The Generative Grammar of the Immune System

In times of confinement due to SARS-CoV-2 virus, it is inevitable to dedicate some hopeful thoughts to the immune system. And it also seems like a good time to use our thoughts about the immune system (and its surprising effectiveness in fighting pathogens) to understand a little better the equally surprising efficiency with which our brain develops language proficiency from environmental data.

The idea of comparing the immune system with generative grammar is not original. The comparison was expressly established by immunologist Niels K. Jerne, who titled his Nobel Prize acceptance lecture The Generative Grammar of the Immune System (revealing that the title of this post is not original either).

Jerne’s analogy is rather metaphorical: it is based on the Chomskyan notion of a generative grammar understood as a system of knowledge that allows the speaker to generate sentences appropriate to new situations, and to understand sentences never heard before. As Jerne points out, the immune system is generative because it is able to find an adequate response to any antigen it faces, even though the antigen has never been encountered before.

What is relevant, and mysterious, is how these processes happen: how the immune system “learns” to create the right molecule (the antibody) for whatever antigen (stimulus) it faces, and how the human brain “learns” a generative knowledge system from the linguistic stimuli of the environment.

In both cases, and independently in each field, science has shown that the word learn is probably not the right one, at least in the usual sense. In fact, the relationship between the immune system and the language faculty that I am going to consider is not metaphorical. It is based on a parallelism pointed out by Massimo Piattelli-Palmarini (in a remarkable 1989 article) in relation to two basic models of biological evolution and development: the “instructive” model and the “selective” model.

The essential idea of an instructive model is that external stimuli give their character to the receiving system, that is, in some way the system extracts its structure from external stimuli. In a selective model, on the contrary, the stimuli also act on a system, but the system is already structured, so that it detects environmental stimuli and selects the relevant ones. The difference between both models can be further clarified if we turn to the definitions of molecular biologist Antoine Danchin (quoted by Piatelli-Palmarini): ‘Instructive theories postulate the existence of a causal agent, exterior to the system, that directs its evolution […] Selective theories, on the opposite, leave to contingent interactions the only driving force that makes systems evolve’.

A more informal way of understanding the difference between the two models, also suggested by Piattelli-Palmarini, may be to consider the difference between buying a coat in a department store or ordering it from a tailor. In the first case, the customer selects the coat that best suits her from among the available sizes (a selective process), while in the second case the tailor measures the client and makes a custom-made one (an instructive process). The common sense notion of learning corresponds to the instructive model.

Piattelli-Palmarini suggests that the common fate of Jerne and Chomsky has been that in both cases they have favoured the transition from instructive models to selective models. In fact, Jerne received the Nobel Prize in medicine in 1984 precisely for his contribution to a selective theory of the immune system.

In Chomsky’s case, as is well known, this transition occurs with the rejection of the empiricist conception of language acquisition, according to which the brain builds the knowledge system (the internal language of each person) extracting its structure from the linguistic data of the environment, through general learning procedures (such as induction, analogy or imitation). This view is opposed by Chomsky’s so-called “argument of the poverty of stimulus”: the information provided by the linguistic stimuli that the child receives is insufficient to justify the entire structure of the knowledge system finally obtained, so it must be postulated that some essential principles of the structure of that knowledge system are already part of the organism, that is, they are innate.

If we return to the coat analogy, the question would be whether the child’s brain makes a custom-made coat to the stimulus of the environment or, on the contrary, selects one of the available coats, depending on the “size” of the stimulus. Chomsky’s answer is the second one, so unintuitive that even today it is fiercely rejected. But the answer that modern immunology offers, while fully accepted, is no less counterintuitive than Chomsky’s answer. Let’s see why this episode in modern biology can enlighten us when it comes to language acquisition.

As Jerne points out in his lecture, the immune system is an organ in the body of any vertebrate. This organ is made up of cells called lymphocytes, which in a human being are around 1% of the body mass, in an estimated number of about 1012 units (that is, more numerous than neurons).

