On Having No Head: Cognition throughout Biological Systems

Claude Bernard performed many anesthetic experiments. He expanded experimental materials from animals to plants. He showed that the Mimosa plant (Mimosa pudica), closing leaves upon touch, was unresponsive when exposed to a diethyl ether atmosphere which immobilized mice. Claude Bernard arrived at the conclusion that plants and animals share a common biological principle that is disrupted by anesthetics. He hypothesized that similarly as animals, also plants are able to actively sense their environment. He called this ability plant “sensitivity”. In order to test his ideas, he performed anesthesia on plants and the results of these experiments were presented in 1878 in “Leçons sur les phénomènes de la vie communs aux animaux et aux végétaux” (Bernard, 1878; Bancroft and Richter, 1930). Later, sensitivity of plants to anesthetics was confirmed not only for Mimosa and Dionea, but also for many other plants (Milne and Beamish, 1999; De Luccia, 2012; Gremiaux et al., 2014).

Similarly as neurons, plant cells are excitable and plant-specific action potentials serve for long-distance communication and integration of plant bodies. Action potential also control rapid plant organ movements such as closing the Dionea traps or touch-induced movements of Mimosa leafs (Volkov et al., 2010; Böhm et al., 2016; Hedrich et al., 2016). Our preliminary data with Dionea traps suggest that anesthetics block action potentials (Yokawa et al., in preparation). Moreover, action potentials control also nutrient transporters in Dionea prey-stimulated traps (Böhm et al., 2016; Hedrich et al., 2016). In the root apex, the transition zone is very active not only in electric activities (Masi et al., 2009), and synaptic-like cell-cell communication (Baluška et al., 2003, 2004, 2005, 2009a,b, 2010; Baluška and Mancuso, 2013), but also in sensory-based control of root growth navigation associated with high electric activity. Root apex navigation is based on complex computations as roots sample continuously huge amounts of abiotic and biotic information from their environment in order to find water and nutrient rich zones in soil; and to avoid dry, toxic and dangerous zones. Our data suggest that root navigation is controlled via computations accomplished at the root apex synapses and associated with electric activities (Masi et al., 2009).

Plants are emerging as excellent biological computational systems. For example, leaves maintain stable temperature near their surfaces despite large fluctuations of temperature in the atmosphere (Helliker and Richter, 2008; Pincebourde and Woods, 2012). They relay in leaf stomata which acts as plant thermostats tissue, with individual stomata acting as autonomous units showing collective behavior (Hetherington and Woodward, 2003; Peak et al., 2004). In the case of plant leaves, stomata are simultaneously the sensors of external information, the processing units that calculate gas exchange rates and sensitively regulate their controls. Plants solved the dilemma of optimal gas exchanges via elegant parsimonious computational techniques in which input, output, and processing are all accomplished by using the same hardware.

Additional nice examples of plant computation include the ability of plants to compute starch synthesis and degradation rates (Scialdone et al., 2013; Webb and Satake, 2015), root apex computation of numerous abiotic and biotic parameters to navigate optimally root growth in complex environment of patchy soil (Baluška et al., 2009a,b, 2010; Masi et al., 2009; Baluška and Mancuso, 2013), as well as computations accomplished via Dionea leaf traps (Volkov et al., 2010; Böhm et al., 2016). Action potentials are relevant for most (perhaps all) of plant-specific computations (Masi et al., 2009; Volkov et al., 2010; Böhm et al., 2016; Hedrich et al., 2016).

In the root apex transition zone, cells and their membranes oscillate in almost all their activities (Baluška and Mancuso, 2013). These root apex transition zones resemble presomitic mesoderm segmentation clocks underlying vertebrate embryo segmentation (Moreno-Risueno et al., 2010; Traas and Vernoux, 2010; Moreno-Risueno and Benfey, 2011).

A number of non-neural cells have been shown to exhibit memory, with respect to somatic position (Carlson, 1983; Chang et al., 2002; McCusker and Gardiner, 2014) or differentiation (Xiong and Ferrell, 2003), implemented via long-term stable changes in bioelectric state (Marder et al., 1996; Turrigiano et al., 1996; Rosen and Cohen, 2006) and transcriptional profile (Kragl et al., 2009; Wang et al., 2009). These are now beginning to be understood via physiological modeling and dynamical systems theory that views memories as attractors in transcriptional, bioelectric, or epigenetic state space (Huang et al., 2005; Cervera et al., 2015; Law and Levin, 2015).

