Biological basis of rapidness-sensitive discrimination between harmful and harmless antigens: implementation and experimental validation of the predictive immune memory model

The adaptive immune system must decide, for each antigen it encounters, whether to attack or tolerate. Yoshido and Honda (preprint 2021; iScience 2023) proposed the first learning-based mechanistic model of this decision, the "predictive immune memory" model, in which conventional and regulatory T cells act as predictors, memory T-cell formation is driven by prediction errors (possibly computed via cytokines), and discrimination between harmful and harmless antigens depends on both antigen concentration and the rapidness of antigen increase. The model is purely computational. Whether such prediction-error computation exists in vivo, which biological machinery implements it, whether the parameters required for discrimination have molecular counterparts, and whether rapidness-based discrimination survives extension to proliferating pathogens and multi-antigen exposure all remain open. The problem is to establish the biological basis of rapidness-sensitive, learning-based antigen-risk discrimination — by identifying and experimentally validating the implementing mechanism (predictive coding or a mechanistic equivalent), or by refuting rapidness-based danger sensing under controlled antigen concentration and dose.

Background

The adaptive immune system constantly encounters molecular substances (antigens) from pathogens, food, commensal microbes, and the body's own tissues, and must decide for each whether to mount a potentially damaging inflammatory attack or remain tolerant. Two antagonistic white-cell populations are central to this decision: conventional CD4+ T cells ($T_{conv}$, helper cells that orchestrate inflammation) and regulatory T cells ($T_{reg}$, cells that suppress immune responses). After an encounter, long-lived memory T cells of both types persist for years, so what is learned during one exposure shapes the response to the next. How this learning-like updating works mechanistically is unresolved.

Classical explanations assign the decision to antigen type: Matzinger's danger model holds that responses are triggered by "danger signals" such as tissue-damage or pathogen-associated molecular patterns rather than by foreignness alone. A complementary line emphasizes the temporal profile of exposure: the discontinuity theory of Pradeu and colleagues proposes that sudden or intermittent antigen exposure provokes vigorous responses, whereas gradual and persistent exposure induces weak responses or tolerance. On the modeling side, Sontag showed that an immune-regulation circuit built from an incoherent feedforward loop (a network motif in which a fast activating branch races a slower inhibitory branch) responds only transiently to constant or linearly increasing antigen but in sustained fashion to exponentially growing antigen — so a circuit can in principle read out how fast a threat grows. Experimentally, antigen kinetics is known to matter at several scales: slow-release antigen formulations change immune reactivity, sustained antigen availability in germinal centers (the structures in lymphoid tissue where B cells mature antibody responses) enhances antibody responses to vaccination, and T cells selectively filter oscillatory antigen-like signals on the minutes timescale. Fold-change detection — responding to relative rather than absolute changes in a signal — is an established circuit design principle in cell signaling. However, none of these results provides a mechanism by which the immune system learns from its own experience to combine antigen concentration with the speed of antigen increase into a danger assessment that is stored as immunological memory.

Yoshido and Honda supplied the first such learning-based model. They imported predictive coding — a framework from machine learning and neuroscience in which a system maintains predictions about its inputs and updates them using prediction errors, defined as the deviations between predictions and observations — into T-cell population dynamics. In their ordinary-differential-equation model, $T_{conv}$ cells act as a predictor of the antigen amount $a$ (the prediction being $m_c T_{conv}$), while $T_{reg}$ cells act as a predictor of excessive immune response; the production rates of memory $T_{conv}$ and memory $T_{reg}$ cells are driven by the corresponding prediction errors, for example $E_c = e_c\,|a - m_c T_{conv}|_+$ where $|x|_+$ denotes the positive part. Numerical simulation showed that the model discriminates antigens by both concentration and input rapidness: a rapid increase to a high antigen concentration provokes a strong response, whereas a slow ramp to the same final concentration induces a suppressive, tolerance-like state, because the prolonged experience of low concentrations generates positive prediction error preferentially for memory $T_{reg}$ production. The model also reproduced history-dependent changes in discrimination, such as allergy onset followed by remission under allergen immunotherapy. Concentration-dependent discrimination, however, holds only in a parameter regime: it requires that the $T_{reg}$-side evaluation of $T_{conv}$ activation, a function $g(T_{conv})$ with scaling parameter $K$, overestimate activation relative to the $T_{conv}$-side evaluation $m_c T_{conv}$ at low antigen concentrations — a condition on $K$ versus $m_c$ for which the authors give a functional but no molecular interpretation, and the rapidness threshold (input time constant $\tau \approx 500$ in model units) has no calibrated biological timescale. The authors explicitly state that the biological implementation is hypothetical: they propose that cytokines (secreted signaling proteins) carry quantitative information about cell population sizes and antigen amounts, and that prediction errors are computed through intracellular signal transduction in T cells, but validation would require quantifying the time series of T-cell populations carrying each T cell receptor (TCR, the antigen-recognition receptor of T cells) together with cytokines. The model also treats only a single, externally supplied antigen that neither proliferates nor is eliminated, whereas real pathogens self-replicate and real exposure histories involve many antigens simultaneously.

