Thoughts from the forest floor: a review of cognition in the slime mould Physarum polycephalum

In the two-armed bandit test, plasmodia were placed between two environments that differed in food site availability and profitability along their length, thereby constituting the ‘arms’ of a two-armed bandit (Fig. 1 g). Plasmodia could extend pseudopodia into each arm to explore them, and then cease exploring one option to favour exploiting the other, once a decision had been made. By providing a range of different testing scenarios, the study demonstrated that Physarum compares the relative qualities of available options, integrates over sequential samplings to perform well in unpredictable environments, and combines information on both reward frequency and magnitude to make correct adaptive decisions (Reid et al. 2016). Increasing the level of difficulty, the researchers proposed 10 different heuristic rules of varying complexity that Physarum potentially could use to accurately exploit information to maximise food intake. These ranged from extremely simple rules (autocorrelation: move in the same direction as the previous timestep) to the only provably optimal method for solving the bandit problem, the Gittins index (select the arm with the highest index, which takes account of future expected rewards from both exploration and exploitation of an arm, based on a Beta prior over its expected Bernoulli reward probability, and a discount parameter applied to future rewards). Comparing the performance of each model to the experimental data via Bayesian model selection, Physarum proved to operate at the mid-level of complexity, where the probability of exploring each arm is proportional to the number of rewards previously encountered on that arm (Reid et al. 2016).

This heuristic is computationally far simpler than the provably optimal strategy, yet it performs nearly as well, and can be employed on a fully decentralised basis by reinforcing exploitation of locally sensed areas of profitability in the environment. It is also mathematically and conceptually similar to the matching law (Poling et al. 2011), where relative rates of responding to a stimulus match relative rates of reinforcement for the stimulus—a pattern widely observed in vertebrates from pigeons (Herrnstein 1961) to rats (Sanchis-Segura et al. 2005), coyotes (Gilbert-Norton et al. 2009) and humans (Alferink et al. 2009), when certain reinforcement schedules are applied.

Studies such as those above have gone some way to describing the how of Physarum decision-making, but what do we know of the where? When examining the physical mechanisms of decision-making in the brain, researchers can determine the location and pattern of electrical signals within and between different brain regions. Ray and colleagues (2019) devised an analogous method to examine the sites of decision-making in Physarum. Using a single tubule of plasmodium stretched between two food sources, the authors measured the amplitude and frequency of contractions at 50 equally spaced locations along the length of the tubule, as the organism decided which food source to exploit (Fig. 1i). Using the information-theoretic measure of Transfer Entropy (Schreiber 2000), the researchers could then determine how different regions of the slime mould responded to the information, and how this information was transferred to other regions. When both food-source options were identical, contractile regions nearest each of the options act as information sources, while those at the tubule midpoint act as information destinations. When food options differed in quality, the regions near the rejected food source transfer information towards the regions near the chosen food source. This appears counter-intuitive to other models of information flow within Physarum, but the authors point out that their results do not indicate whether information transfer occurs via increasing or decreasing contraction properties, and their method does not allow for establishing a direct causal relationship between contraction properties and decision-making.

Despite this caveat, the amount of information transferred between tubule regions was fourfold greater when there was a fivefold difference in food source quality, which provides some evidence for a causal link between information transfer and decision-making.

In contrast to the many definitions of decision-making, memory has a single generally accepted definition: the means by which information is stored and retrieved (Kilian and Muller 2002; Sweatt 2009). Thus, the phenomenon is widespread among both biotic and abiotic systems, from the brain allowing you to read this article, to the computer I am using to write it. Even so, the majority of cognition research couches memory in a specifically neural context, perhaps in part due to memory being a key prerequisite for that perennial favourite of cognition researchers: learning (discussed below). Thus, it is unnecessarily surprising that Physarum has been demonstrated to exhibit multiple forms of memory.