Moving up in organization, several tissues have been suggested to exhibit memory. One is bone, which has many similarities to a neural network, both molecularly and functionally (Turner et al., 2002). For example, the neurotransmitter glutamate plays a role in cell-to-cell communication among bone cells. Glutamate of course is a key neurotransmitter for learning and memory in the hippocampus. Bone cells exhibit habituation (to repeated mechanical stimuli) and sensitization (to mechanical loading) – two of the most basic components of memory. Skull bones react quite differently to mechanical loading and hormones than do long bones, and it has been speculated that the past history of weight bearing imparts long-term cellular memory to the bone cell network, manifesting as differential responses to a variety of stimuli. A model involving long-term potentiation via the NMDA receptor has been proposed to explain memory of past stresses, and its subsequent influence over growth control, has been proposed (Spencer and Genever, 2003; Ho et al., 2005). Muscle comprises of some of the largest cells of animals, and also process, store and retrieve information via muscle-specific memory which can last from 15 years up to the entire lifetime in humans (Bruusgaard et al., 2010; Gundersen, 2016).

A most interesting set of studies have examined the phenomenon of cardiac memory. This is a clinically important pathway, in which specific changes of heartbeat pattern can persist stably (Otani and Gilmour, 1997; Goldberger and Kadish, 1999; Rosen and Plotnikov, 2002). This phenomenon has been modeled as a simple memory-like quantity that determines the relationship among the durations and amplitudes of action potentials (Otani and Gilmour, 1997). Most importantly, a specific mathematical model has been proposed for cardiac memory, taken after Hebbian plasticity in the brain (Chakravarthy and Ghosh, 1997; Zoghi, 2004).

The most recent addition to this body of work is the study of pancreas physiology (Goel and Mehta, 2013), which studied gap junctions (electrical synapses used for ionic communication in the brain, heart, and other organs), and their role in secretion of insulin from the pancreatic islets of Langerhans in response to glucose stimulation. Gap junctions synchronize oscillations of resting potentials among beta cells, and thus control insulin secretion. Past measurements of gap junctional conductance was unable to explain systemic properties, such as diminished junctional coupling in type-2 diabetes. In contrast to the prevailing tendency to focus on bottom-up views of the molecules involved and their interactions, Goel and Mehta viewed the process top–down, as a learning-like adaptation. Modeling gap junctions as links in a network of beta cells, subjected to homeostatic plasticity, they elucidated the system-level properties of this tissue, explaining why reductions in gap junction-mediated coupling in diabetes is necessary for an increase in blood insulin levels following hyperglycemia. It is not yet known if these mechanisms also underlie classical studies by Pavlov and others (Gantt, 1974, 1981; Gantt et al., 1987) on the classical conditioning of body organs to sugar, adrenaline, histamine, and other physiological stimuli.

We next consider larger-scale multicellular systems, at the level of organs or whole bodyplans (Levin, 2012b). Many species’ bodies exhibit pattern memory during regeneration (Baddour et al., 2012; Lobo et al., 2014), and even transplanted organs maintain spatial information, such as transplanted eyes which send out optical axons to penetrate the brain on the side corresponding to its former location in a donor animal (Koo and Graziadei, 1995). It should be noted that one challenge to multicellularity is the ever-present danger of cancer – defection of somatic cells from the anatomical goals of the organism toward more primitive “every man for himself” behavior of individual cells and tumors (Johnston et al., 1992; Vincent, 2012; Chen and He, 2016). The interplay between the tumor and host has been analyzed using game theory (Dingli et al., 2009; McEvoy, 2009), consistent with each being an autonomous system with internal and external information channels, goals, and functional capabilities. Control networks regulating cancer have been analyzed from the perspective of learning (Gyurko et al., 2013), which represents an interesting new area for further research. Interestingly, recent data implicate in carcinogenesis the same bioelectric mechanisms that orchestrate pattern regulation and keep cells away from tumorigenesis (Chernet B. and Levin, 2013; Chernet B.T. and Levin, 2013; Yang and Brackenbury, 2013; Bates, 2015).

