Does Shared HLA Make Immune Responses Portable Across Ancestries?

When two unrelated people encounter the same virus, their immune systems often converge on nearly identical T-cell receptors — the "public repertoire" phenomenon. This convergence is the reason repertoire-based diagnostics work at all: a deep-learning model can read the mix of T-cell receptor sequences in a blood sample and infer past infection or autoimmune disease, but only because disease-associated sequences recur across individuals. Every one of these models, however, has been trained and tested almost exclusively on people of European descent, and we do not know why. Two candidate explanations exist. Either convergence is driven by the HLA molecule — the genetically encoded platform that displays viral fragments to T cells and that is shared across all human populations — in which case a diagnostic trained on Europeans should travel well wherever the relevant HLA allele exists. Or convergence also depends on ancestry-specific background: germline variation in the receptor genes themselves, local haplotype structure, and population-specific pathogen histories, in which case these tools will fail silently in most of the world, the same way European-derived genetic diagnostic criteria already fail in ethnic minorities, as Jones et al. demonstrated for monogenic diabetes testing Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing.

This proposal treats convergent immune response as the core mechanism and population generalizability as both the open scientific question and the evaluation regime. We will decompose receptor convergence into an HLA-explained component and an ancestry-background component, and then use that decomposition to build the first repertoire classifier that is explicitly conditioned on HLA genotype and explicitly protected against ancestry confounding. The closest existing work stops short on all three fronts: Tan et al. linked germline SNPs to TCR-cluster abundances but only in 337,151 European-ancestry UK Biobank participants Interpretable GWAS by linking clinical phenotypes to quantifiable immune repertoire components; Xue et al. built a two-million-cell paired TCR atlas and showed that public clonotypes share HLA alleles, but did not stratify convergence by genetic ancestry A pan-disease and population-level single-cell TCRαβ repertoire reference; and the type 1 diabetes consortium showed that European risk HLA haplotypes sculpt convergent motifs, but never asked whether those motifs transfer to non-European HLA architectures Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood. Nobody has asked the direct question: do two individuals of different continental ancestries who carry the exact same HLA allele mount the same convergent receptor response?

Core Claim

The claim is decomposable into a testable biological statement and a testable engineering statement. Biologically, we claim that convergent receptor sharing across ancestries is largely, but not entirely, explained by HLA matching — and that the unexplained residue, the part carried by ancestry-specific germline background and exposure history, is large enough to break today's classifiers. Operationally, we will estimate this with a single interaction coefficient: for a pair of individuals $(i, j)$, let $s_{ij}^{m}$ indicate that both carry convergent motif cluster $m$ above a frequency threshold, let $h_{ij}$ indicate a shared restricting HLA allele for that motif, and let $d_{ij}$ measure genetic ancestry distance between the pair. We fit

$$\text{logit}\, P(s_{ij}^{m}) = \beta_0 + \beta_H h_{ij} + \beta_A d_{ij} + \beta_{HA}\, h_{ij} d_{ij},$$

where $\beta_H$ is the boost from HLA matching, $\beta_A$ is the baseline decay of sharing with ancestry distance, and the interaction $\beta_{HA}$ is the quantity of interest: it measures how much the value of an HLA match itself degrades when the two individuals come from different ancestral backgrounds. If $\beta_{HA} \approx 0$, convergence is HLA-driven and portable; if $\beta_{HA} < 0$ substantially, ancestry background restricts convergence even under identical HLA restriction, and every European-trained repertoire diagnostic is suspect outside Europe. The engineering claim follows directly: a classifier that conditions on HLA genotype and is trained to be invariant to ancestry-predictive background features should recover most of the cross-ancestry performance that a naive transfer loses.