Tubules within the greater plasmodial network that lie close to a food source were recently shown to be thickened and reinforced, by local release of a softening agent that facilitates its own transport through the network (Kramar and Alim 2021). Local growth occurs at the expense of tubes more distant to the food source, or which flow in directions that do not align with the attractant’s location. This suggests that alteration of tubule properties can act as a form of physical memory encoding the location of a food source, and that this encoded memory can be ‘read out’ upon discovery of a new nutrient stimulus encountered in the same direction. In this context it could be argued that tube thickening equates to ‘memorising’ the attractant’s location, and retraction of tubes that lead elsewhere equates to ‘forgetting’.

Physarum’s ability to sense its own trail of extracellular slime allows it to essentially build up a map of its explored environment. Rather than storing this spatial information inside itself, the information is encoded in the external environment, and retrieved whenever it encounters the trail (Reid et al. 2012). While showing that this memory system is not necessary when navigating in simple environments, the authors demonstrate that when navigating complex environments, such as escaping a U-shaped trap to obtain a food source, Physarum’s ability to utilise its externalised spatial memory dramatically enhances its navigational efficiency (Fig. 1j). Sims and Kiverstein (2022) argue that this constitutes an example of extended cognition, where at least some of the heavy lifting of cognitive processing is performed by entities located within the local environment and external to an agent’s body (Cheng 2018; Clark and Chalmers 1998; Gillett et al. 2022). Furthermore, they suggest that Physarum’s use of ECS (extracellular slime) corresponds to cognitive niche construction, “the process of actively building structures in the local environment that aid learning and problem-solving.” (Sims and Kiverstein (2022), citing (Clark 2008; Wheeler and Clark 2008)).

Physarum’s use of ECS in more complex mazes than a U-shaped trap has been explored in a follow-up study by Smith-Ferguson and colleagues (2017). Small plasmodia were challenged to navigate through (1) ‘open’ mazes with obstacles to avoid, but not arranged in thin bounded channels as in a labyrinth; (2) ‘simple’ bounded labyrinth mazes with only three paths and 2 decision points; and (3) ‘complex’ bounded labyrinths with 3 decision points and several long, dead-end pathways. The experiments were then repeated in mazes with agar pre-coated in ECS to disable the focal plasmodium’s ability to utilise memory. External memory was found to enhance navigational efficiency in the open and simple bounded mazes but not in complex mazes. In the complex mazes, plasmodia that happened to make a choice leading to a dead-end were prevented from quickly retracing their steps to get back on the correct track. Thus, at least in these artificial scenarios invented by experimenters, the simple heuristic of avoiding areas previously explored can be a handicap rather than an advantage.

Physarum can also build up a memory of periodic events and anticipate their predicted approach. When subjected to conditions of cold, dry air, plasmodia respond by slowing their locomotion (Saigusa et al. 2008). If this negative stimulus is applied in short bouts at regular intervals such as each hour, plasmodia would respond—after only three such intervals—by spontaneously slowing down locomotion on the fourth interval, even when the negative stimulus was not applied. Continued absence of the negative stimulus led to resumption of normal, sustained locomotion, but the same anticipatory response could be evoked, even six hours later, after a single application of cold dry air. These results indicate that Physarum possesses some cellular mechanism for memorising periodicity and recalling this periodicity at a later time. Rats have been shown to time intervals using a self-sustaining endogenous oscillator (Crystal 2006); hence it is possible (though yet untested) that Physarum utilises its own endogenous oscillations to time intervals.

With the ability to store and retrieve information about past events, organisms have the potential to change their behaviour based on their recalled experience. If that behavioural change imparts a fitness benefit, then natural selection should favour those organisms. Many definitions would classify these organisms with the ability to learn (Ginsburg and Jablonka 2009). Hence, many recent studies have focussed on defining the learning capabilities of Physarum, beginning with one of the simplest forms of non-associative learning: habituation.