It is important at this point to distinguish between the innate and the adaptive immune systems. The innate immune system includes many more elements than lymphocytes and is a kind of generic first barrier (that is, it acts independently of the pathogen that attacks the body). It includes, among other elements, the skin, mucous, gastric acids, inflammation factors, and macrophages. The so-called adaptive immune system, which is also innate, is the one that Jerne refers to, and the one that interests us here, because it has a generative component. Simply put, lymphocytes (B lymphocytes in particular) are capable of generating specific antibodies for each type of antigen. It is called adaptive because it adapts its response to the type of antigen it faces. There are no generic antibodies, but each antibody is specific for each possible antigen. It is also adaptive in the sense that the organism keeps memory of previous attacks, so that in a subsequent infection the response is faster and more effective (which is the basis of the immunisation that we want to achieve with the future vaccine against SARS-CoV-2).

But this adaptive immune system is also innate. In fact, this is the most surprising thing: this immune system works as a selective model and not, as would be expected, as an instructive one.

Intuitively, since the immune system is capable of creating a specific antibody for each possible antigen, including those that it has not found before and those that are new (that is, that have never been found before in previous evolutionary stages, because they have been created in the laboratory), it is logical to think that the immune system behaves according to an instructive model: the antigen molecule determines the structure of the antibody molecule that has to bind it (so that later the invader is destroyed). That is, the system would somehow “learn” what is the structure of the harmful molecule to make it a “tailored coat”.

But no. What immunologists (led by Jerne) discovered is that things are not like that. The immune system has already prepared the possible antibodies that it will need. Jerne points out that human lymphocytes can create about 10 million different antibodies! (The incommensurability of the figure can be better assessed if we consider that the human body creates “only” about 10,000 different types of proteins, including all hormones, enzymes, etc.).

According to Jerne, the immune system is “complete”, that is, it can respond with specific antibodies ‘to any molecule existing in the world, including molecules that the system has never before encountered’. Contrary to intuition, the adaptive immune system works as a selective system.

As Piatelli-Palmarini points out, the idea of ‘ready-made repertoires was for a long time considered to be, literally, unthinkable in immunology’. For this reason, for decades it was thought that the specific antibodies that the organism generates are the result of an instructive and non-selective process, so that the antigen “dictates” to the body how the antibodies that recognize it should be. As the same author relates, Paul Ehrlich had proposed a selective theory of antibody formation in 1897. But this conception received a death blow when in 1912 Karl Landsteiner provided experimental evidence against this point of view showing that an organism could develop very specific antibodies against artificial substances, that is, compounds that have never been found by any organism in any type of “natural” virus or bacteria. It was then considered that it was impossible for organisms to already have antibodies against artificial substances, and selectivism was considered not only improbable, but even inconceivable: Antibodies against invented or artificial substances could not be innate, since they could never have been useful in the past and therefore preserved by natural selection. As Piattelli-Palmarini says, ‘if these antibodies were innate, then the immune system, so necessary to survival, would be profligate, wasteful, maladaptive’, an unacceptable idea in a typically adaptive (instructive) conception of natural selection.

Of course, after Landsteiner’s experiments, no immunologist thought of selective theories. Until 1955, when Jerne again proposed a selective theory. And he did it because he wanted to understand how the system learned to make the appropriate antibodies for each pathogen.

It is inevitable to observe a parallelism with generative grammar in this regard (parallelism that arises even without a time travel to find the Ehrlich of linguistics in rationalist grammar, as Chomsky did in his controversial 1966 Cartesian Linguistics). Evidence based on common sense and observation of the progressive learning of children’s language, apparently parallel to the development of other capacities, leads to the well-known conclusion in terms of instructive learning. But, just as it happens in the field of evolutionary theory or immunology, we are left without an explanation of how the development processes really happen: how an organism “learns” to have wings, how it “learns” to generate adequate antibodies to each pathogenic element, or how it “learns” to speak the language of its environment.

Nowadays all immunologists accept that an organism has innately incorporated the capacity to generate the repertoire of antibodies, and that this repertoire is capable of finding an antibody for each antigen that is presented, either natural or artificial, whether it is part of its evolutionary history or not. An important conclusion of this state of affairs is that if the organism were to encounter an antigen that, against all probability, did not have its “image” in the expected 10 million antibodies, the organism simply could not do anything against it. To quote again Piattelli-Palmarini, ‘a really new antigen would be literally invisible to the immune system: the organism would develop no response to it. This is, typically, a selective theory’.