This framing is deliberately borrowed from the instrumental-variable literature on ancestry bias, where a genetic instrument discovered in Europeans loses validity in other populations through linkage-disequilibrium breakdown and pleiotropic shifts Causal associations between risk factors and common diseases inferred from GWAS summary statistics. A convergent motif is, in effect, a somatic instrument for past antigen exposure; the question of whether it remains a valid instrument across ancestries is exactly the portability question that Mendelian randomization has already answered pessimistically for germline instruments. We expect the somatic case to be better — because receptor convergence is selected on the antigen, not on the population's haplotype structure — but nobody has measured it.

Research Program

The program has three components. First, we will construct the ancestry-annotated convergence atlas that does not yet exist. The raw material is the pan-disease single-cell TCR atlas of Xue et al., which covers 583 individuals with paired αβ receptors and HLA genotypes A pan-disease and population-level single-cell TCRαβ repertoire reference, augmented by the CMV serostatus benchmark of 684–685 bulk repertoires used by the repertoire-classification literature Modern Hopfield Networks and Attention for Immune Repertoire Classification and the T1D consortium's 2,250 repertoires with whole-genome SNP data Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood. Neither resource is ancestry-annotated in a usable form: the atlas does not stratify convergence by ancestral group, and the CMV benchmark repertoires carry no ancestry metadata at all. We will therefore infer genetic ancestry ourselves, using haplotype-based local ancestry inference at genome-wide and locus-specific resolution — the same machinery that recently showed immune-response loci are among the most heavily selected regions of the genome, with ancestry tracts varying within individuals even inside the HLA region Sparse haplotype-based fine-scale local ancestry inference at scale reveals recent selection on immune responses — and we will complement this with targeted new sequencing of roughly 600 CMV-serostatused repertoires from African, East Asian, and admixed-American donors with matched HLA typing and low-pass genotyping, so that the cross-ancestry arms are not limited by whatever incidental diversity the existing cohorts happen to contain.

Second, we will run the decomposition. Convergent sharing must be measured at the motif level, not the full-sequence level, because the relevant biology is degenerate: gluten-specific recognition is dominated by a short invariant R-motif inside an otherwise variable receptor loop Comprehensive Analysis of CDR3 Sequences in Gluten-Specific T-Cell Receptors Reveals a dominant R-Motif and Several New Motifs, and short amino-acid triplets track antigen-induced perturbations across outbred individuals whose full-length sequences appear completely private Tracking global changes induced in the CD4 T-cell receptor repertoire by immunization with a complex antigen using short stretches of CDR3 protein sequence, a result validated explicitly for degeneracy and privacy by Sun et al. Specificity, Privacy, and Degeneracy in the CD4 T Cell Receptor Repertoire Following Immunization. We will cluster receptors into motif groups using triplet and structural-similarity grouping in the spirit of Tan et al.'s Repertoire Functional Units Interpretable GWAS by linking clinical phenotypes to quantifiable immune repertoire components, annotate each cluster's restricting HLA where known (public CMV and EBV clonotypes, which the atlas showed are the most broadly shared, give the cleanest test cases A pan-disease and population-level single-cell TCRαβ repertoire reference, together with the public HCMV-specific β chains documented across unrelated transplant patients Relationship between T-cell receptor beta chain sequences and human cytomegalovirus infection in allogeneic hematopoietic stem cell transplant recipients), and fit the interaction model above over all pairs, HLA-matched and unmatched, within and across ancestries. The design is inspired by a finding that cuts across the two directions: Renaudineau et al. showed that the same HLA risk haplotype produces different disease outcomes in different ancestries because of background complement-gene variation Across ancestries, HLA-B∗08:01∼DRB1∗03:01 (DR3) and HLA-DQA∗01:02 (DR2) increase the risk to develop juvenile-onset systemic lupus erythematosus through low complement C4 levels; we are asking whether the same logic applies one step downstream, at the level of the selected receptor repertoire.

