer will reach for first, and the evidence says it breaks at two joints. Global statistical shifts do not remove platform artifacts that are non-linear and cell-type-specific — Warnat et al. showed that classifiers trained on one microarray platform and tested on another perform at random chance without proper normalization (and above 85% with rank-based methods), and single-cell technical noise is worse: fragile lymphocytes and robust monocytes are distorted differently. And the SCDC benchmarking showed CIBERSORTx and MuSiC degrade specifically when the single-cell reference carries lab bias. My second experiment measures this pipeline's failure directly on ground-truth mixtures, so the claim will be empirical, not rhetorical.

Why not just train an end-to-end classifier on the bulk profiles? This is a real threat and I want to be honest about it. With carefully center-balanced cross-validation, an elastic net on raw bulk genes might match my predictive accuracy, and if it does — and also survives the leave-one-center-out transfer test — then the proposal's value shrinks to interpretation, and I will report that. But two pieces of evidence say it will not survive: pooled models learn the study center rather than the phenotype (Warnat et al.), and the best bulk-only result on RV144 data gained just 5.6 points of balanced accuracy while leaving the IRF7 signal unattributed (Fourati et al.). The transfer experiment and the attribution output are the two deliverables end-to-end prediction cannot provide, and both are needed for vaccine reformulation decisions.

Why not just measure the cell subsets with flow cytometry? For future trials, partially yes — but the irreplaceable material here is archival. RV144-era biobanks consist largely of frozen bulk lysates and extracted RNA; viable cells for fresh staining are exhausted. And flow panels are predefined sets of 10–20 markers that could not have discovered the unanticipated IRF7 or mTORC1 programs that unbiased transcriptomics surfaced. Where contemporaneous flow data does exist in these trials, I will use it as an independent check on my recovered proportions, which strengthens rather than competes with the proposal.

Why not just normalize transcriptomes the way Khoury and Earle normalized titers? Dividing by a reference panel's mean works for a scalar titer — Earle et al. got a correlation of ρ = 0.93 with efficacy across seven platforms by calibrating binding antibody levels to convalescent sera — because a one-dimensional quantity has one calibration constant. A gene-expression vector has thousands of distortions that differ by gene and cell type, including zero-inflation dropout (BERMUDA). So I adopt their anchoring logic, not their arithmetic: the anchor lives in the latent space of the decomposition, which is exactly the piece that is new.

Proposed Experiments and Decisive Checks

The first experiment stress-tests reference mismatch on the in vitro ground-truth benchmark of three cell lines mixed 6:3:1, where SCDC reports r = 0.991 with matched references. I will first replicate the matched condition as a sanity check, then inject realistic platform shifts — learned as transformation operators from the 7-platform PBMC single-cell benchmark used by scAdapt — between reference and target, and measure proportion RMSE and correlation. No reported result found on Lacuna for deconvolution accuracy under reference–target platform mismatch, which is precisely why this benchmark must be built.

The second experiment tests the state-shift confound on the 38,000 simulated breast-cancer mixtures benchmark of Cabrera et al., where BayesPrism and DWLS were the top performers for granular immune lineages. I will inject a coherent activation program into one lineage, mimicking systemic interferon stimulation, and measure how much of it each method misreports as a composition change. The ablation without $\delta$ should fail identically to the static baselines; that contrast isolates the proposal's key ingredient.

The third experiment reanalyzes RV144 end to end. The placebo arm (40 participants) anchors the batch field; I will hold it out in one run to quantify how much the anchor buys. The scientific questions are concrete: is the protective IRF7 signal (IFN-γ OR 0.883 in Fourati et al.) carried by plasmacytoid dendritic cells, by composition, or by within-cell induction; and is the mTORC1 risk signal (OR 1.33) a monocyte-state effect? Decisive check: cell-resolved features must exceed the published 67.9% balanced accuracy ceiling on fold-matched evaluation. If the signal turns out to be pure composition, the $\delta$ recovery adds nothing and the model will say so — the design is falsifiable in either direction, and both answers inform adjuvant design.

The fourth experiment resolves the Day-1 monocyte expansion in HVTN 097 into classical versus intermediate inflammatory subsets, and tests whether the correlation between Day-1 serum CCL8/IFN-γ and Month-6.5 ADCP (R = 0.45–0.62, Andersen-Nissen et al.) is mediated by composition or by monocytic state — a question that paper explicitly could not answer.

The fifth experiment is cellular immunobridging itself: fit the correlate model on RV144's anchored latent features, transfer to HVTN 097 with only the batch field re-estimated (using its Day-0 baseline samples as the anchor), and predict Month-6.5 ADCP/ADCC from Day-1 features. Raw bulk features serve as the negative control and are expected to collapse across the platform gap, following the cross-platform evidence above. Success is a materially higher cross-trial correlation for anchored latent features than for any unanchored baseline — the high-dimensional analog of the single scalar curve that Khoury et al. showed seven vaccine platforms converge onto. The deliverable beyond the papers is a practical protocol: a pooled PBMC reference RNA aliquot sequenced alongside every batch at every center, the cellular counterpart of the convalescent serum panel, which would make this calibration routine rather than retrospective.