Towards the Idea of Molecular Brains
The rhodopsins consisting of seven transmembrane α-helices belong to the vast family of G protein-coupled receptors [324,368]. The light-induced isomerisation of the 11-cis-retinal is coupled to a cascade of events that relay the signal to guanylate kinases [369,370] or phosphodiesterase [371] that trigger various intracellular responses. The PYP adopt the fold of the Per-ARNT-Sim (PAS) domain [372,373] that senses a vast range of stimuli, from photon to ligand [374]. The trans to cis photoisomerisation of its p-coumaric acid chromophore induces small structural rearrangements of the protein [375,376] and is associated with charge delocalisation around the chromophore [377]. The ubiquitous LOV (light, oxygen or voltage) photosensors domains also adopt the PAS domain structure, but they use the FMN as a chromophore [356,378], and they are associated with various effectors such as histidine kinases, protein involved in synthesis of cyclic cGMP, STAS (anti sigma antagonist), helix-turn-helix. They have also been found to be associated with other sensor domains [379]. Several mechanisms of signalling have been proposed on the basis of their dark and light-activated structures, such as helix Jα unwinding [380,381] or dimerisation [382]. BLUF, bacterial and eukaryotic blue light receptor using flavine adenine dinucleotide (FAD), has a modular architecture comprising a 150aas receptor domain with a ferredoxin-like fold that can be connected with many different effectors [383,384,385,386,387]. They can be found as the single photosensory domain involved in light-regulated protein interactions [388] or in multidomain proteins fused to light-regulated transcriptional effector [389,390], phosphodiesterases [383] or adenylyl cyclases [391,392]. BLUF domains bind FAD/FMN/RF pigment non-covalently for the properties of the isoalloxazine to absorb blue light.
In BLUF, a photo-induced proton-coupled electron transfer (PCET) initiate a rearrangement of hydrogen bonds around the flavin cofactor after illumination, which is transmitted to the surface of the receptor and lead to the activation of the various effectors. Cryptochromes adopt a Rossman-like fold similar to photolyases [393,394] involved in functions ranging from DNA repair to blue light regulation of growth, development and circadian rhythms [395]. Cryptochrome photoreception is based on blue light-induced interconversion between several redox states of flavin as chromophore [396]. While the role of the tryptophan triad is controversial [397], the light-induced oligomerisation of plant cryptochrome is an accepted mechanism for light-regulation interactions with various signalling partners [398,399,400]. Phytochromes are modular multidomain red/far-red photosensory proteins that share conserved PAS, GAF and PHY core photosensory domain that adopt a knotted structure [401] and bind covalently a linear tetrapyrrole chromophore (bilin, phytochromobilin, or biliverdin) in a broad range organisms such as bacteria, unicellular eukaryotes and plants [402,403,404,405]. Photoconversion triggers structural changes in the dimer interactions and the refolding of a “tongue” loop that modulate the activity of C-terminal “output” domains [406,407].
Magnetotactic bacteria have, for example, evolved magnetosomes to perceive and orient along the Earth’s magnetic field [408] and lead to complex collective phenomena of some bacteria subjected to a magnetic field [409,410]. These organelles consist of magnetite (Fe3O4) wrapped in cell membranes organised along the motor axis by cytoskeleton proteins [411,412]. These structures are encoded by conserved gene clusters that have diverged in a large number of magnetotactic bacteria [413]. However, the process may be more complex than expected as a link has been established between the magnetotactic properties of some bacteria and aerotaxis-dependent signal transduction systems [414]. Proteins involved in the magnetotactic response whose genes are in close proximity to genes involved in signal transduction are beginning to be characterised structurally [415]. In animals and among them butterflies, other systems appear to be involved in magnetic field perception, such as radical pairs and cryptochromes [8,9,416,417].
Another important signalling complex has been discovered in some bacteria: the stressosome [418]. This complex is formed by the pseudo-icosahedral assembly of RsbR (sensor), RsbS (scaffold) and RsbT (kinase) proteins [419,420]. This complex responds to environmental changes that could be sources of stress such as ethanol, UV or osmolarity by triggering cascade events leading to the expression (release) of alternative sigma factors (sigma B), which, in turn, activates over 150 stress-related genes. The molecular details of this phosphorylation cascade are not yet well understood.
