On Having No Head: Cognition throughout Biological Systems

The central nervous system (CNS) underlies memory, perception, decision-making, and behavior in numerous organisms. However, neural networks have no monopoly on the signaling functions that implement these remarkable algorithms. It is often forgotten that neurons optimized cellular signaling modes that existed long before the CNS appeared during evolution, and were used by somatic cellular networks to orchestrate physiology, embryonic development, and behavior. Many of the key dynamics that enable information processing can, in fact, be implemented by different biological hardware. This is widely exploited by organisms throughout the tree of life. Here, we review data on memory, learning, and other aspects of cognition in a range of models, including single celled organisms, plants, and tissues in animal bodies. We discuss current knowledge of the molecular mechanisms at work in these systems, and suggest several hypotheses for future investigation. The study of cognitive processes implemented in aneural contexts is a fascinating, highly interdisciplinary topic that has many implications for evolution, cell biology, regenerative medicine, computer science, and synthetic bioengineering.

Survival in a complex, dynamic, and highly competitive environment requires biological systems to make numerous decisions with respect to possible activities (Conrad, 1996; Holcombe and Paton, 1998). Evolutionary pressure to optimize decision-making has led to the inevitable exploitation of past history (memory) and information processing (computation). Importantly however, decisions are made at every level of biological organization. For example, multicellular organisms, such as animals and higher plants, exhibit multilayer complex goal-directed behaviors also at their cellular and subcellular levels. Underlying physiological systems must maintain homeostasis and predict future conditions (Freddolino and Tavazoie, 2012) in the face of unpredictable changes in environmental conditions, while cells must coordinate their activity in an exquisite 3-dimensional ballet of embryogenesis and complex organ regeneration. At the extremes of the scale of organization, dynamic self-organizing subcellular components like cytoskeleton and molecular networks (Albrecht-Buehler, 1985; Craddock et al., 2012; for plant cells see Volkmann and Baluška, 1999; Barlow and Baluška, 2000) and colonies of organisms (Shapiro, 1998; Couzin, 2009) perform similar functions in their own contexts. Here, “cognition” refers to the total set of mechanisms and processes that underlie information acquisition, storage, processing, and use, at any level of organization (Lyon, 2015).

Memory is an essential component of these processes, at all levels. For our purposes, memory can be defined as experience-dependent modification of internal structure, in a stimulus-specific manner that alters the way the system will respond to stimuli in the future as a function of its past. This requires a labile yet stable medium, to provide the necessary latency. The process may or may not involve a degree of intelligence, in the sense of the ability to compress prior stimuli into informationally-compact representations (inference). In essence, sensory memory is a message to one’s future self – a view reminds us that memory is thus another instance of biological communication (which, as exchange of signals, is ubiquitous among all levels of biological organization). Put this way, we can see that in principle many biological mechanisms can be exploited for this purpose. The updates in the configuration (or state) of a system, as occurs also during sensory memory formation in all organisms, is formally known as “computation” in computer science.

These concepts are quite general. However, outside of the unconventional cognition community (Calvo and Baluška, 2015) or biological computation community (Adamatzky et al., 2008), it is widely assumed that memory is the exclusive province of brains, or even complex animals. Older work exploring these issues in plants (reviewed by Gremiaux et al., 2014), non-neural somatic tissues (Mackie, 1970), and even inorganic media (Bose, 1926), have been largely forgotten in favor of the remarkable advances in recent cognitive neurosciences with their focus on the brain. Nevertheless, plant cells are known to be able to use action potentials to control their movements and behavior since times when Charles Darwin and Jagadis Chandra Bose turned their interest toward plants (Darwin, 1880; Shepherd, 2005; Baluška et al., 2009a). Currently, surprisingly, higher plants are emerging as behaviorally active organisms, enjoying bio-communication and showing plant-specific cognition and intelligence (Trewavas, 2005, 2014; Karban et al., 2014a,b; Calvo and Baluška, 2015; Calvo, 2016).