Problem Statement

Does the adaptive immune system implement a mechanism that uses the rate of antigen increase — not only antigen concentration — as a danger cue, and if so, what is the biological machinery computing it?

The predictive immune memory model of Yoshido and Honda is currently the only mechanistic model that produces adaptive, memory-based, concentration- and rapidness-dependent discrimination between harmful and harmless antigens, but it is purely computational. The source's own open questions, together with the discrimination conditions it identifies but leaves biologically ungrounded, define the target, and the four components below all bear on one objective — establishing the biological basis of learning-based danger-cue sensing: (1) determine experimentally whether memory T-cell formation in vivo is driven by prediction-error-like quantities (deviations of observed antigen and response levels from predictions set by the current $T_{conv}$ and $T_{reg}$ populations), so that the speed of antigen increase serves as a reliable danger signal rather than an artifact of dose accumulation; (2) identify which molecules and cells implement the prediction-error computation — the source's hypothesis is cytokine-coded population and antigen information integrated by intracellular signal transduction in T cells, analogous to quorum sensing (density-dependent collective regulation via secreted signals); (3) determine whether the parameter conditions required for discrimination in the model, chiefly the overestimation of $T_{conv}$ activation by the $T_{reg}$-side evaluation ($K$) relative to the $T_{conv}$-side evaluation ($m_c$) at low antigen concentration and the rapidness threshold $\tau$, correspond to identifiable and measurable biological quantities; and (4) determine whether the mechanism operates in the biologically essential regimes the minimal model excludes — antigens that proliferate and are eliminated by the immune response (pathogens), and simultaneous exposure to multiple antigens.

A complete answer takes one of two forms. (a) Mechanism identification: an answer identifies and experimentally validates the biological implementation — the predictive-coding mechanism, or an alternative mechanism that likewise accounts for concentration- and rapidness-dependent, history-dependent discrimination — through antigen-exposure experiments whose input profiles reach the same final concentration at different rates, with time-resolved, TCR-specific measurement of $T_{conv}$, $T_{reg}$, and memory T-cell populations and of candidate cytokine mediators, plus perturbation evidence localizing the computation. Theoretical contributions such as grounding the parameters $K$, $m_c$, and $\tau$ in molecular quantities, or extending the model to proliferating and multiple antigens, may form part of such an answer but are not substitutes for it. (b) Refutation: an answer demonstrates experimentally that, once antigen concentration and cumulative dose are controlled across the physiologically relevant range of input rates, immune discrimination and memory formation do not depend on how fast antigen concentration rises — establishing that rapidness is not a danger cue. The four components enumerated above do not all carry the same weight: (1) and (2) state the substance of form (a), whereas components (3) and (4) are optional adjuncts whose omission does not make an answer incomplete.