Beyond protein receptors, single-cell and pluricellular organisms have developed a great variety of sophisticated systems to monitor and respond to temperature by different molecular strategies [421,422,423]. For example, with their nucleic acid thermosensors, also called “RNA thermometers”, bacteria register temperature changes by using temperature-modulated structures in the untranslated region of some mRNAs. Temperature sensing is based on several non-canonical, heat-labile base pairs temperatures [424,425,426,427,428]. Conversely, cold sensor group II introns are self-splicing ribozymes where cold-induced disruption of key tertiary interactions prevents splicing and triggers various cellular responses [429]. In addition, RNA riboswitches may also sense a variety of metabolites [430,431,432,433,434,435,436] and a great variety of protein chemosensors involved in chemotaxis.
However, molecular sensing can also be extended to other molecular mechanisms and functions as diverse as the control of correct pairing during replication by DNA polymerases or the detection of node chirality by type II topoisomerases [437,438,439]. Cells can also sense and control the DNA integrity in monitoring the long-distance migration of charges through the aromatic base-pair stack within the DNA helix [440,441]. In addition, there is growing evidence that many reactive oxidative and electrophilic species (ROS/RES) act as cellular signals and can mediate inter-organelle or intercellular redox information exchange [442,443].
Despite their functional analogies, cellular signalling and nervous systems have long remained two very distinct disciplines. Even today, few bridges are established between these two themes. The first daring analogies between cell signalling and neurobiology emerged with bacterial chemotaxis [51,52]. Still, these articles have remained surprisingly little cited. Bacterial chemotaxis was thus the first step, in the 20th century, of a fundamental question: how can a bacterium orient itself, memorise and become part of history, learn and even decide where it will go according to chemical stimuli that repel or attract it in its environment? We began to realise that bacteria could make decisions in complex situations [444,445], amplify signals, show habituation faculties and retain the memory of past situations [446,447]. Chemotaxis allows the bacterium to choose its orientation according to a source of attractive or repulsive substances [445]. It is integrated into a broader context of perception of external or internal signals based on a principle of signal transduction: two-component systems [448]. In these sophisticated 2-component systems, a “sensor” protein auto-phosphorylates in response to a specific signal coming from specific receptors and transfers the phosphoryl group to a “response regulator” protein that carries out a cellular response [114,449,450,451,452,453]. The constant comparison of the signals perceived at a past and a present moment decides the direction of rotation of the flagellum [454,455]. When the flagellum rotates clockwise, the bacterium oscillates in a kind of random walk. When the flagellum rotates counter-clockwise, the bacterium moves in one direction only [456]. This alternation between “run and tumble” thus allows it to orient itself along a chemical gradient.
The last decades have provided detailed information on the molecular mechanisms responsible for this sophisticated behaviour. They have identified most of the bacterial chemotactic receptors, sensor and response regulator proteins associated with signal transduction. The structure of individual components and clusters of receptors grouped in networks has also been resolved [457,458,459]. These clusters formed by the assembly of various types of receptors have the property of amplifying and integrating all signals. Thus, their cooperation in response to a given chemical signal makes it possible to generate an appropriate response, even in contradictory situations [444]. Their sensitivity is modulated according to what they have previously perceived through the methylation of a protein region (HAMP), which thus keeps the memory of what the bacterium has encountered. These networks of receptors and sensor and regulator responses were, therefore, the first molecular assemblies to have been compared to proto-brains [48,49]. They confer to each bacterium an individual behaviour that varies according to its own history and the different situations it may have gone through [147]. Interesting data have also been produced by genomics and structural studies, which have shown that the organisation of these clusters is universal in bacteria [460]. Although the diversity of receptors for various substances may vary from one species to another, the basic organisation and mechanisms involving two-component systems are common in all chemotactic bacteria [104,461,462]. Thus the networks of the component of bacterial chemotaxis with emergent properties, allowing bacteria to develop complex behaviours, were the first to be compared to molecular brains, which are moreover frequently found at the “head” of bacteria.
