NL
DE
8 May 2020
This brings us to a question that has intrigued immunologists for decades: how do we get this repertoire that detects all those foreign pathogens, but still tolerates the entire “dictionary” of self peptides naturally present on healthy cells?
Well… Not really.
Healthy cells come in many flavours, and all of them present self peptides on their surface. This adds up to roughly half a million self peptides that our T cells should learn to ignore — about five times the number of words in the dictionary from our little game. And just like you can’t possibly memorise an entire French dictionary in a single hour, it seems unlikely that developing T cells can scan the entire “self peptidome” during their short stay in the thymic medulla. In fact, the current best estimates suggest that they see only a small fraction¹. But if negative selection is that incomplete, so that most self peptides do not get their T cells pruned this way, why bother with a process that kills half (!) of the T cells that would otherwise “graduate” from the thymus? Or can negative selection still be useful — even if incomplete?
Not necessarily. If I ask you right now which of the words “indoda” and “fièvre” is French, you can probably give the correct answer in a heartbeat — even if you’ve never heard either word before. Why? Because your brain just told you that “fièvre” kind of looks like “chèvre”, a word you have likely seen on the packs of goat cheese in your local supermarket. Put simply: to recognise that a word is French, you don’t need to memorise my entire dictionary. You may just win our bet after quickly screening it for some examples of typical French words.
This process of “learning by example” is called generalisation, and our brains are famous for it. But if we can learn to recognise French without memorising an entire dictionary, could T cells somehow learn to tolerate self without seeing all self peptides in the thymus?
It turns out that the process of negative selection itself lets the repertoire “learn” ². To show this, we built a simple computer model of a “T cell repertoire” undergoing negative selection. But rather than using T cells recognising peptides, we started with T cells recognising strings (letter sequences) from different languages (for example, one of those T cells might respond to the “èvre” letter combination that occurs in both “chèvre” and “fièvre”).
Instead of T cells recognising “self” or “foreign” peptides, we first modeled T cells recognising strings from a “self” or “foreign” language.
Here’s what happened: after we let these T cells undergo negative selection, they could discriminate between strings from “self” and “foreign” languages — even the ones they had not seen during the negative selection process. In other words, they learned by example.
How? Well, T cells may be specific, but they can still respond to multiple peptides (or in this case, strings) as long as they “look alike” to the T cell receptor’s binding interface. And this cross-reactivity is a game-changer.
Imagine a computer-thymus containing the word “chèvre”. Even if it lacks the word “fièvre”, some of the cross-reactive T cells specific for “fièvre” may still respond to “chèvre” in the thymus — so they are negatively selected anyway. This is how negative selection not only removes T cells reacting to the limited number of self peptides in the thymus, but also T cells responding to their non-thymic lookalikes.
But even though our model of language-learning T cells shows that a T cell repertoire can generalise in principle, it’s not that easy in practice.
And that’s where it gets tricky. Because while French words look more like other French words than they look like words from a different language, we found that the same does not hold for self and foreign peptides. In a way, this makes sense; the “rules of language” that peptides follow likely depend more on their protein’s function than on the organism they come from. A French chair will resemble an Egyptian chair more than a French spoon. A human kinase may look more like a viral kinase than like a human ion channel. Peptides might follow language rules — but these are not necessarily organism-dependent.
In fact, when we applied our computer model to peptides recognisable by T cells (rather than strings), the “learned” self-foreign discrimination disappeared almost completely. This happened because viral peptides often resembled self peptides — more than they resembled other viral peptides. Just imagine I didn’t ask you to distinguish “fièvre” from “indoda”, but from “quèvre” (a fake French word I just made up). Would that still be easy without a dictionary?
The problem here is: these T cells did generalise. They just generalised the wrong thing, and this ended up removing just as many foreign- as self-reactive T cells. So where does that leave the role of negative selection in self-foreign discrimination?
But that just brings us back to the question: what’s the role of negative selection in all of this? Does it really just ensure tolerance against a minority of self peptides, while the other “unseen” self peptides are recognised just as much as foreign peptides? Or… could “central” tolerance somehow still play a more central role in the self-foreign discrimination problem?
It turns out there is hope for a function of negative selection after all. Because although it is hard for negative selection to achieve self-foreign discrimination, it’s not impossible: in our computer model, we showed that choosing the thymic peptides smartly can make a big difference. Just like you probably shouldn’t waste your time learning that “chèvre”, “bièvre”, “lièvre”, “mièvre”, “nièvre”, and “fièvre” are all French, thymic selection gets a bit more efficient if the thymus ensures some variety in the self peptides it presents. And while that still doesn’t make it any easier to distinguish “fièvre” from the non-existent “quèvre”, it turns out that it does improve discrimination between self and foreign peptides on average. And the best news is this: our thymus could accomplish this effect through something as simple as a bias for peptides with rare amino acids.