Here, we survey a wide-ranging literature on memory and sensory systems-based cognition in organisms (biological systems) lacking animal/human-type brains. Our goal is to acquaint readers interested in cognition with numerous aneural model systems in which this subject can be pursued, and to draw the attention of bench biologists working on those systems to cognitive, information-focused perspectives on the mechanisms they are studying. Importantly, in discussing cognitive performance in the various systems, we do not mean the full-blown human-like cognitive performance, or human-type of self-awareness and consciousness. We are avoiding issues of the ‘Hard’ problem of cognitive science, and do not claim anything like higher-order symbolic representations. Our definition is purely functional and minimalist (Calvo and Baluška, 2015), drawing attention to the similarities in computational tasks performed by diverse biological systems, at all levels of complexity, other than animal and human brains. Figure 1 illustrates the full spectrum of cognitive levels and capabilities upon which the various systems we discuss can be placed (Rosenblueth et al., 1943). Our review begins with a consideration of the familiar substrate of cognition: neural dynamics, and of mechanisms that blur the boundaries between neural and non-neural cell functions. We then proceed through progressively more divergent cognitive systems, considering molecular networks, single cell behaviors, networks of cells in various tissues, and organism-wide information processing during regenerative repair. We conclude with some common threads of cognition across levels of organization, which suggest a unified perspective on these highly diverse systems.

Recent work has begun to encompass cognition in ex vivo systems, with studies that have shown training and learning in cultured minimal neural networks (DeMarse and Dockendorf, 2005; Dranias et al., 2013; Pimashkin et al., 2013). Even in vivo, it is increasingly recognized how much processing happens before signals get to the brain of the central nervous system (CNS); a recent example is the discovery that neurons in the skin perform edge detection (Pruszynski and Johansson, 2014).

Importantly, CNS neurons do not embody cognition due to any magical, unique property. Their computational powers derive from the dynamics of networks of linked elements that propagate and integrate signals, and the ability to alter connectivity among those elements (network topology) based on prior activity. In fact, these basic properties are present in biological systems at many complexity scales (from subcellular protein networks to coupled tissues). Might they too underlie some aspects of cognitive-like information processing? Indeed, neurons did not invent their special tricks – they merely optimized them for speed to drive adaptive behavior. These functions, and the molecular mechanisms that implement them – ion channels, electrical synapses (gap junctions), and neurotransmitter molecules are all ancient (Goldsworthy, 1983; Baluška, 2010; Brunet and Arendt, 2016; Moroz and Kohn, 2016). Neural networks evolved from far older signaling pathways that orchestrated development, physiology, and other cellular functions long before the CNS arrived on the evolutionary scene (Buznikov et al., 1996; Levin et al., 2006; Keijzer et al., 2013). Already simple cells of bacteria enjoy sensory systems feeding into cognitive-behavioral circuits and showing many other neural features (Miller and Koshland, 1977; Koshland, 1980; Lyon, 2015). Electrical long-distance signaling and information exchange via spatially propagating waves of potassium is synchronizing bacterial biofilms (Beagle and Lockless, 2015; Nunes-Alves, 2015; Prindle et al., 2015). Integrated bacteria within the biofilm community appear to act as some kind of ‘microbial brain’. Obviously, the neuronal communication has bacterial origins (Baluška and Mancuso, 2009).

The main principles by which neural networks store and process information – plasticity, excitability, and experience-dependent change (Daoudal and Debanne, 2003) are readily applicable to numerous cell types. Indeed, the computational powers of glia and other non-spiking cells in the brain are increasingly appreciated in their contributions to mammalian cognition and intelligence (Oberheim et al., 2009; Goldman et al., 2015). Astrocyte networks perform computations (Schummers et al., 2008), and models of memory have long been proposed that rely on non-spiking neurons (Aur, 2012), revealing that neural-specific, discrete action potentials are not a pre-requisite for memory dynamics.

At the same time, tissues other than neurons are able to conduct the kind of signaling impulses that are considered the sine qua non of cognition. For example, excitation and impulse propagation have been shown in skin (Roberts and Stirling, 1971; James and Soffe, 2011). The evolution of neurons from excitable precursors has been reviewed elsewhere (Mackie, 1970; Baluška and Mancuso, 2009; Baluška, 2010; Moroz and Kohn, 2016), as have the many similarities between neurons and other cell types (Bharti and Arnheiter, 2005; Yaar and Park, 2012). In this overview, we cast our net even broader, examining examples of cognition outside of the CNS domain of life (Calvo and Baluška, 2015; Lyon, 2015; Calvo, 2016), with or without spiking, in cellular networks of complex metazoans, or within single-cells. We also review some of the mechanisms that underlie this cognition which is inherent to cellular life at all levels of biological complexity, and suggest a few novel experimental directions that may exploit the deep lessons suggested by the ubiquitous nature of aneural cognition.

