A person’s immune system carries a molecular record of both genetic inheritance and past encounters with the outside world. That record is partly encoded in the vast collection of receptors displayed by B cells and T cells, known collectively as the adaptive immune receptor repertoire. Now, an analysis of these receptors in people with celiac disease has revealed a surprising twist: a machine-learning system that could distinguish patients from controls was not primarily detecting the disease itself. Instead, it was reading a powerful genetic risk factor that shapes the immune system before it has encountered gluten.

The study, published in Immunogenetics, examined naïve immune cells—B and T cells that have not yet been activated by their target antigens. This distinction is crucial. Receptors found on activated or memory cells can reflect previous infections, vaccinations, dietary exposures or autoimmune inflammation. Naïve repertoires, by contrast, offer a window into how inherited biology and early immune development have constructed the starting pool of potential immune responses. The researchers asked whether these unexperienced receptor populations could distinguish people with celiac disease from unaffected controls, and which molecular features were responsible for any successful classification.

Celiac disease is an autoimmune disorder in which exposure to gluten triggers immune-mediated damage in the small intestine. The condition is strongly associated with particular variants of human leukocyte antigen, or HLA, a family of proteins that displays peptide fragments to T cells. The HLA-DQ2.5 allotype is especially important: it is present in a large majority of people with celiac disease, although carrying it does not guarantee that disease will develop. HLA molecules act like molecular presentation platforms, holding fragments of proteins in a configuration that allows T-cell receptors to inspect them. In celiac disease, HLA-DQ2.5 can present modified gluten peptides that stimulate pathogenic T cells.

The researchers used three independent naïve adaptive immune receptor sequencing datasets. The main T-cell dataset contained samples from 103 people with celiac disease and 103 controls. One B-cell dataset included 102 people with celiac disease and 102 controls, while a second, previously published B-cell dataset contained 48 patients and 44 controls. The samples came from Norwegian cohorts, including the population-based Trøndelag Health Study. Naïve B cells and T cells were separated from blood-derived peripheral blood mononuclear cells using fluorescence-activated cell sorting, or FACS, before their receptor genes were sequenced.

The resulting sequences were analyzed with ImmuneML, an open-source platform designed for machine learning on adaptive immune receptor data. The models were trained to classify repertoires using several kinds of information: the frequency of variable, joining and diversity genes; short sequence patterns called k-mers; disease-associated reference sequences; and the extent to which receptor clones were shared between individuals. T-cell receptors are assembled from gene segments during lymphocyte development, and their variable regions help determine which molecular structures they can recognize. The complementarity-determining region 3, or CDR3, is particularly diverse and often contributes directly to antigen binding. K-mer models reduce longer sequences into short overlapping patterns, allowing algorithms to detect recurring sequence motifs.

The naïve CD4-positive T-cell repertoires produced a moderately strong disease-classification signal. Logistic-regression models based on variable-gene frequencies achieved a median area under the receiver operating characteristic curve, or AUROC, of 0.72. An AUROC of 0.5 represents random classification, whereas 1.0 represents perfect separation. Models using variable-gene frequencies achieved a median balanced accuracy of 0.707 across three assessment splits, while a positional k-mer approach combining CDR3 information with variable-gene identity reached a median balanced accuracy of 0.699. These results suggested that the T-cell receptor repertoire contained reproducible differences between the two groups, but they did not establish that the differences were uniquely caused by celiac disease.

To identify the source of the signal, the researchers tested whether the same naïve T-cell data could predict HLA-DQ2.5 status. The answer was strikingly affirmative. Models based on T-cell receptor variable and joining gene frequencies classified people carrying at least one copy of HLA-DQ2.5 from those with other HLA-DQ allotypes with a median AUROC of 0.972 and a median balanced accuracy of 0.927. In other words, the naïve T-cell repertoire carried a highly recognizable imprint of this inherited HLA state. The variable gene TRAV12-2 repeatedly emerged as one of the strongest predictive features, consistent with earlier evidence linking its usage to HLA-DQ2.5.

The biological explanation lies in the thymus, the organ where developing T cells are tested and selected. T-cell receptors are generated through genetic rearrangement, creating enormous diversity, but not every newly formed receptor is allowed to enter the circulation. T cells that react too strongly to the body’s own molecules are removed or restrained, while those capable of interacting appropriately with self-HLA molecules receive survival signals. Because HLA variants alter the molecular surfaces encountered during this selection process, they can influence which T-cell receptor combinations survive. This is known as a trans-acting effect: a gene outside the receptor locus changes the receptor repertoire indirectly through cellular selection. HLA-DQ2.5 may therefore leave a detectable signature in naïve T cells long before gluten-driven inflammation begins.