Does this actually happen? That, we don’t know yet. If our immune system can indeed generalise, this would mean that the brain is not the only “intelligent” organ in our body. But no matter how it arises, our immune system’s ability to discriminate self from foreign is impressive. Most days, my T cells are winning a bet I never could — and that makes them plenty smart to me.
¹ It remains unknown exactly how many self peptides T cells do see, but the estimates here and here suggest they don’t see all (or even nearly all), and these two studies (1,2) seem to confirm that by showing that way more self-reactive T cells survive selection than originally thought.
² Actually, the idea that some mechanisms of the immune system can allow for learning inspired an entire machine learning field in the nineties.
Learning French
Suppose I made you a bet. I give you a French dictionary — one of those old heavy tomes with about 100,000 words — and leave you alone with it for an hour or so. (I am assuming here that you speak no or very little French; if you’re a fluent speaker, just imagine any language that seems like gibberish to you). When your time is up, you return the dictionary and I show you a bunch of words in either French or any other language you don’t know. If you can tell me which ones were in the dictionary, you win: let’s make it ten dollars per word. But for every word you guess wrong, you pay me ten dollars. Would you take that bet? Hold on to your answer — we’ll get back to this later.This game may seem silly, but our immune system’s T cells have to solve a problem just like it.
T cells, trees, and the big Question
T cells face a daunting task. They must provide highly specific immunity against any pathogen we may encounter — be it Covid-19, malaria, or next year’s flu. But to avoid autoimmunity, they should also not respond to any of our healthy cells. To meet both of these demands, our T cell “repertoire” contains millions of cells that are all slightly different. A single T cell can never be both highly specific and provide broad immunity (that would be like a rock band trying to keep its niche audience and to become more mainstream — impossible). But many hands make light work: while each individual T cell in our repertoire recognises only a limited number of specific peptides, together, they may still detect more foreign peptides than there are trees on Earth.This brings us to a question that has intrigued immunologists for decades: how do we get this repertoire that detects all those foreign pathogens, but still tolerates the entire “dictionary” of self peptides naturally present on healthy cells?
The thymic dictionary of “self”
Part of the answer lies in the education newly developed T cells receive in the thymus. In the process we know as negative selection, T cells are exposed to many different self peptides from the human proteome. As long as they don’t respond to any of these, they pass the test and enter the bloodstream; but if they do respond to self, they are silenced before they can cause auto-immunity. Makes sense, right? Just make a lot of different T cells, get rid of any that happen to be self-reactive, et voila — we have a repertoire that detects foreign but tolerates self. Problem solved.Well… Not really.
Healthy cells come in many flavours, and all of them present self peptides on their surface. This adds up to roughly half a million self peptides that our T cells should learn to ignore — about five times the number of words in the dictionary from our little game. And just like you can’t possibly memorise an entire French dictionary in a single hour, it seems unlikely that developing T cells can scan the entire “self peptidome” during their short stay in the thymic medulla. In fact, the current best estimates suggest that they see only a small fraction¹. But if negative selection is that incomplete, so that most self peptides do not get their T cells pruned this way, why bother with a process that kills half (!) of the T cells that would otherwise “graduate” from the thymus? Or can negative selection still be useful — even if incomplete?
Goat cheese and generalisation
To answer that question, we return to our game. You know no French and have only one hour to learn some. Mission impossible?Not necessarily. If I ask you right now which of the words “indoda” and “fièvre” is French, you can probably give the correct answer in a heartbeat — even if you’ve never heard either word before. Why? Because your brain just told you that “fièvre” kind of looks like “chèvre”, a word you have likely seen on the packs of goat cheese in your local supermarket. Put simply: to recognise that a word is French, you don’t need to memorise my entire dictionary. You may just win our bet after quickly screening it for some examples of typical French words.
This process of “learning by example” is called generalisation, and our brains are famous for it. But if we can learn to recognise French without memorising an entire dictionary, could T cells somehow learn to tolerate self without seeing all self peptides in the thymus?
Why T cells could indeed learn French
There is one obvious problem here: unlike you and me, T cells do not have brains. So how could they learn anything — by example or otherwise?It turns out that the process of negative selection itself lets the repertoire “learn” ². To show this, we built a simple computer model of a “T cell repertoire” undergoing negative selection. But rather than using T cells recognising peptides, we started with T cells recognising strings (letter sequences) from different languages (for example, one of those T cells might respond to the “èvre” letter combination that occurs in both “chèvre” and “fièvre”).
