Reconsidering the evidence for learning in single cells

The question of whether single cells can learn led to much debate in the early 20th century. The view prevailed that they were capable of non-associative learning but not of associative learning, such as Pavlovian conditioning. Experiments indicating the contrary were considered either non-reproducible or subject to more acceptable interpretations. Recent developments suggest that the time is right to reconsider this consensus. We exhume the experiments of Beatrice Gelber on Pavlovian conditioning in the ciliate Paramecium aurelia, and suggest that criticisms of her findings can now be reinterpreted. Gelber was a remarkable scientist whose absence from the historical record testifies to the prevailing orthodoxy that single cells cannot learn. Her work, and more recent studies, suggest that such learning may be evolutionarily more widespread and fundamental to life than previously thought and we discuss the implications for different aspects of biology.

The emergence of learning was a major event in evolutionary history, allowing organisms to adapt to their environment on time scales much faster than genetic selection. But when exactly did learning emerge, and in what form? The vast majority of research on learning has been conducted in multicellular organisms, mostly vertebrates with complex nervous systems. This focus has led to biological models of learning based on synaptic plasticity (Martin et al., 2000). A plastic synapse, also known as a Hebbian synapse, changes its conductance in response to the close temporal pairing of pre- and post-synaptic spiking. In spike-timing-dependent plasticity, this temporal pairing is measured in milliseconds. A change in synaptic conductance is a change in the magnitude and sign (excitatory/inhibitory) of the postsynaptic currents produced by a presynaptic spike. In connectionist models of learning and memory, synaptic conductances are represented by signed connection weights with a scalar effect on the postsynaptic signal.

However, unicellular organisms capable of complex behaviors (see Dexter et al., 2019; Jennings, 1906) existed for over a billion years before the appearance of multicellular organisms (Wright and Lynn, 1997). This raises the question of whether unicellular organisms are also capable of learning, despite lacking nervous systems, as an evolutionary solution to the information processing challenges common to all living systems. The answer to this question has the potential to profoundly reshape our understanding of learning in multicellular organisms, as we discuss below.

To avoid ambiguity, we begin by fixing some terminology. Broadly construed, learning refers to any persistent and adaptive modification of an organism’s behavior as a function of its experience (see Barron et al., 2015, for further discussion). This behaviorally focused definition avoids reference to underlying mechanism at the cost of confounding a variety of processes (e.g. maturation, fluctuating motivational states) that we would not wish to include (Domjan, 2014). For that reason, it is common to include an appeal to an hypothesized mechanism. In empiricist philosophy, behaviorist psychology and connectionist cognitive science, the hypothesized mechanism is the formation of associations (philosophy and psychology), connections (connectionism) and alterations in the strengths of plastic synapses (neuroscience). The information-processing tradition in cognitive science, sometimes called the computational theory of mind, by contrast, conceives of learning as the extraction of information from sensory experience and the storage of that information in a memory (Gallistel and King, 2011). The memory is the channel by which information extracted from present experience is communicated to the computational operations that inform behavior in the indefinite future. The difference in the two theoretical frameworks can be summarized as saying that the first approach focuses on the acquisition of information and the second approach focuses on the information itself.

In this article, we will focus on a canonical form of learning, Pavlovian conditioning (Pavlov, 1927), while acknowledging the evidence for other forms of learning in unicellular organisms (e.g. habituation; Wood, 1969; Tang and Marshall, 2018). In Pavlovian conditioning, an initially neutral stimulus (the conditioned stimulus, or CS) comes to elicit a conditioned response (CR) when it reliably predicts the occurrence of a motivationally attractive or aversive stimulus (the unconditioned stimulus, or US). Despite its apparent simplicity, Pavlovian conditioning is a sophisticated form of learning that eludes a simple characterization in terms of association formation between stimuli (Rescorla, 1988; Gallistel and Gibbon, 2002; Gershman et al., 2010), which is the standard textbook story. It is an open question whether unicellular organisms exhibit a comparable level of sophistication in their learning capabilities. We note that the question of learning, and especially of Pavlovian conditioning, in non-neural organisms continues to elicit both interest and controversy (Eisenstein, 1975; Reid et al., 2016; Markel, 2020a; Gagliano et al., 2020; Gagliano et al., 2016; Markel, 2020b; Baluška and Levin, 2016).