This implies that the organism cannot learn from the environment. Note the similarity of this conclusion to what Chomsky has called Plato’s problem (that is, how we know so much with so little external information). In a strictly adaptationist interpretation of natural evolution (in an “instructive” context), the notion of natural selection is associated with the loss of what is unnecessary. Given that lymphocytes that generate antibodies against artificial substances could not have adaptive value, the most reasonable way to explain their presence was derived from a process of “learning” or “instruction” from the environment. The same is true in the field of language: the existence of an innate faculty of language (FL) is considered implausible compared to the most economical option of a reduced number of multifunctional learning principles that build each language in people’s minds by copying them from the environment.

But we should already know that common sense is not the key to scientific understanding (although it is very useful for everyday life). There is every reason to think that in cognitive science there has also been a transition from instructive learning models to selective learning models. Today it is assumed that there is an innate component in cognitive functions such as vision, memory or emotions, so language should not be an exception, especially considering that generative linguistics was a pioneer in that transit in the 1950s.

Of course, in the case of language we cannot literally apply the analogy with the immune system. It cannot be postulated that the human genome carries the more than 6,000 existing languages from which one (or more than one) is selected depending on the stimuli of the environment. Languages are not innate. What could be innate are the principles that determine how they are built. The same is true in immunology: the 10 million different antibodies could not be encoded in just 20,000 genes in the human genome.

Indeed, there are numerous aspects of languages that speakers learn from the environment by imitation, analogy, and induction. The point is that not everything that constitutes a human language (that is, a knowledge system) can be learned in this way. Many aspects of language cannot be learned from the environment if the organism does not severely restrict the possible options, both in the field of lexicon (what types of possible meaning can a word have?), and in phonology (what phonetic features are available to the articulatory and perceptual organs?), as well as in syntax (what computational procedures can the brain implement?).

The FL includes at least three major components: a conceptual system (responsible for the lexical meaning), a computational system (responsible for the syntax that combines concepts to create new concepts and thoughts), and a sensorimotor system (which in the case of oral languages determines a specific range of usable phonetic features, syllabic structures, etc.).

All these systems are the result of natural evolution, they have fixed properties, they are shared by all human beings, and all of them, each one in its domain, severely restrict what a possible human language is (what kind of words, what kind of phonemes and prosody, what kind of phrases can they have). All of that is excluded from instructive learning, fortunately.

The development of language in a person can be modelled as the transition from the initial state of FL (FL0) to the steady state (FLE). The steady state FLE is a language, a person’s language organ. The transition from FL0 to FLE is crucially conditioned by stimuli from the environment (particularly the primary linguistic data perceived by children). An environment in which Japanese is spoken will make children speak Japanese, and an environment in which Spanish is spoken will make children speak Spanish. They are two very different languages, but they are both restricted by the same common principles, which are specific to the systems integrated within FL0. The function of environmental data is therefore to select a certain number of open options, that is, not specified by FL0: for example, which phonetic features will be distinctive, how many colours will have their own lexical label, or if the direct object will go before or after the verb.

The so-called Principles and Parameters model (Chomsky 1981) is a precise embodiment of a selective theory of language acquisition. According to this model, FL0 consists of a set of principles, which are common and invariable. But these principles have unspecified aspects that have to be determined during the ontogenetic development of FL in order for the system to become operative. A common metaphor for characterizing that model (see, e.g. Chomsky 2000) is that of the switch panel. According to this view, FL0 can be seen as a fixed network connected to the switch panel: the network represents the principles, while the switches are options to be determined by experience. There are switches that determine the position of others (parameters), and the task of acquiring a language would essentially consist of choosing the setting of those switches based on experience. More specifically, the role of linguistic stimuli is not to provide an inductive basis for discovering principles, but to provide the information necessary to set switches in the right combination. In this theoretical context, the stimuli of the environment (the antigens) provide information on which options of the parameters (antibodies) have been chosen in the surrounding language, that is, learning consists of selecting the appropriate parameters for each possible language.