The interplay of neural and non-neural signaling has been shown in several regenerative systems. Neural inputs are required for amphibian limb regeneration (Singer, 1952; Kumar and Brockes, 2012), although curiously, this is not a hardwired requirement but must be learned: limbs that grew without the presence of a nerve later do not require nerve to regenerate, unlike normal limbs. This phenomenon has been termed “nerve addiction” (Yntema, 1959a,b; Filoni et al., 1995), extending the principle of experience-dependent long term change to limb regeneration. Neural inputs are also needed for maintenance of tissue structure in the rodent tongue (Takeda et al., 1996; Sollars et al., 2002), suppression of tumors in rabbits (Pawlowski and Weddell, 1967; Pawlowski, 1970), and regulation of specific pattern in distal tail regeneration in amphibia (Mondia et al., 2011).

Neural and non-neural information systems must cooperate especially when pattern formation and memory intersect. A unique model system for these studies is planaria, the free-living flatworm (Gentile et al., 2011); this is a unique model system that regenerates every part of its body (Reddien and Sánchez Alvarado, 2004) and also possesses a true centralized brain (Sarnat and Netsky, 1985; Pagán, 2014) and learning capabilities (Wells, 1967; Sheiman and Tiras, 1996; Nicolas et al., 2008). In this model species, the dynamics of behavioral memory can be studied during complete brain regeneration [in the axolotl, this can also be done, albeit with only partial brain regeneration (Pietsch and Schneider, 1969)]. Classical studies (McConnell et al., 1959; Corning, 1966), as well as more recent work performed using automated analysis methods (Shomrat and Levin, 2013), showed that memories in planaria survive decapitation – tail fragments trained on a task regenerate brains and then show evidence of recall of the original information. This requires the body to store learned information and imprint it on the nascent brain after it is rebuilt. The mechanisms of this interaction are completely unknown, but offer an unprecedented opportunity to study transfer between neural and somatic memory systems.

Memory, and often the intermediate processes of computation, requires that “stimuli produce a permanent record written on the irritable substance” (Semon and Simon, 1921). What underlying mechanisms have been implicated in non-neural memory and related processes?

One of the best candidates for mechanisms underlying information processing at the single cell level is the cytoskeleton (Albrecht-Buehler, 1985; Craddock et al., 2010; Sahu et al., 2013), which has all of the necessary properties: it is a large, complex structure that is readily modified by a variety of molecular pathways (writing data), is interpreted by numerous motor proteins and other machinery (reading data), and implements a rich set of discrete transition states that could implement computational operations (Hameroff and Watt, 1982; Lahoz-Beltra et al., 1993; Volkmann and Baluška, 1999; Craddock et al., 2012). The cytoskeleton has long been a favorite locus of information storage and control in single-celled organisms, where it regulates behavior (Eisenstein, 1967; Hamilton, 1975) and serves as a non-genomic repository for permanent somatic changes such as cell surface chirality changes (Nelsen et al., 1989). The dynamic actin cytoskeleton behaves as excitable medium (Khan et al., 2012).

Another medium for information processing is within chemical networks, such as reaction-diffusion (RD) dynamics that underlie pattern formation in embryogenesis (Kondo, 2002; Kondo and Miura, 2010; Raspopovic et al., 2014). Recent work has revealed that RD systems and similar excitable chemical media can be designed so as to execute specific computations, and are being used for the design of minimal cognition controllers (Dale and Husbands, 2010) and other kinds of computation including planning (Adamatzky et al., 2003; Adamatzky et al., 2008; Costello et al., 2009). Remarkably, it was shown long ago (Rosen, 1968) that Rashevsky’s 2-factor systems (a model for neuronal excitation) is formally equivalent to Turing’s RD scheme for self-organizing morphogenesis (Turing, 1952). Grossberg then described extensive parallelism between signal processing in chemical gradients during development and neural memory and visual processing (Grossberg, 1978). RD systems are Turing-complete (Scarle, 2009) and support semantical interpretations (Schumann and Adamatzky, 2009), making them an excellent candidate for complex computations. Recent work used in silico evolution of chemical networks to show that simple, plausible reactions can be found which perform associative learning and Bayesian behavior which includes memory traces (McGregor et al., 2012). These data are especially exciting in that they imply that associative learning can readily evolve in metabolic, gene regulatory, or intracellular signaling networks.

