Identification of a type 1 diabetes–associated T cell receptor repertoire signature from the human peripheral blood

A new study has identified, for the first time, type 1 diabetes–specific motifs within the T‑cell receptor (TCR) repertoire, using artificial intelligence to detect rare autoimmune signals in unprecedentedly large immune datasets. The findings, now accepted in Science Advances, offer a potential new metric for autoreactivity and lay groundwork for future TCR‑based diagnostics and therapeutics.

"In this study, we analyzed more than 275 million T-cell receptor sequences from over 2,200 individuals and used artificial intelligence to identify immune signatures specifically associated with type 1 diabetes. Traditionally, autoimmune risk in type 1 diabetes has been inferred mainly from inherited genetic risk factors. Our work demonstrates for the first time that disease-specific sequence patterns can be detected directly within the immune system, providing a real-time molecular signature of autoimmune activity," first author Puneet Rawat says. He is a researcher at the Institute of Clinical Medicine at the University of Oslo.

"These findings represent a shift in how we can study autoimmune disease. Genetics tells us who may be at risk, but it does not capture how the immune system is behaving in real time. By using AI to read immune repertoires directly, we open new possibilities for earlier diagnosis, improved monitoring of disease progression, and eventually more personalized interventions. This approach has the potential to transform not only type 1 diabetes research, but autoimmune disease research more broadly," Professor Victor Greiff at the Institute of Clinical Medicine at the University of Oslo says.

Decoding the immune system’s most elusive signals: AI for detecting autoimmune TCR signatures
The human immune system is an incredibly complex network designed to protect us from germs, infections, and anything that doesn’t naturally belong in the body. One of its most remarkable features is its ability to produce a huge variety of immune cells, each equipped with specialized tools to recognize and neutralize threats. Among the most important of these tools are T-cell receptors (TCRs). TCRs are present on the surface of T cells, like microscopic keys that help the immune system identify invaders and coordinate the right defensive response. Every individual carries more than 1 billion (10⁹) distinct TCRs, each one slightly different in structure and function. The TCR landscape is also highly personal and dynamic, with any two people sharing only a very small fraction of their repertoires (1). However, this remarkable flexibility can occasionally go off track due to genetics, random biological processes, or environmental influences. In rare cases, a few TCRs mistakenly recognize the body’s own cells as foreign and initiate an attack against healthy tissue, giving rise to autoimmune disease. A recent study of 22 million individuals in the UK showed that autoimmune diseases affect more than 10% of the population (2). In the United States alone, the direct medical costs associated with autoimmune conditions have been estimated to exceed $100 billion per year, although this figure is believed to substantially underestimate the true burden when lost productivity and indirect costs are considered (3). In Norway, around 5-7% of the population has an autoimmune disease with diabetes alone costing around 11.79 Billion NOK in 2022 (4). These diseases remain incredibly difficult to treat since the underlying cause, often located in the TCR repertoires, remains unidentified.

The challenge: finding rare harmful autoimmune signals
Identifying exactly which TCRs are responsible for this harmful behavior is immensely difficult in autoimmune diseases due to low signal frequency, typically only one in ten thousand to one in one hundred thousand T cells are disease-associated (5, 6). The challenge is more like searching for a few misprinted words (autoimmune TCR signals) hidden within an entire library (the individual's TCR repertoire). To make matters even more difficult, each person's “library” is written in a slightly different language, shaped by their genetics, environment, and unique immune experiences. Traditional scientific methods are often overwhelmed by the sheer number of unique TCRs, making it extremely difficult to separate the harmful signals from the millions of harmless ones. This complexity has motivated researchers to turn toward artificial intelligence to uncover patterns that would otherwise remain invisible.

Studying type 1 diabetes at unprecedented scale
Our study centered on Type 1 diabetes (T1D), a condition in which the immune system destroys the insulin-producing cells of the pancreas. T1D is strongly influenced by genetics (particularly by a group of genes called HLA) and the presence of autoantibodies (signaling early autoimmune activity). HLA genes help educate the immune system on what is “self” and what is “non-self.” Certain HLA types, especially DR3 and DR4, are known to dramatically increase the likelihood of developing T1D. To investigate the immune basis of T1D, we sequenced an unprecedentedly large dataset of more than 275 million unique TCR sequences from 2,250 individuals (7). For each person, we also had rich metadata information about their HLA genes and whether they had autoantibodies, both of which are crucial for investigating the causes of T1D.

Using AI to read immune repertoires directly
For decades, it has been assumed that understanding autoimmune signatures requires genetic information. However, genetics represents a static form of risk assessment. It tells us who may be predisposed to T1D, but not how the immune system behaves as the disease develops or how it might respond to treatment (8). To establish a strong genetic foundation, we first used statistical methods to map high-risk HLA alleles to specific TCR patterns (9). This let us identify TCRs that were more commonly found in individuals with these known HLA risk genes. These TCRs formed a genetic “baseline” for T1D susceptibility. This step reaffirmed existing knowledge: certain TCR features are strongly influenced by inherited genetic factors.

Autoimmune diseases do not follow a predictable script. Some people with high-risk genes develop T1D, while others never do. To uncover signals that shift as the disease progresses, we needed a way to examine the immune repertoire without relying on genetic markers. To do this, we trained a deep-learning model called DeepRC (10), based on modern Hopfield networks (11). DeepRC is specifically designed to search through millions of sequences and detect patterns too subtle or complex for conventional analysis. DeepRC treats each person’s immune system as a large, unlabeled collection of TCRs. It does not use any genetic information. Instead, it looks only at the sequences themselves and learns how they differ between people with T1D and those without the disease. During training, DeepRC discovers small yet consistent patterns that occur more frequently in T1D repertoires. We then refined the T1D-positive signals by identifying the overlap between the AI-discovered sequences and the genetic baseline, which removed much of the inherited HLA-driven background and improved classification performance. These refined signals also showed consistent occurrence in T1D repertoires in an independent test dataset and in samples from pancreatic lymph nodes and spleen (the tissues most directly involved in T1D) strengthening the evidence that we had uncovered meaningful biological signals. Taken together, these findings demonstrate that we have, for the first time, identified T1D-specific motifs within the TCR repertoire.