The first task of any animal body is to assemble the progeny of a fertilized egg cell into a specific 3-dimensional pattern during embryogenesis. Then comes the need to maintain anatomical integrity over the lifespan, despite individual cell senescence, injury, and neoplastic conversion. Thus, long before animals developed brains to execute adaptive behaviors, cells had to have ways to coordinate their activity in an exquisite ballet that self-assembles, and then continuously remodels and repairs, a complex anatomical form. Some animals (e.g., salamanders) can regenerate their limbs, eyes, jaws, hearts, and portions of the brain (Sanchez Alvarado and Tsonis, 2006). Mammals have reduced powers of regeneration, but deer regenerate antler bone (adding up to 1 cm per day) every year, while humans regenerate their livers, and children regenerate their fingertips. Tails grafted onto the sides of salamanders slowly remodel into limbs (a structure more appropriate to their new location), and mammalian embryos can be split in half or combined together, resulting in normal embryos (reviewed in (Mustard and Levin, 2014)). All of these capabilities require significant information storage and processing, and many take place prior to (or without) the presence of the CNS.

Embryogenesis, regeneration, and metamorphosis stop precisely when the correct anatomical shape has been produced; this is a process akin to goal-directed behaviors, in the sense that the system can pursue multiple paths toward the same (anatomical) goal state, can accommodate unpredictable external perturbations (is not hardwired but flexible), and rests when it is satisfied (can recognize when its goal is achieved). All of these examples show the remarkable information processing that cells carry out, in order to create and maintain specific shapes (Levin, 2012b). Analogously to how brains implement goal-seeking behavior via information processing, non-neural cell networks process information about current and future anatomical shape. While the brain operates muscles and glands in service of activity in ecological space, the computational processes of non-neural somatic networks control cell behaviors (differentiation, migration, proliferation) to optimize the body’s movement through morphospace (Stone, 1997; Rasskin-Gutman and Izpisua-Belmonte, 2004; Newman and Bhat, 2009).

A primary goal of developmental biology, synthetic bioengineering, and regenerative medicine is to learn to understand and control patterning networks, for applications in birth defects, organ regeneration, and cancer reprogramming (Ingber and Levin, 2007; Doursat et al., 2013). In particular it is crucial to tame the endogenous closed-loop pattern regulatory systems (flexible remodeling and regeneration pathways), as these offer the opportunity to exploit modularity to achieve needed changes in growth and form without micromanaging the details. What mechanisms underlie the ability of tissues to measure large-scale shape, detect deviations from a “remembered” correct target morphology, implement remodeling toward repairing that shape, and know when to stop (Levin, 2011)? Recent work has shown that as in the brain, these control networks make use of ion channels, gap junctions (electrical synapses), and neurotransmitters (Levin, 2012a; Tseng and Levin, 2013). A parsimonious hypothesis is that this is no coincidence, and that the brain learned its prodigious computational tricks from far more ancient pathways, co-opting developmental bioelectricity and optimizing it for the speed needed for behavior. While the brain operates on millisecond-scale bioelectric spiking, developmental bioelectricity involves steady, slow changes in ion fluxes, resting potentials, and electric fields.

A long history of work implicated bioelectric events in patterning (Jaffe, 1981; Nuccitelli, 2003; McCaig et al., 2005). Recent advances in molecular physiology have revealed that gap junctions, ion channels, and neurotransmitter pathway molecules – workhorses of cognitive processes in the CNS – are broadly expressed throughout the body, beginning prior to fertilization. Analogously to the brain, non-neural tissues continuously regulate resting potential (Vmem) and local field potentials (extracellular electric fields), as well as regulate the movement of neurotransmitters among cells (Pullar, 2011; Bates, 2015).