Instead of T cells recognising “self” or “foreign” peptides, we first modeled T cells recognising strings from a “self” or “foreign” language.
Here’s what happened: after we let these T cells undergo negative selection, they could discriminate between strings from “self” and “foreign” languages — even the ones they had not seen during the negative selection process. In other words, they learned by example.
How? Well, T cells may be specific, but they can still respond to multiple peptides (or in this case, strings) as long as they “look alike” to the T cell receptor’s binding interface. And this cross-reactivity is a game-changer.
Imagine a computer-thymus containing the word “chèvre”. Even if it lacks the word “fièvre”, some of the cross-reactive T cells specific for “fièvre” may still respond to “chèvre” in the thymus — so they are negatively selected anyway. This is how negative selection not only removes T cells reacting to the limited number of self peptides in the thymus, but also T cells responding to their non-thymic lookalikes.
But even though our model of language-learning T cells shows that a T cell repertoire can generalise in principle, it’s not that easy in practice.
Foreign peptides and fake French
The (problematic) assumption we have made so far is that “self” and “foreign” are somehow intrinsically different — not just in the sense that individual peptides differ from each other, but that they follow different underlying rules. It was easy for you to tell the difference between “indoda” and “fièvre” because, as a rule, you expect French words to end in “èvre”, but not in “doda”. In fact, it is those underlying rules that you “learn by example” when you scan a French dictionary. But what underlying rules distinguish self from foreign peptides? What underlying language do these peptides come from, and must our T cells learn from the thymic examples?And that’s where it gets tricky. Because while French words look more like other French words than they look like words from a different language, we found that the same does not hold for self and foreign peptides. In a way, this makes sense; the “rules of language” that peptides follow likely depend more on their protein’s function than on the organism they come from. A French chair will resemble an Egyptian chair more than a French spoon. A human kinase may look more like a viral kinase than like a human ion channel. Peptides might follow language rules — but these are not necessarily organism-dependent.
In fact, when we applied our computer model to peptides recognisable by T cells (rather than strings), the “learned” self-foreign discrimination disappeared almost completely. This happened because viral peptides often resembled self peptides — more than they resembled other viral peptides. Just imagine I didn’t ask you to distinguish “fièvre” from “indoda”, but from “quèvre” (a fake French word I just made up). Would that still be easy without a dictionary?
The problem here is: these T cells did generalise. They just generalised the wrong thing, and this ended up removing just as many foreign- as self-reactive T cells. So where does that leave the role of negative selection in self-foreign discrimination?
How a smart thymus helps T cells learn
First of all, it confirms something we already knew: negative selection can never establish robust self-foreign discrimination all by itself. This is no surprise — the very existence of autoimmunity proves that negative selection is not infallible. Immunologists have known this for years, and have uncovered a rich ecosystem of other mechanisms our bodies use to keep auto-immunity at bay. These so-called peripheral tolerance mechanisms are highly diverse but share a common goal: to plug the “leaks” in the central tolerance established by the thymus. The finding that self-foreign discrimination is difficult just makes those leaks a lot bigger, further stressing the importance of peripheral tolerance.But that just brings us back to the question: what’s the role of negative selection in all of this? Does it really just ensure tolerance against a minority of self peptides, while the other “unseen” self peptides are recognised just as much as foreign peptides? Or… could “central” tolerance somehow still play a more central role in the self-foreign discrimination problem?
It turns out there is hope for a function of negative selection after all. Because although it is hard for negative selection to achieve self-foreign discrimination, it’s not impossible: in our computer model, we showed that choosing the thymic peptides smartly can make a big difference. Just like you probably shouldn’t waste your time learning that “chèvre”, “bièvre”, “lièvre”, “mièvre”, “nièvre”, and “fièvre” are all French, thymic selection gets a bit more efficient if the thymus ensures some variety in the self peptides it presents. And while that still doesn’t make it any easier to distinguish “fièvre” from the non-existent “quèvre”, it turns out that it does improve discrimination between self and foreign peptides on average. And the best news is this: our thymus could accomplish this effect through something as simple as a bias for peptides with rare amino acids.
Does this actually happen? That, we don’t know yet. If our immune system can indeed generalise, this would mean that the brain is not the only “intelligent” organ in our body. But no matter how it arises, our immune system’s ability to discriminate self from foreign is impressive. Most days, my T cells are winning a bet I never could — and that makes them plenty smart to me.
¹ It remains unknown exactly how many self peptides T cells do see, but the estimates here and here suggest they don’t see all (or even nearly all), and these two studies (1,2) seem to confirm that by showing that way more self-reactive T cells survive selection than originally thought.
² Actually, the idea that some mechanisms of the immune system can allow for learning inspired an entire machine learning field in the nineties.