Pavlovian conditioning is particularly interesting for our purposes because of the prevailing theory that it is mediated by synaptic plasticity. This theory has been criticized on several grounds: that synaptic plasticity cannot account for the behavioral features of Pavlovian conditioning (Gallistel and Matzel, 2013); that it is too unstable to implement long-term memory storage due to molecular turnover (Crick, 1984; Mongillo et al., 2017); that it can be experimentally dissociated from behavioral measures of memory (Chen et al., 2014; Ryan et al., 2015); and that behaviorally relevant information is not stored in a readable format (Gallistel, 2017). Furthermore, a synaptic memory substrate requires that computations operate via the propagation of spiking activity, incurring an energetic cost roughly 13 orders of magnitude greater than the cost incurred if the computations are implemented using intracellular molecules (Gallistel, 2017).

As an alternative (or complement), it has been proposed that memory may be stored using a cell-intrinsic substrate, such as polynucleotide sequences (e.g. RNA), post-translational histone modifications, or DNA methylation patterns (Landauer, 1964; Crick, 1984; Day and Sweatt, 2010; Abraham et al., 2019; Gräff and Tsai, 2013; Gallistel, 2017). These theories posit a mapping from experienced quantities, like the duration of the interval between the onset of the conditioned stimulus and the onset of the unconditioned stimulus, to the hypothesized changes in molecular level structures that encode them. The biochemical processes that produce these changes in response to relevant synaptic input and that later convert the encoded information to appropriately timed output signals remain unknown, although recent work revealing a CamKII-dependent tunable event timer provides a possible biochemical mechanism for interval timing (Thornquist et al., 2020).

The possibility of a cellular-level mechansim for storing acquired information with delayed behaivoral consequences is exciting from an evolutionary perspective because it suggests that the mechanisms for memory storage in complex multicellular organisms may have been inherited from much simpler organisms, possibly even protozoa, that share the same intracellular molecular repertoire. Synaptic plasticity is clearly not an option for protozoa, so if evolution hit upon a way to implement learning in these organisms, it is natural to conjecture that such a mechanism would be conserved across phyla, given its computational and energetic advantages. A convincing demonstration of Pavlovian conditioning in a unicellular organism would render less surprising recent results that imply that the encoding of the CS-US interval in Pavlovian eyeblink conditioning is intrinsic to the cerebellar Purkinje cell (Johansson et al., 2014).

At present, this evolutionary argument is speculative, and comparative studies have often focused on animals (Ginsburg and Jablonka, 2010). Moreover, the question is complicated by the fact that cell-intrinsic mechanisms interact in complex ways with synaptic mechanisms in multicellular organisms. The most well-established theory is that cell-intrinsic mechanisms, such as histone modification and DNA methylation, serve an enabling rather than a storage function (Kandel, 2001; Lisman et al., 2018). By regulating gene expression, they enable the synthesis of synaptic plasticity-related proteins. This type of theory is clearly not applicable to memory storage in single cells, which lack synapses.

Before we can begin addressing the molecular storage mechanisms in single cells, we must address a prerequisite question: do single cells learn? And if so, what exactly are they capable of learning? This question has been tangled in controversy since the early 20th century. In this paper, we chart the history of past attempts to address the single cell learning question, focusing in particular on the remarkable, and largely forgotten, studies of Beatrice Gelber. At the time of their publication, these studies were criticized for failing to rule out various confounding explanations. We revisit these criticisms, arguing that some of them are misplaced, while some of them can be reinterpreted in a more positive light. In our concluding remarks, we discuss some broader implications for the cell biology of learning.

At the turn of the 20th century, the study of protozoa was given a jolt with the publication of the book Behavior of the Lower Organisms by the zoologist Herbert Spencer Jennings. At that time, an influential view, championed by the physiologist Jacques Loeb (see Pauly, 1981), held that protozoan behavior was driven by various ‘tropisms’ (heliotropism, galvanotropism, chemotropism, geotropism, etc.). The concept of a tropism was inherited from the study of movement in plants, where it was observed that some plants reoriented leaf surfaces such that the incident light was symmetric (a form of heliotropism). Loeb and his students applied this concept more broadly to algae and animals, arguing that many movements could be conceptualized as reactions to an asymmetrically impinging force (e.g. light, electrical currents, chemical gradients, gravity), resulting in either approach or avoidance. Like Loeb, Jennings believed that there were deep connections between patterns of animal behavior and those of simpler organisms. However, they understood the nature of these connections, and the nature of behavior itself, quite differently. Unlike Loeb (who was essentially a behaviorist in his sympathies), Jennings sought mechanistic ‘sensorimotor’ explanations of behavior, on the belief that the mechanisms in animals were an elaborated form of similar mechanisms found in simpler organisms.