To give a very simplified example: there are languages in which cases (nominative, accusative, genitive, etc.) are expressed morphologically (such as Japanese or Basque), while there are languages in which they are not expressed (such as English or Spanish, ignoring pronouns). Suppose the innate principle of FL0 is that “every noun phrase (NP) that enters a sentence must have an assigned case”. Let us assume, for example, that a caseless NP would be invisible to a certain component of the mind, producing an ungrammatical sentence. Of course, such a principle (let’s call it the case principle) is not genetically encoded nor shaped by natural selection. We must consider it as a hypothesis that the scientist formulates, and that maybe it is a consequence of the effect of interactions between other more basic principles or mental processes that we simply do not yet know. What is relevant now is that the case principle does not specify that cases should be expressed morphologically (audibly) in NPs. In fact, there are languages in which case is expressed and languages in which it is not. There are languages in which certain cases are expressed and others are not, languages in which cases are expressed in pronouns or determiners but not in nouns, or languages in which certain cases as expressed with suffixes and others with prepositions. But all this variation is systematic and constrained by underlying uniformity. Thus, there is a universal hierarchy of cases (determined by FL0), roughly represented as follows (see Blake 1994):

nominative > accusative > genitive > dative > locative > ablative > others

As Caha (2009) has shown, cases can be conceived of as interpretable functional categories that form a fine-grained and universal functional sequence on top of NPs. Variations observed in the surface patterns are sensitive to this universal hierarchy. So, if a language marks the dative case, it also marks the cases on the left (that is, genitive, accusative and nominative). If in a language the dative case is expressed with a preposition and not with a suffix, then all cases to its right do as well. There are no languages in which only the nominative and the dative are marked, or in which the genitive is marked and the accusative is not. 

The role of environmental stimuli is not to provide very young children with information so that they learn what cases are, and deduce the case principle or the case hierarchy (something that they could not do in any way). Rather, the stimuli provide information for the children to (instructively) learn which NPs should be case marked in the language they are acquiring, that is, information for them to select the appropriate parametric values.

If a child were to find a language inconsistent with principles of FL0, she would either not be able to learn it spontaneously, or she would adapt it to these principles (as happens when children develop creole languages from pidgins).

The instructive theories of the first half of the 20th century in the field of immunology assumed that the synthesis of antibodies could only be done from a set of formless “globulins” willing to receive the “primer” by external antigens. Similarly, the instructive theories of language learning still assume today that the genesis and development of language is governed solely by general principles of learning, analogous to those we employ to develop specifically human but non-instinctive systems of knowledge, such as playing chess, solving quadratic equations or understanding social institutions or cultural phenomena, such as the baroque or romanticism.

The transition from instructive to selective models has been notable in biology in recent decades, both in the field of evolutionary theory and in developmental biology or, as we have seen, in immunology. Of course, this fact is not an argument when it comes to language, but it would not be smart not to take it into account, if what we are studying is an attribute of our species.

As Chomsky has said on some occasions, there are many things that we cannot learn, simply because we are not designed for it, but that is actually fortunate, because that fact guarantees that we can learn many other things.

Spanish version at

One thought on “The Generative Grammar of the Immune System

  1. I love that Jerne lecture, and I remember finding it quite remarkable that he takes the position that if only immunology could be more like generative grammar, it could be a proper science. That was quite striking when I first read it because it seems like the inverse of the standard position today. This inversion says something about the state of science generally and linguistics specifically.

    I also found your coat analogy instructive, but because I see the two options as two varieties of a selective procedure. Obviously, buying a coat off a rack at a store is a matter of selecting an item off of a list, but, in a way, so is having a coat tailor-made. When you get a coat tailor-made you select style, fabric, colour, etc. and then the tailor takes measurements. All of these inputs are parameters which, when combined with the principles of coat-making, yield your desired coat. A set of parameters can also be written out and, assuming a finite number of possibilities for each parameter, all the other possible sets of parameters can be written out. Thus, the process of having a coat tailor-made is reduced to choosing from a list.

    The analogy is instructive, I think, because instructive models, while being “obviously true”, seem to be extremely hard, perhaps impossible to conceptualize.

    Liked by 1 person

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s