When you work on mathematical modelling, you tend to resist the idea of assigning special attributes to certain parts of your system just for the sake of explaining the data. Special cases should emerge from the basic principles of the system, not because there is something inherently special about that case per se. By that logic, there are no special cases, there are only less and more frequent cases, both existing in the state space of your core mechanisms. If this is not the case, either a dimension is missing (which may feel like an inelegant way to resolve the issue, but sometimes it really is necessary), or there may exist a completely different set of mechanisms that would explain that (n+1)th observation. Probably the best example is the transition from Newtonian to quantum mechanics: Quantum mechanics doesn't just extend Newtonian mechanics - it replaces its foundational assumptions.Nothing in biology makes sense except in the light of evolution.
- Theodosius Dobzhansky
I am not an immunologist by any standards, but I think the (n+1)th observation that led to the rise of the the self-nonself school of thought1 was self-tolerance. Once you understand how powerful the immune system is, it’s kind of insane that we don’t accidentally destroy ourselves all the time. So the fact that we’re still alive had to be explained - and the answer came in the form of a special subset of CD4⁺ T cells: the T regulatory cells, or Tregs (which got the Nobel Prize 2025 in Medicine by the way!).
So the first step to not destroying yourself is to recognize yourself (Γνῶθι σεαυτόν!). Tregs were given the role of recognizing self antigens - unlike other T helper cells, whose main job is to recognize foreign antigens - and instead of launching an inflammatory response, they were supposed to suppress it, so that we don't die. So along both dimensions - the identity of the antigen and the direction of the response - they do the opposite of what other CD4⁺ T cell subtypes are known for (The real question here is how “self” and “foreign” are even defined - which turns out to be much trickier than it sounds: a nice book about this confusion is here, and it still is a confusion.)
The way Tregs are differentiated introduces an if/else logic into the mechanisms of T cell differentiation: if a T cell binds to a self antigen too strongly, it dies (negative selection – this makes sense even if you don't have Tregs in this framework, because you simply don’t want helper cells that strongly bind your own tissues); not too strongly, it becomes a Treg (strong enough to do its job, but not enough to kill - this is where it gets weird); if very weakly, it’s allowed to circulate as a conventional T cell, with the hope that it will bind to a pathogen and become a helper cell (also sensible). At least these are the rules for the ones differentiated in the thymus, the thymically derived Tregs (tTregs). Then you also have Tregs that are induced in the periphery (pTregs, or iTregs for induced), which see their antigen outside the thymus and follow yet another set of rules – depending on environmental cues and antigen persistence. Even if we allow this complicated logic, it does not fully explain why we do observe Tregs against commensal microbes – which are by this definition, not self. These rules leave you with a complicated decision tree that starts to feel arbitrary. Makes you wonder whether there is another explanation that makes both Tregs and T helper cells emerge from the same set of principles.
The Danger Model proposed by Matzinger [1] offers a slightly more satisfying explanation - it reframes the special cases of self-tolerance (and its loss, autoimmunity) by placing Tregs within a broader framework and redefining how the immune system decides to mount a response or not2. In this model, the immune system responds not necessarily against what is foreign (non-self), but to what is dangerous - not perceived as dangerous, but literally, actively causing harm: killing your epithelial cells, releasing toxins, etc. Within this framework, Tregs are just another class of effector T cells, but just the less aggressive type in terms of inflammatory response. The decision about which class should dominate is made – and communicated – by the tissue [2], likely based on how much inflammatory response and thus how much collateral damage it can tolerate while clearing the pathogen. This helps explain why you have a dominantly Treg + IgA response in the eye (which cannot regenerate that fast, and would be pretty bad to lose in terms of survival – unless you must lose it to survive since being blind is better than being dead), as opposed to a Th1 response (pretty aggressive) in the liver, which can regenerate way better. In Matzinger's view, we should call all T cells “regulatory”, since they constantly regulate each others’ behavior.
