Learning of signalling networks — at the level of their components

Molecular mechanisms of neuronal learning became well established [1]. However, much less is known about the regulation of learning at the individual, non-neuronal cells. Recent findings gave further evidence that learning, indeed occurs in unicellular organisms, as well as in individual cells of various tissues other than neurons, even in rather sophisticated forms [2]. In our paper we define cellular learning as an adaptive response to a stimulus, when the stimulus is repeated in a short time. This leaves out many classical models of learning (such as Pavlovian conditional learning) from our discussion. However, such a simplification greatly helps the identification of molecular mechanisms, which become increasingly obscured when long-term, multistep adaptation phenomena are examined, such as cell differentiation or tumour development. Several experiments in budding yeast, Arabidopsis or rice cells, mouse fibroblasts or murine CD8+ memory cells showed the formation of molecular memory resulting in a faster, larger, more sensitive and/or more robust response after the second signal than the first [35]) to humans (e.g. RACK1 [36]). These proteins, once they became activated, maintain larger pathway segments pre-organized, ready to respond to the second stimulus faster, and stronger. We note that similar signalling cascade memories may be postulated in each signalling pathway. As examples the JNK and Hippo pathway cascades are enhanced by the scaffolding proteins JIP1 and MOB1A, respectively [37, 38]. These scaffolds may prime these pathways giving a stronger second response after an initial stimulus.

The third example expands the above idea of pathway organization and consequent cellular memory formation to networks other than signalling networks, such as metabolic networks. Analysis of non-Markovian chemical reaction networks on gene expression showed that molecular memory of protein synthesis and degradation may induce feedback, bimodality and switch behaviour, and may fine tune gene expression noise, all components of molecular memory [39]. Even bacteria use their inner membrane as a scaffold [91].

Learning, desensitisation, stress and ageing in system-level signalling: phases of the same response?

There are only a few reports of time-dependent changes in cellular signalling upon shorter versus longer extracellular signals. One of these was made on rat pancreatic islets, where 300 mg/dl, high concentrations of extracellular glucose induced a stronger response after 3 h, which turned to glucose insensitivity after 6 h [69]. This is clearly a two-step cellular response, where the pancreatic beta-cells first learned the presence of glucose and made a preconditioned, stronger response to them. However, after a longer time, an overload occurred and the cells turned to insensitive to glucose. Note that under natural conditions high glucose is only a temporary, postprandial event. A similar effect was observed, when 3T3-L1 cultured fat cells were exposed to insulin receptor auto-antibodies. Acute administration of anti-receptor antibodies induced a more efficient deoxyglucose uptake, while prolonged exposure led to insulin insensitivity [92]. Here again, high levels of the original agonist, insulin are also only transient, postprandial events.

If we take the examples of (1) cellular learning and development of cellular memory after a few repeated stimuli [10, 25, 54, 55]; (2) the desensitisation of signalling after a prolonged exposure to the signal [61,62,63,64,65,66,67,68,69,70] and (3) the result of the above studies [69, 92] (where in an extended timescale first learning and then desensitisation was observed) together, the conclusion can be drawn that, in fact, cellular learning and desensitisation may be consequent phases of the same response. The cell first becomes more ‘alert’ and more ‘ready’ to respond to environmental changes. However, after a prolonged stimulus its signalling network becomes ‘saturated’ and starts to ‘protect itself’ (Fig. 2). We may also add stress [74,75,91] to this spectrum, where ‘fatigue’ of the signalling network is induced by both as examples of overloading short-term (stress) and long-term (ageing) changes (Fig. 2). We note that comparative studies of agonist-, stress- and ageing-induced desensitisation are missing. Therefore their combination on Fig. 2 is only illustrative and hypothetical.

Fig. 2
figure 2

Repeated signals induce cellular learning; persistent signals lead to cellular desensitisation; permanent signals (such as the accumulated signals and damage in ageing) overload the signalling network and provoke cellular stress. Note that on the contrary to the few studies showing a change from cellular learning to desensitisation in the same system [69, 92], comparative studies of agonist-, stress- and ageing-induced desensitisation are missing. Therefore, this figure is only hypothetical, illustrative and by no means quantitative

Full size image

Discrimination between repeated and unexpected signals: perhaps as also a property of single cells?