The decisive experiment was to control for the HLA-DQ2.5 imbalance between the celiac and control groups. When the researchers restricted the analysis to HLA-DQ2.5-positive individuals, the apparent diagnostic power of the naïve T-cell repertoires disappeared. Balanced accuracy fell to approximately 0.5, the level expected from random guessing, across the tested models. New models trained only within the HLA-DQ2.5-positive group also failed to make reliable disease classifications. This result indicates that the original machine-learning success was largely exploiting the unequal distribution of HLA-DQ2.5 between patients and controls, rather than identifying a separate molecular signature of celiac disease.

The finding is a warning about a subtle but important problem in biomedical machine learning. An algorithm can classify two groups accurately while relying on a feature that is correlated with disease but is not itself a direct marker of disease activity. In this case, HLA-DQ2.5 is a major risk factor for celiac disease, so its presence naturally makes the patient and control groups look different. But a model that recognizes HLA-associated repertoire architecture may be measuring inherited susceptibility rather than intestinal injury, gluten exposure or active autoimmunity. Without careful stratification, such a model could appear to be a disease diagnostic while actually functioning as an indirect HLA classifier.

The study also found a sharp contrast between T-cell and B-cell repertoires. None of the naïve B-cell receptor models accurately distinguished patients with celiac disease from controls in the principal dataset. Their AUROC and balanced-accuracy scores were consistent with random classification. The researchers also attempted to reproduce an earlier report that had identified celiac disease using naïve B-cell receptor repertoires. They analyzed both processed data from that study and raw sequencing reads reprocessed through a comparable pipeline, but neither approach produced better-than-random classification. Differences in cohort composition, sequencing and preprocessing, or undocumented machine-learning parameters could explain the discrepancy.

The failure of naïve B-cell receptors to classify disease does not mean that B cells are irrelevant to celiac disease. Activated B cells and plasma cells produce antibodies against deamidated gluten and against transglutaminase 2, an autoantigen associated with the disorder. These disease-related cells can carry highly stereotyped receptor features, and some celiac disease-specific antibodies show limited somatic hypermutation, suggesting that certain responses may draw on germline-encoded receptor configurations. However, such signals may be much easier to detect in antigen-experienced or tissue-resident cells than in the broad, unactivated naïve B-cell pool circulating in blood.

The T-cell results likewise do not overturn previous work showing that effector or memory T-cell repertoires can distinguish celiac disease. Once gluten-specific cells have encountered antigen and expanded, their receptors can become enriched for recognizable sequence and gene-usage patterns. The present study addresses a different question: whether disease-linked differences are already visible before antigen-driven activation. Its answer is that naïve CD4-positive T cells do reflect inherited HLA biology, but the researchers found no robust evidence for a strong, HLA-independent celiac disease signature in these cells.

The study’s cohorts were substantial for immune-repertoire research, but the authors emphasize several limitations. Restricting analysis to HLA-DQ2.5-positive participants reduces the number of available samples and therefore statistical power. Naïve receptor repertoires also vary considerably from one person to another, partly because receptor rearrangement is stochastic. Subtle effects from genetic variation within the T-cell receptor or immunoglobulin loci could therefore be obscured. In addition, the analyzed samples came from a relatively ethnically homogeneous, low-migration region of Norway, so the results may not apply uniformly to populations with different genetic backgrounds and HLA frequencies.

Even so, the work offers a clear lesson for efforts to develop blood-based immune diagnostics. Before treating a predictive sequence pattern as disease-specific, researchers must determine whether it reflects age, ancestry, treatment, sampling site, cell composition or a germline genetic factor such as HLA. Cross-validation and independent test sets can measure how well a model generalizes, but they cannot by themselves reveal whether the model has learned the intended biology. Carefully balanced cohorts and analyses that explicitly control for genetic risk alleles are essential.

For celiac disease, the immediate implication is not a new clinical test, but a more precise understanding of how susceptibility is built into the immune system. HLA-DQ2.5 appears capable of shaping the naïve T-cell repertoire through thymic selection, potentially creating a pre-existing landscape in which gluten-reactive responses are more likely to emerge. Future studies using larger, HLA-balanced cohorts could investigate other celiac-associated HLA allotypes, including DQ8 and DQ2.2, and test whether weaker signals from non-HLA genetic variants can be separated from the dominant HLA effect. The immune system may be carrying clues to autoimmune risk from the moment its T cells are selected—but decoding those clues will require algorithms that know the difference between a cause, a correlate and a genuinely diagnostic signal.