Why this matters
This application of AI marks an important shift in how we understand autoimmune disease. Unlike HLA genes, which reflect lifelong predisposition, AI-discovered TCR signatures enable tracking of early immune events, including the rise of autoantibodies. This new perspective opens the door to more sensitive early diagnostics, improved monitoring of disease progression, and new strategies to predict autoimmune activity. These findings have been accepted in Science Advances (7), and the large-scale T1D dataset is being used in a Kaggle competition (https://www.kaggle.com/competitions/adaptive-immune-profiling-challenge-2025; in principle accepted in Nature Methods) to enable the development of next-generation AI models for immune repertoire analysis. This effort will catalyze the development of community-wide AI benchmarks and accelerate methodological innovation for immune signal identification.


1. N.-P. Weng, Numbers and odds: TCR repertoire size and its age changes impacting on T cell functions. Semin. Immunol. 69, 101810 (2023).
2. N. Conrad, S. Misra, J. Y. Verbakel, G. Verbeke, G. Molenberghs, P. N. Taylor, J. Mason, N. Sattar, J. J. V. McMurray, I. B. McInnes, K. Khunti, G. Cambridge, Incidence, prevalence, and co-occurrence of autoimmune disorders over time and by age, sex, and socioeconomic status: a population-based cohort study of 22 million individuals in the UK. Lancet 401, 1878–1890 (2023).
3. F. W. Miller, The increasing prevalence of autoimmunity and autoimmune diseases: an urgent call to action for improved understanding, diagnosis, treatment, and prevention. Curr. Opin. Immunol. 80, 102266 (2023).
4. J. M. Kinge, H. Øien, J. L. Dieleman, B.-A. Reme, A. K. S. Knudsen, G. Godager, G. Selbæk, J. C. Frich, E. Barış, C. J. L. Murray, S. E. Vollset, Forecasting total and cause-specific health expenditures for 116 health conditions in Norway, 2022-2050. BMC Med. 23, 116 (2025).
5. C. Saggau, P. Bacher, D. Esser, M. Rasa, S. Meise, N. Mohr, N. Kohlstedt, A. Hutloff, S.-S. Schacht, J. Dargvainiene, G. R. Martini, K. H. Stürner, I. Schröder, R. Markewitz, J. Hartl, M. Hastermann, A. Duchow, P. Schindler, M. Becker, C. Bautista, J. Gottfreund, J. Walter, J. K. Polansky, M. Yang, R. Naghavian, M. Wendorff, E.-M. Schuster, A. Dahl, A. Petzold, S. Reinhardt, A. Franke, M. Wieczorek, L. Henschel, D. Berger, G. Heine, M. Holtsche, V. Häußler, C. Peters, E. Schmidt, S. Fillatreau, D. H. Busch, K.-P. Wandinger, K. Schober, R. Martin, F. Paul, F. Leypoldt, A. Scheffold, Autoantigen-specific CD4+ T cells acquire an exhausted phenotype and persist in human antigen-specific autoimmune diseases. Immunity 57, 2416–2432.e8 (2024).
6. P. Bacher, A. Scheffold, Flow-cytometric analysis of rare antigen-specific T cells: Flow-Cytometric Analysis of Rare Antigen-Specific T Cells. Cytometry A 83, 692–701 (2013).
7. P. Rawat, M. R. Shapiro, L. D. Peters, M. Widrich, K. Mayer-Blackwell, K. Motwani, M. Pavlović, G. al Hajj, A. L. Posgai, C. Kanduri, G. Isacchini, M. Chernigovskaya, L. Scheffer, K. Motwani, L. O. Balzano-Nogueira, C. M. Pettenger-Willey, S. Valkiers, L. Jacobsen, M. J. Haller, D. A. Schatz, C. H. Wasserfall, R. O. Emerson, A. J. Fiore-Gartland, M. A. Atkinson, G. Klambauer, G. K. Sandve, V. Greiff, T. M. Brusko, Identification of a type 1 diabetes-associated T cell receptor repertoire signature from the human peripheral blood, medRxiv (2024)p. 2024.12.10.24318751.
8. M. Nakayama, A. W. Michels, Using the T cell receptor as a biomarker in type 1 diabetes. Front. Immunol. 12, 777788 (2021).
9. K. Ishigaki, K. A. Lagattuta, Y. Luo, E. A. James, J. H. Buckner, S. Raychaudhuri, HLA autoimmune risk alleles restrict the hypervariable region of T cell receptors. Nat. Genet. 54, 393–402 (2022).
10. M. Widrich, B. Schäfl, H. Ramsauer, M. Pavlović, L. Gruber, M. Holzleitner, J. Brandstetter, G. K. Sandve, V. Greiff, S. Hochreiter, G. Klambauer, Modern Hopfield Networks and Attention for Immune Repertoire Classification, arXiv [cs.LG] (2020). http://arxiv.org/abs/2007.13505.
11. H. Ramsauer, B. Schäfl, J. Lehner, P. Seidl, M. Widrich, T. Adler, L. Gruber, M. Holzleitner, M. Pavlović, G. K. Sandve, V. Greiff, D. Kreil, M. Kopp, G. Klambauer, J. Brandstetter, S. Hochreiter, Hopfield Networks is All You Need, arXiv [cs.NE] (2020). http://arxiv.org/abs/2008.02217.