The transcriptional control machinery that guides embryogenesis has also been modeled as cognitive processes. Gene regulatory networks can be modeled as neural networks (Watson et al., 2010), with genes representing nodes and functional links representing inductive or repressive relationships among those genes. That landmark study showed that changes to the connections in the regulatory net represent a kind of Hebbian plasticity (as genes whose expression is up-regulated in specific environments tend to become co-regulated and thus expressed together). In part due to this fire-together-wire-together process, a GRN will develop an associative memory of phenotypes selected in the past. This view sheds important light on the relationship between homeostasis and evolvability and shows that a transcriptional network can develop memory and recall capabilities often thought to be reserved for classical cognitive systems. As a consequence of memory, genetic networks can exhibit predictive ability, enabling anticipatory behavior with respect to physiological stimuli (Tagkopoulos et al., 2008). A similar result was obtained for protein networks, showing that signaling via the tumor suppressor P53 could be modeled as a neural net (Ling et al., 2013), while MAP kinase pathways implement specific decision-making processes (McClean et al., 2007). Embryos make use of genetically encoded cellular memory, for example in the case of HOX gene expression patterns, which constitute a form of positional memory – “an internal representation by a cell of where it is located within a multicellular organism” (Chang et al., 2002; Rinn et al., 2006; Wang et al., 2009), and hysteresis in Hedgehog protein signaling (Balaskas et al., 2012), all of which are used to guide the subsequent activity of cells as a function of prior “experience”.

Additional memory media include the extracellular matrix (Becchetti et al., 2010; for plant cell walls see Humphrey et al., 2007; Seifert and Blaukopf, 2010; Hamann, 2015) and chromatin complex markings (Francis and Kingston, 2001; Maurange and Paro, 2002; Ringrose and Paro, 2004), both of which are ideal media for recording traces representing specific environmental and/or physiological events. These are examples of internal stigmergy – activity that leaves traces in a labile intracellular or extracellular medium which can be read as memories in the future by cells making decisions for migration, differentiation, apoptosis, or signaling (Theraulaz and Bonabeau, 1999; Ricci et al., 2007).

Importantly, many cell types communicate electrically, not just excitable nerve and muscle (McCaig et al., 2005; Levin, 2007a,b, 2012a; Bates, 2015). Recent molecular data show that developmental bioelectricity is an important modality by which cell networks process information that instructs patterning during regeneration, development, and cancer suppression (Levin, 2014a,b,c). Thus, one obvious candidate for cognition outside the brain is via the same mechanism used in the brain – bioelectrical networks (Levin and Stevenson, 2012; Mustard and Levin, 2014). Indeed it is likely that the processing in the brain is a direct extension (and speed optimization) of far older mechanisms used originally for morphogenesis (Buznikov and Shmukler, 1981; Levin et al., 2006). Developmental bioelectricity in animal systems features slowly-changing, continuous voltage changes as opposed to millisecond discrete (binary) spiking usually studied in the brain. However, the brain also includes non-spiking neurons (Victor, 1999) that have computational compartments similar to the membrane voltage domains observed in embryonic and other non-neural cells (Levin, 2007b; Adams and Levin, 2012). It has recently been proposed (Levin, 2012b, 2013; Mustard and Levin, 2014) that non-neural tissues support the same two types of plasticity as seen in the brain: changes of connectivity via electrical synapses (gap junctions) which corresponds to synaptic plasticity, and changes of ion channel function which corresponds to intrinsic plasticity (Marder et al., 1996; Turrigiano et al., 1996; Daoudal and Debanne, 2003).