References

1. Kana Yoshido, Naoki Honda, 'Adaptive discrimination of antigen risk by predictive coding in immune system', bioRxiv preprint (2021). doi:10.1101/2021.12.12.472285, https://doi.org/10.1101/2021.12.12.472285
2. Kana Yoshido, Naoki Honda, 'Adaptive discrimination between harmful and harmless antigens in the immune system by predictive coding', iScience 26(1):105754 (2023). doi:10.1016/j.isci.2022.105754, https://doi.org/10.1016/j.isci.2022.105754
3. Yuna Kotsubo, Akane Hara, Rena Hayashi, Yoh Iwasa, 'Age-dependence of food allergy due to decreased supply of naïve T cells', Journal of Theoretical Biology 602-603:112060 (2025). doi:10.1016/j.jtbi.2025.112060, https://doi.org/10.1016/j.jtbi.2025.112060
4. I. Eizenberg-Magar, L. Dayan, B. Chain, Y. E. Antebi, 'T cell decision-making decodes the dynamic antigenic landscape', Frontiers in Immunology 17:1806850 (2026). doi:10.3389/fimmu.2026.1806850, https://doi.org/10.3389/fimmu.2026.1806850
5. Thomas Pradeu, Sébastien Jaeger, Eric Vivier, 'The speed of change: towards a discontinuity theory of immunity?', Nature Reviews Immunology 13:764-769 (2013). doi:10.1038/nri3521, https://doi.org/10.1038/nri3521
6. Thomas Pradeu, Eric Vivier, 'The discontinuity theory of immunity', Science Immunology 1(1):aag0479 (2016). doi:10.1126/sciimmunol.aag0479, https://doi.org/10.1126/sciimmunol.aag0479
7. Eduardo D. Sontag, 'A Dynamic Model of Immune Responses to Antigen Presentation Predicts Different Regions of Tumor or Pathogen Elimination', Cell Systems 4(2):231-241 (2017). doi:10.1016/j.cels.2016.12.003, https://doi.org/10.1016/j.cels.2016.12.003
8. Pål Johansen, Tazio Storni, Lorna Rettig, Zhiyong Qiu, Ani Der-Sarkissian, Kent A. Smith, Vania Manolova, Karl S. Lang, Gabriela Senti, Beat Müllhaupt, Tilman Gerlach, Roberto F. Speck, Adrian Bot, Thomas M. Kündig, 'Antigen kinetics determines immune reactivity', Proceedings of the National Academy of Sciences 105(13):5189-5194 (2008). doi:10.1073/pnas.0706296105, https://doi.org/10.1073/pnas.0706296105
9. Hok Hei Tam, Mariane B. Melo, Myungsun Kang, Jeisa M. Pelet, Vera M. Ruda, Maria H. Foley, Joyce K. Hu, Sudha Kumari, Jordan Crampton, Alexis D. Baldeon, Rogier W. Sanders, John P. Moore, Shane Crotty, Robert Langer, Daniel G. Anderson, Arup K. Chakraborty, Darrell J. Irvine, 'Sustained antigen availability during germinal center initiation enhances antibody responses to vaccination', Proceedings of the National Academy of Sciences 113(43):E6639-E6648 (2016). doi:10.1073/pnas.1606050113, https://doi.org/10.1073/pnas.1606050113
10. Geoff P. O'Donoghue, Lukasz J. Bugaj, Warren Anderson, Kyle G. Daniels, David J. Rawlings, Wendell A. Lim, 'T cells selectively filter oscillatory signals on the minutes timescale', Proceedings of the National Academy of Sciences 118(9):e2019285118 (2021). doi:10.1073/pnas.2019285118, https://doi.org/10.1073/pnas.2019285118
11. Miri Adler, Pablo Szekely, Avi Mayo, Uri Alon, 'Optimal Regulatory Circuit Topologies for Fold-Change Detection', Cell Systems 4(2):171-181 (2017). doi:10.1016/j.cels.2016.12.009, https://doi.org/10.1016/j.cels.2016.12.009
12. P. Matzinger, 'Tolerance, Danger, and the Extended Family', Annual Review of Immunology 12:991-1045 (1994). doi:10.1146/annurev.iy.12.040194.005015, https://doi.org/10.1146/annurev.iy.12.040194.005015
13. Yuanshan Zhang, 'Chronic Antigen Exposure and Immune Tolerance Dynamics: A Unified Treg–Tpex Phase Transition Framework', Research Square preprint (2026, non-peer-reviewed). doi:10.21203/rs.3.rs-10300749/v1, https://doi.org/10.21203/rs.3.rs-10300749/v1