Signaling mediated by bioelectric events plays a crucial, instructive role in pattern formation (Funk, 2013; Levin, 2014b). Ion channel-mediated changes in Vmem not only affects individual cell behaviors such as proliferation, differentiation, apoptosis, and migration (Sundelacruz et al., 2009), but also determines large-scale parameters such as organ size, shape, and axial patterning of the entire body (Beane et al., 2011; Perathoner et al., 2014). In a range of model systems, Vmem regulates the formation of the brain, eye, wing, and face, and controls patterning along the anterior-posterior and left-right axes during embryonic development (Levin et al., 2002; Dahal et al., 2012; Pai et al., 2015a). Moreover, experimental control of bioelectric gradients has enabled induction of regenerative ability in non-regenerative contexts (Tseng et al., 2010; Leppik et al., 2015), induced reprogramming of gut tissue into complete eyes (Pai et al., 2012), and normalized tumors (Chernet B. and Levin, 2013; Chernet B.T. and Levin, 2013). Electrical synapses (gap junctions, GJs) and neurotransmitters like serotonin are a key component of several patterning systems, having been implicated in embryonic left–right asymmetry, bone patterning, tumor suppression, and brain size control (Levin and Mercola, 1998; Iovine et al., 2005; Chernet et al., 2015; Pai et al., 2015a). As in the brain, these elements often work together, such as the bioelectrically controlled movement of serotonin through GJs during left-right patterning and control of nerve growth (Levin et al., 2006; Blackiston et al., 2015). The molecular pieces are now being identified, but the idea of neurotransmitters being ancient “pre-nervous” developmental signaling molecules is an old one (Buznikov and Shmukler, 1981).

The analogy between the brain and somatic pattern control (Figure 2) makes several specific predictions. One is that ion channels, GJs, and neurotransmitters should play a role in development; this has been amply demonstrated by the identification of patterning channelopathies (Levin, 2013), functional experiments in regenerative and developmental biology (Stewart et al., 2007), and the teratogenic effects of numerous psychoactive drugs (Hernandez-Diaz and Levin, 2014). Another key prediction concerns the encoding of instructive information. In the brain, genetics establish the hardware – genes encode the available components and thus define the limits of cellular activity. However, the information content of the brain is not directly encoded by the genome, but rather arises dynamically through environmental stimuli (learning) and self-organizing dynamics of the electrochemical circuitry (plasticity). Is this the case in pattern formation as well?

Can “long term somatic memory” be edited, in the context of a wild-type genome, leading to a permanent change? A first example of this was shown in a different species of planaria (Nogi and Levin, 2005), where targeting GJs for just 48 hours in a chunk of tissue caused it to regenerate 2 heads – one at the former anterior end (normal), and one at the posterior-facing end (which would normally grow a tail). Strikingly, these 2-headed worms continue to regenerate as 2-headed when cut in subsequent rounds of regeneration, in plain water, months after the GJ blocking reagent is long gone from the tissue (Oviedo et al., 2010). The target morphology – the shape to which this animal regenerates upon damage – has been permanently re-written by temporarily editing the physiological network. This finding has clear similarity to plasticity [well-known to be exhibited by electrical synapses (Pereda et al., 2013)]: a brief induced change of GJ connectivity becomes stabilized to a long-term change (Levin, 2014a). This interaction between bioelectric activity and voltage-gated GJs makes developmental bioelectrical networks especially suitable as a labile yet stable memory medium (Palacios-Prado and Bukauskas, 2009). Another brain-like property exhibited in this effect is its holographic nature: in each round of cutting, the ectopic head (perhaps “epigenetically reprogrammed”) is removed, and a middle fragment of the gut still knows it must make 2 heads if cut. The patterning information is distributed non-locally throughout the network.

This field is advancing rapidly in its mechanistic details at the cellular level: the genetics of endogenous ion channels causing the gradients, the transduction mechanisms that control transcription after Vmem change, and the gene expression changes downstream of bioelectrical signaling are all becoming clear (Yang and Brackenbury, 2013; Pai et al., 2015b). Techniques, such as optogenetics (Adams et al., 2013, 2014), are allowing imposition of specific voltage patterns onto tissue in vivo. As in the brain, where optogenetics is used to insert memories directly into brains (Ramirez et al., 2013; Liu et al., 2014), these techniques will be crucial to learn to rewrite pattern memories during regeneration or embryogenesis. However, as in neuroscience, there is more than one level at which progress needs to be made. A mature understanding of the brain requires synthesis of data from people working on the genetics and biochemistry of specific neurotransmitter receptors and their downstream molecular signaling, with the insights of workers at the level of circuits, behavior, cognitive science, and psychology.