In addition to computation via changes in resting potential, which is a primary regulator of pattern memory in embryogenesis and regeneration (Adams, 2008; Funk, 2013; Levin, 2014b), as well as of processing in the brain (Sachidhanandam et al., 2013; Yamashita et al., 2013), ion pumps such as the ubiquitous sodium-potassium ATPase, have been suggested as computational elements (Forrest, 2014).

Most of these bioelectrically active systems are based on ion dynamics at membranes which modify bioelectric fields via activities of ion channels and transporters (Taylor, 1974; Wayne, 1993, 1994; Hille, 2001). These membrane-associated electric fields feed-back on membranes and associated proteins (Jaffe, 1977; Tsong and Astumian, 1986; Westerhoff et al., 1986; Bezanilla, 2002, 2006, 2008). They also control endocytosis and vesicle trafficking (Antov et al., 2005; Baluška and Wan, 2012). Relevantly, even biochemical reactions are under electric control (Aragoněs et al., 2016; Xiang and Tao, 2016), as is transcription (Pai et al., 2015b) and chromatin modification (Carneiro et al., 2011; Chernet and Levin, 2014).

While the dominant model of neural-based cognition relies on the signaling dynamics among networks of neurons, it’s becoming increasingly appreciated that single neurons can execute subtraction, addition, low- and band-pass filtering, normalization, gain control, saturation, amplification, multiplication, and thresholding with respect to the input-output relations they implement (Koch and Segev, 2000). Memory and computation is thus not exclusively a multi-cellular phenomenon, and is not restricted to somatic neural cells. Recent computational studies have revealed conditions under which cells expressing ion channels can keep a stable memory with respect to resting potential, and these conditions do not specifically require neuronal cell identity – they can be fulfilled by numerous cell types, somatic as well as free-living (Ramanathan and Broach, 2007; Cervera et al., 2014; Law and Levin, 2015).

The amoeba of Dictyostelium discoideum migrate by extending pseudopods in an alternating pattern. The specific pattern of the pseudopods’ zig-zag behavior was recently shown to be predictable by viewing the cell surface as an excitable medium. In this model, the appearance of a pseudopod makes the local cortex temporarily more excitable (a kind of potentiation), while globally new pseudopods are inhibited. This model thus includes a memory of previous pseudopod locations, and quantitatively fits data from cell tracking experiments and the known chemotactic sensitivity of these cells (Cooper et al., 2012).

Budding yeast also keep a history which influences their future behavior – a memory of past events. They avoid pheromone-induced cessation of cell cycle after a deceptive mating attempt (failure to reach a putative partner cell within a specific time period). The mechanisms of this are beginning to be unraveled (driven by the dynamics of the maternally segregating G1/S inhibitor Whi3), and the authors term the macromolecular assemblies that mediate this memory “mnemons”, cellular structures that encode previous environmental conditions (Caudron and Barral, 2013). With respect to the search for the molecular substrate of specific memories, this yeast work may be ahead of similar efforts in the brain (Ungar, 1972, 1974a,b).

The flexible and versatile responses of bacteria to their environment has drawn significant attention of synthetic, molecular, and evolutionary biologists, as well as those interested in unconventional computational media (Miller and Koshland, 1977; Koshland, 1980; Ben-Jacob, 2009; Norris et al., 2011). Single bacteria are able to migrate toward beneficial targets, and away from noxious stimuli. The control algorithm for this behavior has long been the subject of investigation, with respect to the short-term memory needed for following gradients (Vladimirov and Sourjik, 2009) as well as “infotaxis” policies that do not require gradient sensing (Vergassola et al., 2007). Especially exciting are the recent findings that bacterial communities (biofilms) process information and make decisions about nutrient distribution and metabolism as an integrated whole, using ion channels (Prindle et al., 2015) and a kind of volume transmission as occurs in the brain (Agnati et al., 2006; Fuxe et al., 2013; Zhang et al., 2013). Ciliates (protozoa) exhibit learning and a form of memory, which even survives loss of nuclei and some cytoplasm (Gelber, 1962; Applewhite et al., 1969; Hamilton, 1975; Clark, 2013). The mechanism is unknown, but may involve electrical signaling (Applewhite, 1972; Kunita et al., 2014).