Third, we will build the portable classifier and the diagnosis of why naive transfer fails. The baseline architectures are the massive multiple-instance learning models that currently define the field — DeepRC, which retrieves rare disease-witness sequences from bags of hundreds of thousands of receptors using Hopfield-network pooling Modern Hopfield Networks and Attention for Immune Repertoire Classification, its noisy-label reformulation NLL-IRC A Noisy-Label-Learning Formulation for Immune Repertoire Classification and Disease-Associated Immune Receptor Sequence Identification, and the current strongest multimodal model LaDM³IL A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification. All three assume the training and test repertoires come from the same background distribution; our modification is twofold. We add an explicit HLA-conditioning input (the individual's typed HLA genotype, which the motif–HLA coupling demonstrated in T1D Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood suggests the model needs but never receives), and we add an adversarial objective that removes from the learned receptor-embedding any information predictive of genetic ancestry while preserving disease discrimination — so the classifier cannot score points via ancestry-correlated background k-mers that happen to associate with the label in the European training set. The public Hopfield-layers implementation ml-jku/hopfield-layers gives us DeepRC directly; the others will be reimplemented from their descriptions. As a secondary analysis, we will re-derive Tan et al.'s gdRFU procedure — Lasso models predicting TCR-cluster abundance from TRB- and HLA-locus SNPs — and project these genetic instruments onto our non-European genotyped cohorts, measuring first-stage instrument strength in the target ancestries; no reported result found on Lacuna for cross-ancestry gdRFU instrument F-statistics, which is precisely the hole this fills.

Baselines

The naive reference point is the burden test: count how many known disease-associated public clones appear in a repertoire and classify on the total. On the CMV benchmark this achieves a ROC-AUC of 0.699, as reported by Widrich et al. Modern Hopfield Networks and Attention for Immune Repertoire Classification — barely better than chance given how sparse the witness signal is, and it is exactly the method that would be deployed first in a new population because it requires no training. Any cross-ancestry method we propose must beat this comfortably in every target ancestry, or the exercise is pointless.

The strongest existing competitor is the current state of the art evaluated under the transfer protocol it has never faced: LaDM³IL, which reports a CMV AUC of 84.88% in the standard same-distribution setting A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification, and DeepRC, which reports ROC-AUC 0.831 on the same CMV benchmark Modern Hopfield Networks and Attention for Immune Repertoire Classification and, tellingly for our setting, AUROC 0.79 in its discovery T1D cohort and 0.77 on held-out T1D cohorts Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood — a held-out drop measured across cohorts that were ancestrally homogeneous, meaning that drop reflects batch and sampling noise alone, and any additional drop we measure across ancestry is attributable to ancestry itself. We will train both models on the European arm of our assembled data with their published architectures and a matched hyperparameter-search budget, then evaluate zero-shot on each non-European arm. Beating this baseline means achieving cross-ancestry AUC strictly above the transferred LaDM³IL and DeepRC in every target ancestry, and within roughly 0.03 of the within-ancestry retrained upper bound, which we obtain by training and testing the same architecture entirely within each target ancestry; that upper bound also tells us how much of any remaining gap is a data-volume problem rather than a bias problem.

The key ablation is our own architecture stripped of its two additions, run in two configurations: HLA conditioning removed but the ancestry-adversarial term kept, and the reverse. If the full model closes the transfer gap but the HLA-stripped variant does not, the portability problem is a missing-covariate problem and the interaction term $\beta_{HA}$ should come out near zero. If the adversarial term is what matters, the problem is that the baselines were quietly exploiting ancestry-correlated background sequence, a finding that would matter well beyond this proposal.

Why not just…?

Why not just run multi-ancestry GWAS and polygenic scores, skipping repertoire sequencing entirely? This is the cheapest alternative and it fails on the file's own evidence. Conventional GWAS is blind to signals that receptor clustering reveals: Tan et al. found four TCR clusters associated with better cancer survival in lung, brain, ovarian, and lymphoma cancers "where conventional GWAS cannot identify any significant locus" Interpretable GWAS by linking clinical phenotypes to quantifiable immune repertoire components, and convergent T1D motifs emerge in autoantibody-positive individuals before clinical onset, capturing ongoing immune activation that static germline genotyping cannot Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood. Moreover, the polygenic-score route inherits the very problem we study: Jones et al. proved that diagnostic criteria derived from White populations fail in ethnic minorities for diabetes Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing. That said, germline genetics is not our rival but our substrate — we need genotyping for local ancestry inference, and the rfuQTL portability analysis directly imports the instrument-validity toolkit from that literature Causal associations between risk factors and common diseases inferred from GWAS summary statistics.