Classic work explored the extensive parallels between chemical gradients during development and signal processing in the visual system (Grossberg, 1978), and indeed early quantitative models of patterning (explaining self-regulatory features like proportion regulation) were based on visual system function (Hartline et al., 1956; Gierer and Meinhardt, 1972). More recent efforts include the notion of memory for position during regeneration (Chang et al., 2002; Kragl et al., 2009; Wang et al., 2009) and development (Beloussov, 1997) and for signaling hysteresis during development (Balaskas et al., 2012), excitable cortex memory models of pseudopod dynamics (Cooper et al., 2012), and neural network models of chemical signaling (Ling et al., 2013) (which showed formal isomorphisms between gene regulation networks and Hebbian learning in neural nets) (Watson et al., 2010; Ling et al., 2013). In addition to classical neuroscience concepts, more exotic group cognition models have been applied to patterning (Gunji and Ono, 2012), while a few recent studies investigated the decision-making and formal computational capabilities of RD systems – a chemical signaling modality often used to model morphogenesis (Adamatzky et al., 2003, 2008; Costello et al., 2009; Dale and Husbands, 2010, which is now known to be Turing-complete (Scarle, 2009) and support semantic interpretations (Schumann and Adamatzky, 2009). Despite these fascinating efforts to identify elements of cognitive-like processing in well-known elements of pattern formation, developmental biology is still firmly centered in a mechanistic perspective, seeking explanations in terms of pathways and not information (systems that know things and make decisions based on that understanding).

However, it is crucial to note that attributing true knowledge and memory to biological systems is not mystical thinking – computational neuroscience shows us a clear proof of concept that information-level, cognitive approaches to cellular networks are viable, and in fact necessary, strategy for understanding a system at all of its salient levels.

Thus, neuroscience offers developmental biology more than just tools and molecular mechanisms: it offers a unique paradigm, otherwise unavailable to molecular and cell biologists, of the emergence of higher levels of organization that have both causal potency and experimental tractability. The field is in need of new formalisms and conceptual tools for linking the dynamics of physiological circuits with downstream patterning outcomes. Developmental biology is currently focused entirely in a bottom-up mode, with molecules being the preferred level of explanation. Neuroscience teaches us that we must look upward as well as downward, for emergent levels with their own rules and advantages (Friston et al., 2015). For example, training an animal to a particular complex behavior is far more efficient than attempting to elicit the same behavior by manipulating individual neurons in their brains. We now know that beneficial changes at the genetic and chemical levels can be induced by cognitive therapies – top–down control of tissue structure and function induced by specific thoughts and experiences. If patterning tissues are “primitive cognitive agents”, in the sense that they can be profitably understood as memory-bearing, information processing, goal-directed cybernetic systems (Pezzulo and Levin, 2015), then a whole new set of approaches becomes available for regenerative medicine. If we understood the bioelectric code, we could interact with it at these higher levels of organization, taking advantage of endogenous modularity and perhaps rationally controlling anatomical outcomes without having to micromanage molecular networks. In this field, cognitive science, unconventional computation, and developmental biology intersect.

A fundamental open direction is the search for a computational pipeline to extract goal patterns from bioelectric state data, parallel to efforts to extract image data from brain measurements (Nishimoto et al., 2011). The flow of knowledge will likely not all be unidirectional: cracking the bioelectric code in patterning tissues is likely to in turn benefit fundamental neuroscience by showing, in perhaps a simpler context how to extract semantic content from bioelectrical cell states in the brain.

How does biological matter give rise to decision-making, memory, representation, and goal-directed activity? Implementation-independence is a core principle of computer science: an algorithm does what it does regardless of what kind of medium is implementing the steps. However, in the biological sciences, the study of memory and other cognitive functions has largely been the province of neurobiology, which studies the information processing and computational functions of one type of system: collections of neurons. Instead, we have surveyed a broad range of systems at various scales, from molecular to organismal, which have their own distinct ability to process information, make decisions, and achieve specific goal states (Figure 3). Neural-like computation, decision-making, and memory have been reported in sperm (Alvarez et al., 2013), amoebae (Zhu et al., 2013), yeast (Caudron and Barral, 2013), and plants (Gagliano et al., 2014), using ubiquitous mechanisms like cytoskeletal elements which appear to be also involved in neural information processing (Sahu et al., 2013). It is clear that neural networks have no monopoly on such functions. Remarkably, it is not only the positive (adaptive) cognitive functions that are widely conserved: some of the same illusions to which advanced brains’ perception and rational reasoning fall prey are being found in systems from slime molds to multi-animal colonies (Beekman and Latty, 2015; Sakiyama and Gunji, 2016).