That said, the class of the response is probably not determined by the tissue's regenerative capacity alone, but by its regenerative capacity relative to the frequency of potential infections. Take the gut, for example: the intestinal lining is one of the fastest-renewing tissues in the body – it certainly regenerates better than the eye or the brain. Gut infections are also no joke, and might be even more fatal than losing an eye: If you don't kill certain diarrhea-causing enteric pathogens fast enough, it can kill you within a week. Granted, we do mount a Th1 + IgG response against such pathogens - because at that point, there is no use worrying about collateral damage anymore – or even about not killing yourself – if the pathogen is going to kill you anyway. The same logic applies to the eye, as Matzinger points out in her lectures2: when shit hits the fan, the immune system defaults to a Th1 response, regardless of the tissue. But if we responded like that every single time a less invasive pathogen entered one of our mucosal sites (now think about all the food you eat and the air you breathe), we would end up destroying these tissues over and over again. Even if they could handle a single hit, the cumulative damage would be catastrophic. So it makes sense that sites like the skin, lung, and gut rely on a Treg + IgA-dominant response as a first line of defense.
This paradigm shift neatly explains how we coexist with symbiotic commensals which are not self yet don’t cause damage, and why the immune system fails to target tumors until they become dangerous3. In this framework, tolerance - and by extension, the notion of “self” - is not a static state learned during infancy. It is a dynamic process that requires constant re-evaluation as long as your thymus produces new T cell clones and releases them to the periphery. The principles of this framework are as follows – this is danger model synthesized with the 3 signal model of adaptive immunity [3]:
1. You strictly need professional antigen presenting cells (APCs) to activate naive T cells. No other cell can do this job. That's why Matzinger called them “professional” – since there are also unprofessional antigen presenting cells (like B cells) that cannot activate a naive T cell, but can activate a memory T cell4. The most common and best-studied professional APC is the dendritic cell (DC). Macrophages are also professional APCs.
2. The first step to activation – Signal 1 –is antigen recognition. A T cell must have a T cell receptor (TCR) that binds with sufficient affinity to the antigen–MHC complex presented on the APC. This binding is necessary for “docking” the T cell onto the APC, but it’s not sufficient for activation.
3. Once this docking happens, the second step to activation is the co-stimulatory signals - Signal 2 - provided by the APC.
3.1. These signals are only upregulated when the APC has itself been activated by danger signals - like intracellular components spilling into the environment due to tissue damage, toxins from pathogens, or other stress signals. These are called Danger Associated Molecular Patterns (DAMPs), and termed Signal 0, as an extension of the 3 signal model by the Danger Model [3]. Signal 0 has to be present for any APC to be activated. One way to look at it is to think of APC as the transmitter of Signal 0 to Singal 2 to communicate what is going on in the environment to the cell type it is talking to.
3.2. In the absence of DAMPs, APCs remain in a resting state and do not provide co-stimulation.
3.3. If a T cell encounters its antigen (and “docks”) in the absence of co-stimulation, it either becomes anergic or is deleted.
4. If both Signal 1 and Signal 2 are present, the T cell becomes activated. Then comes Signal 3, which determines what type of T cell it becomes. The APC provides contextual cytokines (basically chemical signals) that determine whether that T cell will turn into a Th1, Th2, Th17, Treg, etc.
5. Activated cells need to be re-stimulated to persist. Once a T cell is activated it does not stay active indefinitely. After a certain amount of time - or certain rounds of doing its job - it needs to be re-stimulated by its cognate antigen – this time it can also be done by an unprofessional APC though (this is why vaccines work - and yes, they do work regardless of politics). If that doesn’t happen, it undergoes contraction: either dying off or transitioning into a memory state.
When you follow these rules, tolerance emerges from the system without the need for any “special” category of T cells. In the absence of Signal 2 (co-stimulation), even if a T cell recognizes its antigen (Signal 1), it doesn’t get activated. Instead, it becomes anergic or undergoes deletion, which is a way to maintain tolerance.
This framework is a more mechanistic, and frankly more appealing. Thanks to the Rule 5, you can’t really justify the need for a dedicated cell type like Tregs to actively resolve inflammation after an infection is cleared - a role often cited as one of the main reasons Tregs are considered “special” - because effector responses contract on their own unless danger signals persist.
Good, but this framework mostly supports the idea of maintaining tolerence by deletion, and doesn't really explain how any naive T cell can differentiate in the absence of danger, which is the case for (at least some) Treg differentiation. Plus, the data clearly shows that Tregs are essential for maintaining tolerance, so tolerance by deletion alone is abviously not enough. So we need to figure out how do Tregs emerge from the set of rules above, and this is where I am still battling.