Desensitisation protects the system from the overload of inputs. At a low level of complexity overload can be understood that too many of the same signal within a certain time (where the system may adjust its thresholds defining the “too many” and the “within a certain time”). At a higher level of complexity overload also occurs, if the system is not able to make ‘groups’ of similar input patterns. In fact, our brain defines objects (features, categories, concepts, etc.) as groups of correlating ‘suspicious coincidences’. Moreover, recognition of (and reduced response to) similar input patterns helps to highlight unexpected signals, which is essential for survival. If a layer of Hebbian learning units becomes connected by modifiable anti-Hebbian feed-backs, the resulting system is able to learn this discrimination and to recognize other principal components of an incoming, complex signal than only its first principal component [93, 94]. A well-known biological example is that of the mormyrid electric fish, which is able to eliminate predictable inputs produced by its own, regular motor output. However, this response is a general feature of cerebellum-like, laminar structures, where anti-Hebbian outputs of a deeper layer modulate outer layers (Fig. 3A) [95]. Thus using anti-Hebbian learning prevents excessive noise (i.e. regular, correlating, expected input) from masking important (i.e. unexpected) sensory information. Most sensory systems work based on the principle of fold-change detection, which allows for a proportional response to the fold-change of a signal (the unexpected) relative to the background (the repeated, regular, expected) [96]. From the complexity of learning responses of non-neuronal single cells [2] and the presence of distributed decision making in cellular signalling [97], we may expect that the widespread occurrence of anti-Hebbian learning in signalling networks (see examples above) is involved in the discrimination between repeated and unexpected signals in single, non-neuronal cells, too. Horizontal activation at a receptor-proximal level, as well as mutual inhibition at a receptor-distant level in signalling networks also point toward this expectation. For instance, in the TNF-induced NF-κB signalling, the well-studied upstream crosstalk conveyed by TNFR-associated factors (TRAFs) acts as horizontal activation at the receptor-proximal level [98]. While downstream, a network motif containing inhibition has been described that can impart fold-change detection to cell signalling circuits: the incoherent type-1 feed-forward loop (I1-FFL) (Fig. 3B) [96, 99, 100]. I1-FFL is one of the most frequently occurring network motifs in transcriptional networks [101]. Besides fold-change detection, I1-FFLs have a role in response acceleration even in yeast [102]. In an I1-FFL, X upregulates Y, while it also upregulates Z, a repressor of Y. This indirect repression of Y, coupled with the direct activation of Y, can be considered an anti-Hebbian learning mechanism. Besides NF-κB signalling, I1-FFLs were also suggested to enable fold-change detection in the nuclear levels of the transcription factors of transforming growth factor beta (TGF-β) signalling, explaining how the cells are able to give the same proportional response, even though the nuclear level of transcription factors can vary greatly from cell to cell [96]. The occurrence of I1-FFLs in major signalling pathways suggests that this learning mechanism may be a rather general feature of signalling networks. Even still, to decide whether discrimination between repeated and unexpected signals is also a property of single cells, future experiments are required.

Fig. 3
figure 3

Hebbian and anti-Hebbian learning layers in neuronal and signalling networks. A Schematic representation of the combination of Hebbian- and anti-Hebbian learning layers, which result in the discrimination between predictable and unexpected inputs. The cerebellum-like laminar structure of the figure is widespread in various animal and human neuronal networks [95] and was also shown to work in computational neural networks [93, 94]. Note that self-inhibitory connections (‘autapses’) are not necessarily needed for the circuit. B Proposed Hebbian and anti-Hebbian learning layers in the NF-κB signalling. Receptor proximally, the signalling of multiple receptors can lead to the activation of the NF-κB pathway through TRAFs. Concurrently, some can lead to the activation of the adjacent AP-1 signalling pathway [98]. This constitutes a horizontal activation in the proposed Hebbian learning layer in the upstream signalling. In the downstream signalling, AP-1 and the non-canonical NF-κB pathway modulate NF-κB target genes (e.g. interleukin-8, interleukin-6). The canonical NF-κB (RelA-p52 heterodimer) has also been shown [100] to upregulate the formation of the transcriptionally inactive p50-p50/p52-p52 homodimers that act as competitors to NF-κB for κB sites in the target genes’ promoters. These interactions, highlighted in red, constitute a Type-1 incoherent feed-forward loop (I1-FFL, see inset) that can be understood as an anti-Hebbian learning mechanism. This system enables fold-change detection of the incoming signal in NF-κB nuclear levels, that is analogous to discrimination between predictable and unexpected inputs. This figure was created with yEd Graph Editor

Full size image

Applications of cellular learning and ‘forgetting’ in pharmacology and drug design

Mimicking cellular learning (memory) became a recent hit in drug design. Chemically-induced proximity between two adjacent signalling proteins (a new drug design paradigm [103, 104]) is actually copying Hebbian learning of the cell [10]. In this ‘cellular learning scenario’ chemical proximity-develo** drugs induce targeted posttranslational modifications of key, otherwise undruggable proteins. In the reversed, anti-Hebbian learning model, chemically-induced proximity promotes the selective degradation of the target [105, 106].

Drug resistance can be conceptualised in a learning network model as habituation. Biological networks may contain nodes, where stimulation breaks the habituation (drug resistance) developed by the network [52]. Limited drug tolerance can be conceptualised as sensitisation in a learning network model, most simply by displaying a hysteresis-type response. Interestingly, breaking of sensitisation was much rarer phenomenon in a model of 35 biological networks than that of habituation [52]. However, the break of allergy-induced sensitisation against drugs became a carefully manageable clinical modality in the last decades — as we will describe in the following paragraphs.