In addition to cells that make their living independently, single somatic cells from metazoan organisms also exhibit memory and decision-making (Albrecht-Buehler, 1985) during directed steering (Albrecht-Buehler, 1982) – a capability that also extends to cell fragments (Albrecht-Buehler, 1980) and even human sperm that adjust their flagellar beat to reach the egg via calcium-dependent tracking of chemical attractants (Alvarez et al., 2013).

The immune system has long been a paradigm of pattern recognition and classification (Carter, 2000). While the mainstream view of immune function is that of an evolutionary system driven by selection, a cognitive perspective has been proposed (Cohen, 1992a,b) as an alternative theoretical framework for understanding the body’s remarkable ability to distinguish self from non-self and adapt via immunological memory. It is interesting that the converse proposal has been made as well, to understand brain dynamics via a selectionist model (Fernando et al., 2012). Thus, in some sense, evolutionary and cognitive dynamics could be parallel (isomorphic?) ways to explain complex systems. If true in general, it may have significant implications for evolutionary theory.

An interesting kind of cognitive process is revealed by drug addiction. The increased tolerance with exposure is desensitization (one kind of basic memory element). Drug addiction reactions have been shown in somatic mammalian cells in culture (Corssen and Skora, 1964; Manner et al., 1974; Higgins et al., 1978), suggesting that this form of memory is not always a body-level phenomenon that necessarily involves the brain.

Physarum polycephalum is a slime mold that has been extensively used in studies of biological information processing (Nakagaki et al., 2004; Saigusa et al., 2008). By computing optimal paths for nutrients throughout its syncytial body, the organism can implement behavior that solves challenging spatial optimization problems, such as solving mazes and finding efficient highway layouts (Nakagaki et al., 2004; Saigusa et al., 2008; Adamatzky and Alonso-Sanz, 2011). This organism shows how the internal dynamics of morphogenesis, even at this primitive step toward a multicellular bodyplan, can implement decision-making and computation. It is particularly interesting that in this system, a kind of variational (least-action, or minimization) principle is explicitly implemented by a biological medium (Friston, 2010; Friston et al., 2015), providing a much-needed “base case” for starting to understand the common features of goal-directed activity across levels of organization from cells to body structures to organism behavior.

Physarum also shows evidence of memory. In their study of the Traveling Salesman Problem (requiring an optimal strategy for connecting regions in space), Zhu et al. (2013) found that when two individuals were created by dividing one individual, they remained correlated in their exploration even though they were spatially separated, suggesting the presence of a long-term memory in the intrinsic dynamics.

Although plants are still considered generally to be outside of neuronal and cognitive organisms, due to their lacking of animal-type of neurons and brains, plant cells have many features which are considered neuronal, including plasma membrane excitability supporting action potentials, acentriolar microtubules, motile Trans Golgi Networks, and synaptic-like actin-enriched cell-cell adhesion domains (Wayne, 1993, 1994; Barlow and Baluška, 2000; Baluška et al., 2003, 2005, 2008, 2009b; Baluška, 2010). Especially cells in root apices are very active in these neuronal-like activities and act as brain-like command centers (Baluška et al., 2004, 2009a,b, 2010; Baluška and Mancuso, 2009, 2013), navigating growing roots in their search for water and mineral nutrients in soil, and active root avoidance or escape from toxic, stressful and dangerous situations (Burbach et al., 2012; Yokawa et al., 2014; Yokawa and Baluška, 2015, 2016).

The classic studies on plants showing animal-like features and activities were accomplished more that 150 years ago by Charles Darwin, assisted with his son Francis Darwin, and Claude Bernard (Darwin, 1880; Bancroft and Richter, 1930; Perouansky, 2012). Later, Jagadis Bose accomplished his sophisticated experiments on plants, confirming and extending the previous results obtained by Charles Darwin and Claude Bernard (Shepherd, 2005). Despite the fact that plant action potentials are known for more than 150 years now, and these are known to control many plant processes (Wayne, 1993, 1994; Masi et al., 2009; Volkov et al., 2010; Sukhov et al., 2011; Böhm et al., 2016; Hedrich et al., 2016), plant action potentials are still ignored by the mainstream. For example, there is no single mention of plant action potentials in the book Plant Physiology by Lincoln Taiz, which represent the most accepted view of plants in biology (Taiz, 2010).