Habituation occurs when an organism decreases its response to a stimulus after repeated or prolonged exposure to that stimulus (Perry et al. 2013; Shettleworth 2009). Physarum plasmodia that are separated from access to a food source by a bridge of repellent quinine-agar show clear aversive responses to the quinine-agar (Boisseau et al. 2016). However, when this stimulus was repeated for 5 days, plasmodia gradually reduced their aversive response, habituating to the negative stimulus. This response was repeated with another repellent, caffeine, but quinine-habituated plasmodia did not reduce their aversive response to caffeine, and vice versa. By demonstrating response specificity, the researchers ruled out the potential confounding factors of fatigue or general sensory adaptation.

A follow-up study demonstrated that the learned response survived the process of cell–cell fusion with a clonemate plasmodium (Vogel and Dussutour 2016). By placing habituated and unhabituated plasmodia next to each other and allowing them to fuse, pseudopodia which crossed the repellent–agar bridge were just as likely to originate from the region of previously unhabituated biomass as the habituated region. While the authors describe this result as unhabituated plasmodia “directly acquir[ing] a learned behaviour from a habituated slime mould”, it is incorrect to label these regions as individual plasmodia after cell–cell fusion has occurred. Indeed, the authors state that “extensive protoplasmic mixing took place”, quickly rendering the fused clones as a single entity of mixed protoplasm from habituated and unhabituated donor plasmodia. A later study provided supporting evidence that the mechanism underlying habituation, and its transferability between plasmodia, is high levels of the repellent stimulus itself (in this case NaCl salt) being taken up into the cell and distributed throughout the protoplasm as a ‘circulating memory’ (Boussard et al. 2021). By contrast, Smith-Ferguson and colleagues (2022) found the opposite result in a study published a year later. Plasmodia repeatedly exposed to NaCl were more likely to avoid the salt. This appears to show an example of sensitisation, another form of learning. The researchers argue this could be due to a greater build-up of NaCl in the latter study; plasmodia in the earlier study had prolonged but non-repeated exposure to the stimulus.

While habituation has been characterised as the “simplest” (Hawkins and Kandel 1984; Rose and Rankin 2001) or “most ancient” (Van Duijn 2017) form of learning, the successful demonstration of any form of learning in a brainless organism is an achievement in itself. The question of learning in unicellular organisms was hotly debated in the early twentieth century, the result of which was the prevailing view that non-associative learning was possible for these ‘simple’ creatures, but not higher forms of associative learning such as Pavlovian conditioning. Later scientific attempts to disprove this notion were generally condemned on grounds of non-reproducibility or misinterpretation (see Gershman et al. (2021) for an excellent review of the topic). Recent success in demonstrating ‘entry-level’ learning abilities in Physarum begs the obvious question of whether Physarum is capable of associative learning. Such a demonstration could have widespread impacts (or maybe not, see “Discussion”).

Only one published study has claimed to demonstrate associative learning in Physarum (Shirakawa et al. 2011). As discussed in the subsequent literature (Dussutour 2021; Krause et al. 2022; Loy et al. 2021), and conceded by the authors themselves, the observed results have more parsimonious explanations. Pairing the food reward stimulus with the conditioned negative stimulus of low temperature, trained plasmodia were observed to move towards both stimuli, while untrained control plasmodia avoided the low-temperature option. Temperature affects many aspects of slime mould physiology, including metabolic rate, movement speed and rates of chemical uptake and sensing, and so could introduce confounding factors that could be interpreted as association. Loy et al. (2021) say the data presented “are not sufficient to assess the effectiveness of the conditioning treatment”. Dussutour (2021) posits that if learning did occur, the plasmodia are likely to have habituated to the low temperature. While the possibility of associative learning in non-neural organisms is daily becoming less extraordinary, common acceptance of the notion amongst animal cognition researchers requires extraordinary evidence, which has so far yet to be produced (but see Carrasco-Pujante et al. (2021) for recent evidence of associative conditioning in unicellular amoebae).