Why not just compare repertoire diversity or V-gene usage across ancestries? This is the obvious first analysis and we will run it as a sanity check, but the evidence says it is uninformative about disease and likely about ancestry-driven convergence too. The T1D consortium explicitly reported no repertoire-level diversity or similarity differences between patients and controls — discrimination was achievable only through subtle HLA-associated motifs extracted by deep learning Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood — and the burden test, the closest analogue among frequency-based metrics, manages only 0.699 ROC-AUC on CMV Modern Hopfield Networks and Attention for Immune Repertoire Classification. If macro metrics showed large ancestry differences, that would itself be an interesting finding, but we expect the ancestry signal, like the disease signal, to live in convergent motif space.

Why not just count identical public clonotypes shared between European and non-European cohorts? Full-length sequence matching is cheap and interpretable, but it systematically underestimates convergence because functional equivalence does not require sequence identity: the celiac R-motif work shows recognition is carried by a short invariant motif inside a variable loop Comprehensive Analysis of CDR3 Sequences in Gluten-Specific T-Cell Receptors Reveals a dominant R-Motif and Several New Motifs, and the triplet-tracking studies show that individuals with completely private full-length repertoires still share degenerate triplet motifs Tracking global changes induced in the CD4 T-cell receptor repertoire by immunization with a complex antigen using short stretches of CDR3 protein sequence. A full-length matching study could therefore report "no cross-ancestry sharing" and be wrong. We will run exact matching as a lower bound alongside the motif-level analysis; a large gap between the two estimates within HLA-matched pairs is itself a headline result.

Why not just retrain a separate classifier for each ancestry? If abundant labeled, ancestry-matched training data existed for every population, per-ancestry retraining would probably work as well clinically, and we say so plainly. It does not exist: disease-associated witness clones occur at rates around 0.01% of a repertoire Modern Hopfield Networks and Attention for Immune Repertoire Classification, which is why even the strongest models rely on data-hungry deep architectures, and LaDM³IL's own robustness result — maintaining accuracy at 10–20% of the original dataset A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification — concerns downsampling within one distribution, not a shift to a new ancestry where the background k-mer frequencies move. Retraining per ancestry also abandons the scientific question, which is whether convergence is HLA-universal, and abandons any population too small to field its own cohort. Our experiments settle this empirically: the within-ancestry retrained upper bound directly measures how much data a per-ancestry model would need before it matches the transferred conditioned model.

Proposed Experiments and Decisive Checks

Experiment one measures the interaction coefficient. Using the ancestry-annotated atlas plus the newly sequenced multi-ancestry cohort, we estimate $\beta_{HA}$ with cluster-level and pairwise bootstrap confidence intervals, separately for viral public clonotypes (CMV, EBV, influenza — the most broadly shared responses in the atlas A pan-disease and population-level single-cell TCRαβ repertoire reference) and for autoimmune motif clusters. The decisive readout: if the confidence interval on $\beta_{HA}$ straddles zero with a tight interval, HLA matching fully accounts for cross-ancestry convergence and the portability thesis holds; if $\beta_{HA}$ is materially negative, we quantify exactly how much of an HLA match's value is lost across ancestry, per motif and per HLA allele.