First: how does Treg differentiation in the periphery (iTregs) work in this model? If danger is the only signal that activates professional APCs, upregulates co-stimulation (Signal 2), and allows for differentiation – and the absence of danger leads to deletion – then how, and more importantly, why, do we observe Tregs specific to harmless commensals at all? These microbes are not causing damage - so by the logic of the Danger Model, there shouldn’t be any activation in the first place. If you follow the rules above, for Signal 3 to work, Signal 2 is a must, and in the absence or danger, we shouldn't have Signal 2 at all.
One possible explanation is that these microbes are still bacteria after all, and ancient innate recognition mechanisms flag them as “potentially dangerous” by default - leading to Signal 2 even in the absence of actual damage (more like an anticipation of damage based on past experiences and a bit of stereotyping), which could then be paired with a non-inflammatory Signal 3. Fair – but this is more like a combination of self vs. non-self discrimination (me vs. most of the bacteria that historically didn't treat me well) with danger signals, not the Danger Model alone.
This also requires Signal 0 threshold to be very low, since you need to activate DCs in the presence of microbes but absence of real danger (like, a bit of stress just because there are microbes around). If it is so low that it's always crossed, then it becomes meaningless as a danger detector. It starts to look more like Signal 0 is required for basic APC function, not danger discrimination. After the APC is activated, most of the job is loaded on Signal 3 to determine the class, which includes the microenvironmental cues.
Plus, why would we need these iTregs anyway, wouldn't deleting these commensal-reactive clones by default resolve the issue?
This doesn’t just apply to the microbiome. How do we end up with iTregs specific to other peripheral tissues whose antigens aren’t presented in the thymus? If those T cells only ever encounter their antigen in the absence of danger - and there isn’t even an evolutionary reason to anticipate damage (since these are not bacterial antigens with historically conserved "nonself" regions) and upregulate Signal 2 - shouldn’t they just be deleted as well?
Even if a few autoreactive clones escape thymic deletion, a homeostatic periphery without no co-stimulation should either delete them or render them anergic. If some of them do get activated in the presence of danger signals - say, during an infection - you might theoretically get autoimmunity, and those autoreactive cells would continue to encounter their antigen (because self antigens are always around) - but after the infection resolves, this encounter happens in the absence of danger signals. Persistent antigen exposure without inflammatory context (remember rule 5 above?) eventually leads to deletion or functional impairment. This is also why even autoreactive memory cells don't live forever. Memory cells don’t require co-stimulation to be reactivated (that’s what makes them faster and more responsive than naive cells), but if they repeatedly encounter antigen without inflammation, they too undergo contraction or functional exhaustion over time.
As redundant as tTregs sound when you lay it out this way, data says otherwise, as discussed earlier. Even when peripheral regulatory mechanisms are intact, animal models develop autoimmunity in the absence of tTregs. Turns out peripheral deletion cannot fully compensate – and tTregs are not redundant after all.
At this point, if my arguments sound circular to you, it's because they are. This entire field is a maze, and that's why it's fun.
Nothing in biology makes sense except in the light of evolution. So to really understand Tregs, I thought I should first understand how - and why - adaptive immunity evolved in the first place. So here’s my version of events for the last 500 million years - at least, as far as it makes sense to me. I will try to avoid tracing back and reconstructing events based on what we know about the immune system now, but rather try to speculate what would be an evolutionarily useful solution as the clock moves forward – grounded in historical evidence, of course. Here it goes:
Let's start with the biggest reservoir of Tregs, the gut. Already before adaptive immunity evolved (and you only had enteric (intestinal) phagocytes5), jawed vertebrates likely had a gut microbiota, albeit a very simple one. I think the reason is, because at this point, the host already evolved to interpret the microenvironment as suggested by the Danger Model. It only increases phagocytic activity when there are DAMPs around. So when you sense danger, associated DAMPs (toxins, intracellular components released, etc.), tell all the phagocytes to kill whatever is around. This is still an improvement over not interpreting the microenvironment at all, and basing your activity only on conserved patterns on bacteria and killing all bacteria all the time (as a primitive self-nonself approach, which wouldn't let you kickstart a microbiome of your own). Instead, if you can evolve to interpret danger signals, you kill safe and useful bacteria during danger, but at least not all the time. This gives you an evolutionary advantage to harbor these microbes (and let them recolonize when it's not dangerous around) and benefit from them metabolically. In my opinion, there is no need for Tregs for this to work. However, being a symbiotic commensal versus a pathogen is not black and white, but a gradient. So this simple mechanism might not allow for a very complex microbiome.