Physarum clearly possesses many of the hallmarks of cognition, including sensing, communication, navigation, decision-making, memory and learning. The increasing popularity of behavioural research in easy-to-use Physarum points to the slime mould emerging as a model for non-neural cognition. Doubtless, as further research progresses more abilities from the traditional cognitive tool-kit will be added to Physarum’s demonstrated repertoire.

Beyond adding demonstrations of further cognitive capabilities, future researchers also need to focus on understanding the intracellular mechanisms of cognitive behaviour in this protist. While there has been extensive measurement and modelling of the oscillatory system (Gao et al. 2019; Nakagaki et al. 1999; Wohlfarth-Bottermann 1979), the work to understand the molecular underpinnings of cognition has only just begun, most notably with the role of an intracellular signalling agent (Kramar and Alim 2021) and the role of chemical retention during habituation (Boussard et al. 2021). This is in stark contrast to the wealth of accumulated knowledge on the molecular machinery of neural cognitive systems. While seeking to close this knowledge gap, researchers must also pay attention to appropriate experimental design for testing cognition in organisms that often operate at such a different temporal and spatial scale to our own (for a detailed outline of these challenges, see Reid et al. 2015).

Classic models of cognitive processing have drawn a line between the cognitive and non-cognitive organisms based on the flow of sensorimotor information. Non-cognitive organisms are defined by reactions to external stimuli without internal feedback between the stimulus receptor and the site of action, while cognitive organisms are those that modulate the receptor via internal neural feedback from the site of action (Fuster and Bressler 2012; Reid and Latty 2016; von Uexküll 1926). The emerging trend of broader phylogenetic inclusivity in cognitive research has led to descriptions of other sensorimotor feedback systems that need not rely on neurons, such as the two-component signal transduction system of the bacterium E. coli (van Duijn 2006), and the coupled-oscillation system of Physarum (Reid and Latty 2016). Indeed, Baluška and Levin (2016) point out that neurons are ill-deserving of their reputation for “magical, unique” cognitive properties, because cognitive computations may arise 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.” This statement itself could serve as a description of Physarum and its mode of action.

Proponents of basal cognition remind us that neural networks evolved from far more ancient signalling pathways; neurons mainly optimised existing mechanisms for speed (Sterling and Laughlin 2015). Taking this perspective offers new avenues for understanding the evolution of not only cognition, but also of nervous systems themselves. Sensorimotor coordination, described as “the process by which organisms adaptively coordinate their sensors and effectors to optimize the external conditions for their metabolism and homeostasis” (van Duijn 2017), is an ancient strategy that enabled complex forms of cognition to evolve. It is also a popular hypothesis for why nervous systems evolved in the first place. This is in part attributed to size: tiny organisms can function adequately using sensorimotor systems based on cilia, while larger, more complex organisms require something like a nervous system to support their muscle-based locomotion. According to this ‘moving hypothesis’ (Llinás 2002), higher motility and larger size led to the exploitation of more heterogeneous environments and development of evolutionary arms races that fuelled a massive explosion in morphological and behavioural diversity. Physarum mirrors this pattern: it has high motility, including an extremely fast rate of cytoplasmic streaming (up to 1 mm/sec) coupled with an ability to reach enormous sizes for a single cell. Physarum’s large size is in part facilitated by its unique contractile mechanism of information transfer, which may in turn have enabled it to access more heterogeneous environments, leading to the exploitation of higher cognitive niches than your average protist.