As seen above, in his 1995 seminal paper, Dennis Bray proposed one of the most exciting hypotheses in biology: circuits formed by protein networks may play a role analogous to the nervous systems in cells [13]. Although the emergence of these “proto-brains” probably coincided with the beginning of life, it took another two decades before they were explicitly named [48,49].
Another analogy with the nervous circuit has been proposed for the r-protein networks observed in the ribosome [44,45]. Interestingly, although both ribosomes and chemoreceptor arrays have been compared to brains, they differ significantly in their symmetry and geometric properties: while chemotactic receptor arrays form regular and symmetrical structures, r-proteins display long unstructured regions and form totally irregular networks (Figure 3). Yet, although asymmetric, the ribosome is considered as a window towards the earliest forms of life that predate the three kingdoms and therefore, the analogy of the ribosome with brains can provide interesting information on both the functional and evolutionary levels.
While in astrophysics looking far away gives the opportunity to glimpse the fossil radiation of the universe, looking into the heart of the ribosome may tell us what the first forms of life might have looked like. The ribosome evolved by accretion around a core that predates the radiation of the three kingdoms and were probably present in LUCA (Figure 2) [463,464,465,466,467]. The ribosomes are thus considered as a relic of ancient translation systems that co-evolved with the genetic code have evolved by the accretion of rRNA and ribosomal (r)-proteins around a universal core [466,468,469,470,471,472]. They then followed distinct evolutionary pathways to form the bacterial, archaeal and eukaryotic ribosomes whose overall structures are well conserved within kingdoms [55,473,474,475]. The complexity of ribosome assemblies, structures, efficiencies and translation fidelity concomitantly increased in the course of the evolution.
Strikingly, with their long filamentous extensions, ribosomal proteins (r-proteins) form neuron-like networks that “innervate” the ribosomal functional centres such as tRNA sites, PTC, and the peptide tunnel [44,45]. In these networks, r-proteins are interconnected by the formation of molecular “synapses”, very small and phylogenetically conserved interfaces between the filamentous extensions of the ribosomal proteins that thread their way through the interstices of the rRNA [45]. These tiny structures are stabilised by their interactions with the rRNA and display a “necessary minimum”, conserved aromatic-basic amino acid motifs that are also shared by larger interfaces. It has been proposed that these highly conserved interfaces have been selected during evolution to play a specific role in inter-protein communication, and they reveal the strictly necessary interacting residues to ensure information transfer from a protein to another (Figure 4). An interesting hypothesis is that these minimalist “molecular synapses” reveal much more general principles in molecular communication. Indeed, these tiny interfaces, which appear in their simplest expression in the ribosome thanks to the spatial constraints of ribosomal RNA (rRNA), could be ubiquitous in macromolecular complexes but drowned out by a “structural” background involving other amino acids for their stabilisation.
The r-protein networks could contribute to both the ribosomal assembly and the “sensorimotor control” of protein synthesis. Many experimental studies have indeed shown indeed that ribosome functional sites continually exchange and integrate information during the various steps of translation. As the numerous studies of the Dinman group have shown: “an extensive network of information flow through the ribosome” during protein biosynthesis [476,477,478,479,480,481,482,483]. For example, several studies have also demonstrated long-range signalling between the decoding centre that monitors the correct geometry of the codon-anticodon and other distant sites such as the sarcin ricin loop (SRL) or the E-tRNA site [473,484]. R-proteins of the ribosomal tunnel also play an active role in the regulation of protein synthesis and co-translational folding [485,486]. Ribosomes also perceive each other through quality sensors of collided ribosomes in eukaryotes [487]. In addition, the ribosomes synchronise many complex movements during the translation cycles [488,489,490]. The recent discoveries of “ribosome heterogeneity” [491] also significantly expand the complexity of the possible ribosome’s network topologies [492] and open new perspectives on “network plasticity” that could also play a role in its behavioural richness.