However, different insults can lead to similar DAMPs. For example, having a pathogen invasion will likely destroy your epithelium, releasing intracellular components around. Similarly, any other environmental toxin or a physical insult to your gut epithelium (swallowing a stone while eating some leaves off the ground) would do the same. Does it make sense to increase the phagocytic activity in both cases?
At this point, we should stop and revise what phagocytic activity exactly does. Phagocytes are good at tracing bacteria and engulfing them, but they also release reactive oxygen species (ROS) to kill the invading pathogen, which are highly toxic to the environment - causing collateral damage and destroying your tissue along the way. If there is really an invading pathogen, it's probably worth it. But if not, it's really not a great idea to destroy your tissue all the time, if in reality what you really need to do is to repair it (e.g., in case of mechanical injury). This strategy is especially relevant for exposed mucosal sites like the gut or the eye that are vulnerable to environmental insults (and where a lot of Tregs and IgA are found as discussed above).
How would you differentiate different types of insults if the DAMPs are similar? An insult means the epithelial integrity is compromised, meaning that your epithelium is leaking whatever above it under the tissues below. That's also where your immune cells lie. So one way to understand whether there is actually an invading pathogen or not is to check what kind of microbes are getting under your skin (literally). If there is a pathogen, it will anyway be quite invasive and show up in high density under your epithelium. In this case, go and initiate a full blown phagocytic activity, that's fair. But if the problem is not a pathogen, you will have DAMPs (cells are still destroyed), but you will mostly have these symbiotic commensals leaking into the lamina propria (the tissue underneath your gut epithelium), at a higher density than normal. So differentiating by the density alone does not really work, but the type of bacteria that is breaching - wouldn't that be a good way to tell the root cause of the DAMPs?
To do this though, you need to remember which bacteria are safe and symbiotic, and then remind the phagocytic cells of it. One way to do this is to register this information during homeostasis (when there is no danger), and then remember it afterwards. There will anyway be occasional breach of these microbes to the lamina propria in the absence of danger. When the phagocyte engulfs this bacteria, and lacks the DAMPs (Signal 2), it should dump this information on another cell - to be reminded by it of this precious information when DAMPs are around and the phagocyte is already activated.
So I think this other cell might be the first Treg ever. I will call it fTreg, because this is not an academic paper and I get to use the notation I like.
Why do we need an additional cell type though? Why not directly do this on the phagocytic cell? Why can't it remember this on its own? There is evidence that this happens to some extent. But to fully do that, they need to develop a complex if/else logic: “if you see these nice commensals plus DAMPs then it's just a physical insult, but if you see these pathogens plus DAMPs then it's really getting infectious here”. Plus, phagocytes should probably act fast as soon as there are DAMPs around. Shutting them down might work better than turning them on slowly with complex decision making mechanisms. But this is just a speculation of mine.
To dump this information on the fTreg, we have to represent a two dimensional vector (dimension one being “bacteria or not”, and dimension two being “DAMPs are around or not”) by using only one dimension, a scalar. Why? Because the idea of fTreg is to be able to remind the phagocyte of what used to be safe back in calm times (homeostasis), regardless of whether the environment is safe or dangerous, now. This means the second dimension is useless, and fTreg can't rely on the context of the environment to make a decision. Since you are taking a binary vector of two and coding it in a scalar now, this means you need more than the conserved region recognition (bacteria vs. not, which is again binary) and be more specific. In Cryptography terms, you need more symbols in your library.
Let's assume fTreg was the first cell type to develop slighlty more diverse (not too much though) receptors6, because it needed a bit more than "bacteria vs not". When things are calm, and the environment is homeostatic, macrophages presented the small amount of commensal antigens they anyway engulf due to the occasional breach of the epithelium, in combination with the anti-inflammatory signals. fTregs, who had little discriminatory capactiy for these commensal antigens took that information in the presence of these homeostatic cues, and registered that information to be presented later. When pathogens hit, macrophages engulf disproportinately more pathogens compared to commensals, which did not activate the fTreg, becuase the receptor does not match. So nothing changed, macrophages continued their standard pathogen response. In contrast, when pathogen-"sterile" insult occured, macrophages got activated due to DAMPs as usual, but this time, they presented commensal antigens all over their surface. fTregs, who knew these commensal antigens back from homeostatic times, recognized them, docked on, and instructed phagocyte to stop all the aggressive things it is doing, but to start repairing - despite the appearently dangeorus environment. The more you refine this recognition of fTregs - the more symbols you have for this n-ary7 digit in your library -, the more you can become commensal-specific. But would that help develop a more complex microbiome?