Taking the more phylogenetically inclusive approach to cognition could benefit our understanding of extant decision-making systems as well. One of the most widely accepted models of decision making (the drift–diffusion model), is based on how neurons interact within the brain. ‘Evidence’ in favour of competing options (in the form of firing rate) builds in competing neurons, until one of them exceeds a decision threshold (Bogacz et al. 2006; Chittka et al. 2009; Livnat and Pippenger 2006). A strong analogy exists here with the oscillation system of Physarum (Reid and Latty 2016). More recently, oscillation patterns in Physarum plasmodia were observed to be highly dynamic, consisting of interlaced regular and irregular contraction patterns, similar to neural activity observed in nematodes and fruit flies (Fleig et al. 2022). These observations, coupled with the ‘non-adaptive’ examples of speed-accuracy trade-offs and irrational decision-making in Physarum (not to mention the evidence for cognitive capabilities in other non-neural taxa), provide strong evidence for fundamental principles of information processing and decision-making that span the majority, if not the entirety, of the phylogenetic tree (Reid and Latty 2016). Embracing this viewpoint could have significant and measurable impacts on the field of cognition. When cognitive science restricts its viewpoint to ‘brains and above’, it at best underestimates the diversity of strategies available, and at worst may remain blind to the real underlying mechanisms of cognition.

The elusive demonstration of associative learning in a non-neural organism is still a passionately sought goal of many labs around the world, with Physarum researchers only the most recent to join the race. The Physarum model system, due largely to its ease-of-use and robustness to experimentation, has indeed been consistently successful at puncturing previously held cognitive prejudices. This success certainly favours Physarum as a non-neural system that has the potential to demonstrate associative learning, but this is far from a foregone conclusion. Smith-Ferguson and colleagues (2022) conclude—probably correctly, given the bias against publication of negative results—that a number of attempts to show associative learning in a wide range of organisms have probably failed. If so, that raises a question: why? If information processing or even cognition are ubiquitous among taxa, why not associative learning? They argue that Physarum simply has no strong selective pressure necessitating any kind of learning mechanism more complex than habituation. However, they concede that Physarum’s inability to make associations has not been thoroughly tested.

If definitive proof of this ‘gold standard of the cognitive’ were to be found in Physarum, what would that mean for the field of cognition? It would certainly cement the ascribed importance of taking a more holistic approach to cognitive science across taxa, and would spur vigorous research interest into the mechanisms of learning in ‘non-neuralia’. However, it could just as likely be viewed as just another neat trick pulled by Physarum and contribute little to the trajectory of animal cognition research into the future. This is especially likely if cognition researchers relegate Physarum behavioural research to a separate, irrelevant domain through the clever use of definitions. Physarum behaviour has been classed as basal cognition (Lyon et al. 2021), embodied cognition (Cheng 2022), extended cognition (Sims and Kiverstein 2022), and minimal cognition (see Vallverdú et al. (2018) for a list of references demonstrating each of the ten biogenic principles of minimal cognition from Lyon (2006), across a wide swathe of Physarum research). While it is conventional and often necessary to strictly define domains of research, definitions can also be misused to exclude entire fields that do not sit well with the established narrative (equating to a ‘No true Scotsman’ fallacy). For this reason, Smith-Ferguson and Beekman (2020) use Physarum behavioural research to argue against using the term cognition at all, “in favour of discussing various forms of information processing”.

In the recent 25th Anniversary retrospective edition of Trends in Cognitive Sciences, the question was raised: “What would make cognitive science more useful?” (Lewis Jr 2022). Lewis identifies as a major hurdle the “large discrepancy between the homogeneous samples that our field studies and the diverse populations that exist in the broader world—discrepancies that distort our understanding of how minds work and why they work in the ways that they do.” This discrepancy has been known for decades (see the famous study by Henrich et al. (2010) on the disproportionate use of WEIRD samples in human psychology). While Lewis’ target was human populations and the human mind, his statement equally applies to all organisms and all minds. Recent research on Physarum and other ‘basal’ organisms has shown us that true understanding of how minds work and why requires greater understanding of the diversity of minds that exist in the broader natural world.

The author would like to thank the editors of this special issue on basal cognition, Pamela Lyon and Ken Cheng, for the opportunity to present this review, and two anonymous reviewers for comments that improved the manuscript. The author was supported by grants from the Australian Research Council (DE190101513 and FT220100669).

Open Access funding enabled and organized by CAUL and its Member Institutions. No funding was received to undertake this work.