A recent interdisciplinary study has shown how r-protein networks have evolved toward a growing complexity through the coevolution of the r-protein extensions and the increasing number of connexions [46]. This revealed that network expansion is produced by the collective (co)-evolution of r-proteins leading to an asymmetrical evolution of the two subunits. Furthermore, graph theory showed that the network evolution did not occur at random: each new occurring extension and connection gradually relates functional modules and places the functional centres in central positions of the network. The strong selective pressure that is also expressed at the amino acid acquisition links the network architectures and the r-protein phylogeny, thus suggesting that the networks have gradually evolved to sophisticated allosteric pathways. The congruence between independent evolutionary traits indicates that the network architectures evolved to relate and optimise the information spread between functional modules (Figure 1). This network archaeology study has also revealed the existence of a universal network that consists of 49 strictly conserved connections that were probably present before the radiation of the bacteria and archaea [493] (Figure 5).
Interestingly, this primordial network is much more developed in the small ribosomal subunit suggesting that the large subunit network complexity developed in later evolutionary stages. These findings, therefore, suggest that LUCA already possessed such type of molecular networks, with long wires and tiny interfaces. Interestingly, these networks also mix the π-systems of rRNA and aromatic amino acids of proteins for forming conserved structural motifs probably involved in signal transduction. It is, therefore, possible that this ancestral mode of communication has then not only evolved in modern ribosomes but in other macromolecular systems for information transfer and processing. The ribosome opens a window on the first information processing networks, which appeared at the origin of life. They probably diverged towards other cell systems that have been compared to brains, such as the multiple nano-brains described in Baluska’s article published in this special issue [40].
Just as the observation of similar behaviours of cells, plants and organisms with nervous systems, it does not seem unreasonable to imagine that the signalling mechanisms that underlie them share some common points. In this regard and because of their mechanistic and functional analogies with nervous systems, plant roots [53,54], bacterial chemotaxis [13,48,51,52] and ribosomal signalling networks [44,45] have been compared to brains. This analogy assumes both a spatial and temporal scale invariance between molecular and neuron networks: while ribosomal signalling networks, which are no larger than a few tens of nanometres, can tell us about the first life informational systems at the origin of life, nervous systems, of the order of a metre in size, corresponding to a major evolutionary transition that occurred 600 million years ago, the appearance of metazoans [59,115,494,495]. The comparison of these systems and their behaviours reveal similarities and differences that we believe can provide interesting insights into the invariants of living information systems and the way they have evolved during major evolutionary transitions (Figure 6).
The perception, transfer and integration of signals by the ribosomes or the cells ensure autonomous behaviour and appropriate decision making in the face of their own fluctuations or those of their environment. However, if for cells the “outside” consists of other cells, the intercellular space or their ecosystem environment, the ribosomes listen to each other, organelles, and they monitor what may be happening in the cytoplasm. Whether anchored in ribosomal rRNA or membrane phospholipids, the sensor proteins of both systems have developed close relationships with their matrix, which is actively involved in the signalling processes through various mechanisms. While rRNA allostery also participates in information transfer within the ribosome [479], phospholipids modulate the responses of the receptors and may also constitute signalling molecules [285].
However, the main difference between ribosomal and cellular signalling lies in the duration of the interaction between the nodes. Once the ribosome is assembled, interactions between ribosomal proteins and functional centres are mostly “permanent” in that most of the r-protein remain connected (Figure 3 and Figure 4). However, the r-protein nodes can also interact transiently with tRNAs, mRNAs, various translation factors and proteins outside the ribosome. In contrast, the cell signalling networks are essentially transient and involve contacts between proteins that diffuse into the cytoplasm. Note, however, that proteins anchored in membrane systems can also establish long-lasting contacts with their partners in signalling clusters or scaffolds (as, for example, the chemoreceptor and CheA and CheW) (Figure 3). However, cells also display systems analogous to ribosomal extensions involved in signal transmission: the filaments of the cytoskeleton [496]. Thus, the ribosome and the cell signalling networks correspond, respectively, to the solid and liquid brains concepts described by R. Solé [62] (Figure 6).