Signal desensitisation plays a major role in anti-cancer therapy, which can be regarded as ‘the archetype’ of therapeutic intervention consequences in a number of other diseases. As an example for the first modality of protocols, the anti-cancer agent, 90 kDa heat shock protein (Hsp90) inhibitors induce a desensitisation of the EGF receptor via p38 MAPK-mediated phosphorylation at Ser1046/1047 of the EGF receptor in human pancreatic cancer cells. Here drug-induced desensitisation of the cancer-promoting growth factor signal is a mode of action to avoid disease [105]. As an example for the second modality of consequences, gastric cancer cells become desensitised to trastuzumab-treatment by upregulation of MUC4 expression and by catecholamine-induced β2-adrenoreceptor activation. Here desensitisation, i.e. the development of drug resistance is an unwanted consequence of drug treatment [106]. As a third modality of therapeutic interventions, response-desensitisation (i.e. breaking the sensitisation of unwanted side-reactions for the drug) is a general goal in cancer therapy, where patients often develop sensitivity towards the administered drugs [107, 108].

Supporting the notion that cancer therapy experiences are ‘pars pro toto’ for other conditions, increased hypersensitivity to drugs (e.g. for aspirin and non-steroid anti-inflammatory drugs, NSAIDs in patients with heart disease or inflammatory diseases; for insulins, penicillin or other antibiotics) became a general phenomenon in the past 25 years due to the widespread and intensive drug use. Therefore, carefully administered drug-desensitisation protocols became more and more important in the clinical practice [109, 110]. Similarly, drug-induced desensitisation of cellular mechanisms of action is also a general phenomenon in a number of non-cancer treatment protocols, including that of β-adrenergic agonists in asthma [111], diabetes [112, 113] or heart failure [114].

Conclusion

In our previous work [10] we gave several examples for cellular learning [35,36,37,38], as well as network responses other than those of signalling networks, such as metabolic reaction networks and metabolons [61,62,63,64,65,66,67,68,69,70]. This may be displayed by cross-desensitisation, where prolonged exposure for a pathway agonist induces the desensitisation of other pathway(s) [72, 73]. A specific condition of generally increased signal intensity and/or complexity is stress, which desensitises a wide range of signals [74,75,91].

Here we propose that cellular learning, desensitisation, stress and ageing may be placed as responses along the same axis of more and more intensive (more and more prolonged, or more and more often repeated) signals (Fig. 2). We pose the question, whether single cells may also display discrimination between repeated and unexpected signals, a common property of neuronal and artificial neural networks (Fig. 3A) [93,94,95]. As a first step in answering this question, we present fold-change detection enabling I1-FFLs as anti-Hebbian learning mechanisms that are potentially general features of signalling networks given their occurrence in prominent signalling pathways like NF-κB (Fig. 3B) and TGF-β signalling [96, 98,99,100,101].

Finally, we summarize applications of signalling network learning and desensitisation in clinical treatments discriminating between five scenarios (Fig. 4): 1.) when cellular Hebbian learning is mimicked by chemically-induced proximity between signalling network components [103, 104]; 2.) when cellular anti-Hebbian learning is mimicked by chemically-induced proximity of protein degradation [104, 105]; 3.) when desensitisation of unwanted signalling (such as that in cancer) is the mechanism of drug action [105]; 4.) when desensitisation of wanted signalling occurs, and should be avoided (in cancer, asthma, diabetes or heart failure [106, 110,111,112,113,114]); and finally, 5.) when sensitisation against a drug occurs by allergic reaction, which also should be minimized (in cancer, inflammatory diseases, diabetes or infections [107, 109, 110]).

Fig. 4
figure 4

Clinical treatments for and against Hebbian and anti-Hebbian learning of the signalling network. a Cellular Hebbian learning is mimicked by chemically-induced proximity between signalling network components. b Anti-Hebbian learning is mimicked by chemically-induced proximity of protein degradation. c Desensitisation (drug-induced anti-Hebbian learning) of unwanted signalling (such as that in cancer). d Prevention of desensitisation of wanted signalling (cellular anti-Hebbian learning) in cancer, asthma, diabetes or heart failure. e Prevention of allergy-induced sensitisation (Hebbian learning) against the drug in cancer, inflammatory diseases, diabetes or infections. This figure was created with BioRender.com

Full size image

We hope that our summary will prompt further investigations of the phenomena, when cells learn (develop cellular memory) by Hebbian learning-type processes, and when they ‘refuse’ to learn more, i.e. become desensitised (display anti-Hebbian learning, i.e. cellular ‘forgetting’) by prolonged exposure to environmental signals, by stress or by ageing. It is an interesting question, how much desensitisation remains specific for the given pathway, and how much it is displayed as cross-desensitisation of other pathways, or as a general forgetting (desensitisation) of many (if not all) pathways. While agonist-induced desensitisation is mostly the former, directed type desensitisation against the same pathway (or selected different pathways), stress- and ageing-induced desensitisation are usually more widespread phenomena involving a larger segment of the signalling network. We predict that network methodologies will greatly help the discrimination between these scenarios.