I honestly don't think it does. It happened later, when B cells and antibody production (especially IgA in the gut) hit the scene. Because if you think about it, Tregs by themselves can only control what is beneath the epithelium, not above it – which is the gut lumen, where your microbiome lies, and the community composition gets complex.
This framework also aligns with the modern evidence on how Tregs promote wound healing by switching the macrophage phenotype from M1 (pro-inflammatory) to M2 (anti-inflammatory) [5], shifting the focus from chasing pathogens to collecting cell debris and accelerating regeneration and tissue remodeling. Even if the insult was pathogenic to begin with, after the pathogen is cleared, but the mechanical problem (the epithelial damage) persists, you need to understand that insult is over, and change your strategy accordingly. Which is another "job" assigned to Tregs in the current scientific literature.
So as a summary, I would like to propose an alternative opinion, which I have no idea how to prove experimentally, but computationally I am working on it: Tregs might have allowed for a more complex microbiome eventually, but I don't think this is the evolutionary advantage that led to their emergence [4]. They first emerged in the mucosal tissues to differentiate the type of insult - pathogen-sterile versus pathogenic -, since multiple types of insults can lead to the release of similar DAMPs. However, if you can tailor the response to the type of injury and avoid unnecessary collateral damage, then you gain an evolutionary advantage. Piggybanking on this regulatory mechanism is probably what allowed for the rest of the complexity of the adaptive immune system to evolve, as beautifully described by Müller et al [4].
References
[1] Matzinger, Polly. The danger model: a renewed sense of self. Science, 2002.[2] Matzinger, Polly, and Tirumalai Kamala. Tissue-based class control: the other side of tolerance. Nature Reviews Immunology, 2011.
[3] Matzinger, Polly. Essay 1: the Danger model in its historical context. Scandinavian journal of immunology, 2001.
[4] Müller, Viktor, Rob J. De Boer, Sebastian Bonhoeffer, and Eörs Szathmáry. An evolutionary perspective on the systems of adaptive immunity. Biological Reviews, 2018.
[5] Bhavana Nayer, Jean L Tan, Yasmin K Alshoubaki, Yen-Zhen Lu, Julien M D Legrand, S. Lau, Nan Hu, Anthony J Park, Xiao-Nong Wang, D. Amann-Zalcenstein, Peter F. Hickey, Trevor Wilson, Gisela A. Kuhn, Ralph Müller, Ajithkumar Vasanthakumar, Shizuo Akira and Mikaël M Martino. Local administration of regulatory T cells promotes tissue healing. Nature communications, 2024.
1 For a nice summary, see “A journey to your self: The vague definition of immune self and its practical implications”, section A Brief History of Immune Self by Koncz et al. ↩
2 This set of lectures is a hidden gem: https://videocast.nih.gov/watch=17095 ↩
3 In the Danger Model, tumors become dangerous only once they begin to cause damage—e.g., outgrowing their blood supply and start dying in an inflammatory way (necrosis). By then, the immune system may already have developed tolerance by deleting most of the autoreactive clones, as explained in Matzinger’s lectures.↩
4The reason is the existence of idiotype-anti-idiotype interactionsthe, stemming from Jerne’s Idiotypic Network Theory - but that requires it's own blog post.↩
5I will simply use the term phagocyte for any type of cell within the body capable of engulfing and absorbing bacteria and other small cells and particles, and then present them on their surface via MHC.↩
6In ref. [4], this is called "limited somatic diversification of selected (safe) receptors. It's an amazing paper to read as well.↩
7An n-ary digit is the generalization of a binary digit (bit), where instead of 2 possible values (0 or 1), you have n possible values (like ternary with 3 values: 0, 1, 2, or decimal with 10 values: 0-9).↩