The structure and duration of the interactions within the network constraints and have a direct implication on the nature of the signals that are transmitted between partners. While the ribosomal “permanent” networks can transmit diverse transient or ephemeral signals that can be grouped into allosteric phenomena, transient networks have had to develop chemical messengers that are durable over time. The long-lasting but reversible post-translational modifications of proteins and the synthesis of second messengers that trigger downstream signalling cascades are signalling solutions to the “impermanence” of networks. An evolutionary consequence of the network transient interactions is a great diversity in the protein modification and signalling codes. Conversely, the evolution of the nervous systems has convergently reinstated networks based on durable interactions between neurons through the establishment of “true” synapses. In consequence, simar to ribosome networks, a unification of signal transmission modes seems to have re-emerged later in evolution: the directional propagation of “ephemeral signals”, the action potentials induced by the discharge of neurotransmitters in the inter-synaptic space. The analogy between the ribosomal network and the nervous system can also be continued with regard to the organisation of functional classes of nodes. For example, the number of nodes in the eukaryotic r-protein network is of the same order of magnitude as the C. elegans neural network, which contains several hundred neurons [497]. These two networks are organised into modules dedicated to distinct functions in processing information (Figure 7).
Further comparisons between the architectures and modular organisation of networks containing an equivalent number of nodes at different size scales could provide valuable information on the general principles of information processing. However, beyond the analogies concerning the connectivity of networks, data on their dynamics and the way in which signals are exchanged are indispensable.
Graph theory and computational approaches have provided various models for understanding how signalling networks are organised and identified a number of emergent properties such as noise reduction, signal amplification [41,42,498,499,500] or even associative learning [501,502]. In the framework of the Bayesian brain, adaptative systems evolve in such a way that they can predict the behaviour of their environment so that they can preserve their integrity. These “brains” make a hypothesis on the evolution of their environment in order to make sense of it and act accordingly to these hypotheses in a way that maximises their benefits. These systems update, in a Bayesian fashion, the hypothesis they make on the world with respect to new observations, which, in turn, changes what they expect the future outcomes of the environment would be. For example, Lee and Mumford proposed a Bayesian interpretation, using hierachical Bayesian models, of how the visual cortex integrates information accounting for neurophysiological evidence previous paradigm could not account for [503].
A model of the Bayesian brain hypothesis is the active inference framework [504,505], where a “brain” receives information from gateways such as sensory organs and transfers it as information onto its internal model of the world in an optimal fashion. It then uses this information to act in adequacy with the hypothesis is made on how the external world evolves accordingly to its observations.
More precisely, observations of the world, received, for example by sensory organs, are seen as being generated by an underlying unobserved process that the brain can only model partially. In this framework, the “brain” specifies internally collections of variables that stand for the causes on which depend the observations, but as this modelisation is imperfect, the inputs depend on the causes in a stochastic manner. This “brain” uses its observations to choose from different hypotheses on the causes that would have generated this input with respect to its own modelisation of the generative process; this is the inference step, a hypothesis is a probability distribution over the causes, and the “brain” proceeds through this step by choosing the best approximation on the posterior on the causes after new input data are observed. The “brain” then makes use of this hypothesis to act, for example, in such a way as to maximise a reward. There are several propositions for how Bayesian computations could be implemented biologically, for example, in the cortex [504,505,506] in a manner consistent with experimental data on connectivity (role of forward and backward connections) and temporality of cortical response [505,507].
We want to stress that what is important is how biological signalling interact with one another if one wants to understand what “computations” underly such process: biologically relevant quantities act as a dictionary complex enough for encoding important variables over which to take decisions and interactions between such quantities have enough structure to capture the interactions between these variables. For example, in [508], classical equations describing the dynamics of rod phototransduction can be interpreted as a belief update algorithm on the presence or absence of a photon reaching the rod. This point of view motivates the relevance of the use of a Bayesian brain framework to other intelligent systems capable of decision making but lacking a nervous system [509]. More generally, cell signalling can, remarkably, be interpreted in a Bayesian fashion, where macromolecules and messengers are the physical support for computing Bayesian a posterioris [74]. For example, it has been observed that the unicellular Physarum polycephalum’s choices when facing two environments with different amounts of reward is dependent in a probabilistic manner on the proportion of reward in each environment; however, information on how signalling occurs is yet to be understood [510].