McCulloch said “Why the mind is in the head? Because there, and only there, are hosts of possible connections to be performed as time and circumstance demand it” (McCulloch, 1951). Given the facts of protein, cytoskeletal, transcriptional, and bioelectric networks, it appears that many different media at various scales have the ability to form and rewire experience-dependent connections. The “dynamical hypothesis” (van Gelder, 1998) asks, what if the brain is better understood as a dynamical system, than a computational one? We invert this hypothesis, and ask what if some dynamical systems are better understood as cognitive agents? The appearance of memory and computation at many levels of biological organization suggests a fractal organization of cognitive subsystems within systems – molecular, cellular, tissue, and body-wide (Figure 2). This has been suggested in the brain [Smythies’ nested doll hypothesis, (Smythies, 2015)] but may indeed exist throughout the biological world. Whether each successive level of organization is in some sense smarter than the ones below it, or whether structures derive their cognitive powers from those of lower levels, remains to be discovered. It should be noted, however, that even in advanced brains, the relationship between cognitive capacity and biological structure is not trivial to pin down, as shown by the occasional example of potent function in the presence of severe structural deficits (Lorber, 1978, 1981; Nahm et al., 2012).

The hypothesis of nested, widely prevalent cognitive layers suggests a rich research program, including: (1) the development of improved methods for reading/writing bioelectrical state information into somatic tissues and sculpting non-neural bioelectric circuits (optogenetics beyond excitable cells and in the synthetic biology of gap junction and neurotransmitter signaling; Adams et al., 2013), (2) continued work on cracking the bioelectric code (bioelectric state information maps onto the topology of various patterning outcomes in tractable model systems such as planaria; Tseng and Levin, 2013), (3) formulation and testing of quantitative, molecular models of LTP, habituation, sensitization, plasticity, and higher-order learning applied to protein interaction networks, gene regulatory circuits, cytoskeletal dynamics, and cell behavior during morphogenesis, (4) use of reagents that impact cognition (hallucinogens, anesthetics; Kawamoto et al., 2005), stimulants, nootropics/cognitive enhancers, etc.) in cellular, developmental, and regenerative patterning assays to probe conservation of pathways between neuroscience and morphogenesis, (5) creation of larger-scale computational models of regeneration and functional experiments in morphogenesis based on goal-seeking and error minimization algorithms with molecularly specified metrics (Slack, 1980; Chao et al., 2008), (6) exploration of molecular models of cognitive concepts (attention, autism spectrum, sleep, visual illusions/hallucinations, addiction) in specific patterning and mispatterning contexts, (7) bioengineering platforms that reward and punish in vitro patterning systems for specific changes in growth and morphogenesis (instrumental learning and top–down control of shape in developmental or regenerative contexts), and (8) a mechanistic investigation of the mechanism of persistence of memories through massive b

rain regeneration, which is likely to reveal the interface between somatic and neural memories (Blackiston et al., 2008; Shomrat and Levin, 2013).

We have avoided here the thorny issues of philosophy of mind that arise from trying to define exactly under what conditions words like “knowledge” are appropriate, in favor of an intentional stance-like pragmatic, engineering approach grounded in cybernetics. The coverage of cognitive terms across biology can expand to the extent that information-centered approaches are shown to be effective in predicting and controlling the behavior of biological systems. The practical implications for biotechnology, unconventional computation, and regenerative medicine are enormous. Equally likely, the lessons we learn from unconventional cognitive systems will deeply impact our most basic understanding of how mind emerges from the brain.

ML and FB both contributed sections to the review according to their specialties. Both provided novel ideas, edited each other’s text, and prepared the finished product.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.