Experiment two is the transfer test. DeepRC, NLL-IRC, and LaDM³IL are trained on the European CMV arm and evaluated zero-shot on each non-European arm, reporting AUC against their published same-distribution anchors of 0.831 Modern Hopfield Networks and Attention for Immune Repertoire Classification, 83.24 A Noisy-Label-Learning Formulation for Immune Repertoire Classification and Disease-Associated Immune Receptor Sequence Identification, and 84.88 A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification, with the T1D held-out drop from 0.79 to 0.77 Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood as the benchmark for how much drop occurs even without ancestry shift. Our HLA-conditioned, ancestry-adversarial model runs under the identical protocol and tuning budget. Success is cross-ancestry AUC above every transferred baseline and near the within-ancestry upper bound; failure modes are informative either way, because the ablations attribute the gap to missing HLA conditioning versus background-distribution exploitation.

Experiment three is the instrument-portability audit. Re-derived gdRFU Lasso predictors, built on the European TRB- and HLA-locus SNPs following Tan et al.'s procedure Interpretable GWAS by linking clinical phenotypes to quantifiable immune repertoire components, are projected onto our non-European genotypes, and we report first-stage F-statistics per RFU per ancestry against the standard weak-instrument threshold of 10. We expect widespread failure of the tag-SNP structure, mirroring the LD-breakdown failure mode documented for conventional Mendelian randomization instruments Causal associations between risk factors and common diseases inferred from GWAS summary statistics; no reported result found on Lacuna for any cross-ancestry gdRFU evaluation, so any systematic measurement here is new. The salvage path we test is direct: imputing HLA genotypes and using HLA alleles themselves, rather than tagging SNPs, as the instruments — consistent with the cross-ancestry conservation of specific HLA haplotypes shown for juvenile lupus Across ancestries, HLA-B∗08:01∼DRB1∗03:01 (DR3) and HLA-DQA∗01:02 (DR2) increase the risk to develop juvenile-onset systemic lupus erythematosus through low complement C4 levels.

The decisive checks that would falsify our claims are, first, a tight-zero $\beta_{HA}$ combined with a large zero-shot transfer drop that our conditioning does not repair — which would mean the failure lies somewhere other than HLA or ancestry background, most plausibly in exposure history, pointing the field toward pathogen-ecology rather than genetics; and second, a near-zero transfer drop for the unmodified baselines, which would mean the ancestry-bias concern, while real for germline instruments, simply does not propagate to somatic receptor repertoires. Both outcomes would be worth publishing; only the second would make this line of work unnecessary.

References

  1. Precision diagnosis for monogenic diabetes requires ethnicity specific criteria for genetic testing
  2. Interpretable GWAS by linking clinical phenotypes to quantifiable immune repertoire components
  3. A pan-disease and population-level single-cell TCRαβ repertoire reference
  4. Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood
  5. Causal associations between risk factors and common diseases inferred from GWAS summary data
  6. Modern Hopfield Networks and Attention for Immune Repertoire Classification
  7. Sparse haplotype-based fine-scale local ancestry inference at scale reveals recent selection on immune responses
  8. Comprehensive Analysis of CDR3 Sequences in Gluten-Specific T-Cell Receptors Reveals a Dominant R-Motif and Several New Minor Motifs
  9. Tracking global changes induced in the CD4 T-cell receptor repertoire by immunization with a complex antigen using short stretches of CDR3 protein sequence
  10. Specificity, Privacy, and Degeneracy in the CD4 T Cell Receptor Repertoire Following Immunization
  11. Relationship between T-cell receptor beta chain sequences and human cytomegalovirus infection in allogeneic hematopoietic stem cell transplant recipients
  12. Across ancestries, HLA-B∗08:01∼DRB1∗03:01 (DR3) and HLA-DQA∗01:02 (DR2) increase the risk to develop juvenile-onset systemic lupus erythematosus through low complement C4 levels
  13. A Noisy-Label-Learning Formulation for Immune Repertoire Classification and Disease-Associated Immune Receptor Sequence Identification
  14. A Label Disambiguation-Based Multimodal Massive Multiple Instance Learning Approach for Immune Repertoire Classification