From a mechanistic point of view, the central question is, how do molecular or brain networks integrate multiple signals and take appropriate decisions? For example, in ribosomal networks, many r-proteins display multiple connections where some hubs such as the eukaryotic uL4 or uS8 can be connected to more than eight partners [45,46] (Figure 1). What are the molecular mechanisms underlying the integration and decision making from these multiple stimuli? The secret is in the modular organisation and the combinatory interaction of modules at different scale levels.
At the cellular level, the combinatorial binding of multiple receptors and adaptators allows signal integration in space and time. For example, RTKs can recognise specific targets and integrate multiple stimuli through regulated multidomain interactions. Thus, multiple signal inputs can be integrated through a combination of recognition modules and permit “coincidence” detection [347,349,511]. At the molecular level, the detection and integration of multiple signals rely on sophisticated allosteric mechanisms observed in particular proteins. Similarly, we will see that it is the combinatorial association of smaller “allosteric modules” within protein domains that make possible the integration of multiple signals.
However, allostery, which, according to Jacques Monod, constitutes “the second secret of life” [512], is an evolving concept. A classic view is that allostery consists of a remote “communication” between two sites of a macromolecule, an “allosteric” or regulatory site and a “functional” or catalytic site. Thus a ligand that binds at a distance from an enzyme active site may modulate (decreasing or increasing) its activity. However, the conceptualisation of “how signals are propagated” along a macromolecule to regulate distant sites has evolved considerably. Allostery has transited from the simple concept where conformational changes were propagated in a “domino-like” motion to a more probabilistic view based on the multiple dynamic states of proteins and “conformation ensembles” [286,513]. According to Nussivov and colleagues, a unified view bringing together a thermodynamic and structural approach now provides a better understanding of how allostery works [121]. Today, we more readily speak of structural coupling between functional and allosteric sites and of “allostetric propagation pathway”, which establishes can form “channels” or “networks” within or between macromolecules (RNA or proteins). Coupling the two distinct sites would mean creating a concerted or correlated movement between them in the dynamics of the protein through a network of interaction between the residues between these two sites. There is now a large repertoire of mechanisms that macromolecules can use to transmit intra- or inter-molecular signals, and in recent decades, the molecular disorder has also contributed to documenting allosteric mechanisms. It is thought that propagations of order-disorder transitions can contribute to signal transmission [514].
In addition, intrinsically unstructured proteins are thought to play a particular role in signal amplification due to the low energy barriers between conformers [515,516,517,518].
Allostery is itself subject to evolutionary processes throughout the history of life. In the same way as sensitivity to physico-chemical stimuli and due to their intrinsic dynamics, allostery is a ubiquitous property of biological macromolecules [519,520,521,522]. Evolution has been able to modulate allosteric properties along particular pathways by optimising the sequences of macromolecules [523]. On the other hand, allosteric regulation and catalysis have emerged from common pathways since they share evolutionary optimisation of the same conformational mobility in protein sequence [524,525]. Evolution has thus selected robust sequences capable of compensating for the deleterious effects of mutations in allosteric pathways [526]. It is now widely accepted that macromolecules have optimised their sequences during evolution to inhabit, represent (“populate”) several states that can switch from one to another (switchable states), e.g., one disordered or unfolded and one ordered.
However, macromolecules also display other signal transmission mechanisms involving their electrostatic properties. For example, the coupling of structural changes between distant domains can be electrostatically driven and rely on the reorganisation of electrostatic charges and potentials [527,528]. For example, the charge reorganisation mediates the communication between the two domains of the bilobed r-protein bL20 [527] and calmodulin [529]. These mechanisms have subsequently been observed in other systems [530] and are referred to as dielectric allostery [531,532]. Dipole coupling and electrodynamic processes may also be involved in signal propagation between distant sites in proteins [533,534].