Most importantly for our purposes, Jennings, 1906 argued that one of the relevant mechanisms underlying protozoan behavior was learning. Because tropisms were typically conceptualized as unlearned stimulus-response mappings, the learning hypothesis was problematic for Loeb’s tropism-based theory of protozoan behavior. For example, Jennings observed that repeated aversive stimulation of the ciliate Stentor roeseli resulted in a characteristic sequence of distinct avoidance behaviors (resting, bending away, ciliary alteration, contraction, and detachment). This sequence could be interpreted as an elementary form of learning, in that the same stimulus came to elicit a different response. Jennings recognized that this implied a change internal to the organism. The change, he argued, was adaptive, a key feature of learning in ‘higher’ organisms:

"The essential point seems to be that after experience the organism reacts in a more effective way than before. The change is regulatory, not merely haphazard." (Jennings, 1906, p. 178).

On this score, Jennings took pains to rule out the hypothesis that the sequence reflected some form of fatigue.

Jennings’ study of elementary learning was largely discredited among the experts because of claims of non-reproducibility (Reynierse and Walsh, 1967). However, this was likely the result of experimental failings, most glaringly the fact that Reynierse and Walsh used a different species of Stentor, and recent experimental work with the correct species has vindicated Jennings (Dexter et al., 2019). The reception of Jennings’ work on S. roeseli illustrates the prevailing tendency to discount complex behaviour in unicellular organisms, to the point of being less critical of experiments which give the expected result. It is also a reminder that even if learning is phylogenetically widespread it still needs to be studied in the ecological context of a particular organism: what one species of Stentor learns may be irrelevant to another.

Jennings arrived at some of the same ideas about associative learning that were concurrently being investigated by Ivan Pavlov in dogs, in particular the idea that the response to an aversive or appetitive stimulus could be ‘transferred’ to a neutral stimulus by virtue of an association between the two stimuli. Nonetheless, Jennings had little evidence for this claim. It was left to later scientists to fill the gap.

Day and Bentley, 1911 placed Paramecia (another ciliate) into a capillary tube, observing that the Paramecia turned and eventually reversed direction. With repeated trials, the Paramecia increased the speed with which they achieved reversal (see also Smith, 1908). Inspired by Day and Bentley, French, 1940 used a larger capillary tube, allowing easier turning, and found that over the course of trials the Paramecia became faster at escaping from the tube by swimming downwards. This pattern has been interpreted in terms of associative learning (e.g. Huber et al., 1974), based on the hypothesis that contact with the tube walls is aversive, so that the Paramecium forms an association between its swimming actions and relief from the aversive stimulus. However, Hinkle and Wood, 1994, using Stentor as their subjects, presented evidence that the learning effect may be confounded with the orientation of the tube: the subjects increased the frequency of downward swimming even when the escape was from the upper end of the tube. More recently, Armus et al., 2006 showed that Paramecia learned to associate the location of cathodal stimulation (which exerted an attractive effect) with a light-conditioned stimulus. Other studies used aversive unconditioned stimuli such as shock or heat (Bramstedt, 1935; Soest, 1937; Hennessey et al., 1979), but the results were inconsistent (Best, 1954; Mirsky and Katz, 1958) and beset by alternative interpretations, such as changes in the chemical composition of the medium (Grabowski, 1939).

Taken together, these studies leave an equivocal picture of single-cell learning capabilities. Moreover, even those studies that successfully demonstrated learning often did not systematically investigate the parameters governing the strength of the effect, its retention, or any hypothesis about the underlying mechanism.

We next turn to one of the most systematic programs of research on learning in Paramecia—one which ignited considerable controversy during its heyday, but was eventually consigned to the scientific waste-bin. The remarkable scientist who single-handedly carried out this program, Beatrice Gelber has since been largely forgotten. The sole photograph we have been able to trace of her comes from a news story (Figure 1). Our goal is not only to rehabilitate her scientific legacy, but also to amplify some of her visionary ideas about the molecular biology of memory.

In the late 1940s, Beatrice Gelber was a divorced mother of 3 living with her grandmother in Long Beach, NY (Census data for 1930–1940). After her youngest child departed for college, she decided to follow her curiosity and enrolled in the graduate Psychology program at Indiana University. Her advisor was Roland Clark Davis, a leading proponent of ‘psychophysiology’, which sought to establish links between physiological and psychological phenomena. Davis primarily conducted electrophysiological experiments on humans (he was a pioneer in the study of the galvanic skin response), but just down the hall B.F. Skinner was conducting experiments on learning in pigeons. This made for a rowdy work environment:

"[T]he lab had a distinctive odor, which came from the multichannel recorder, an early inkless polygraph. Electrically activated metal pens—old phonograph needles through which a sizable current was passed during recording—traced the signal onto thin tar paper that was coated with gunpowder. As the paper flew through the recorder at 50 mm/s , the powder exploded off, leaving a fine black tar recording—and the room filled with smoke… [T]he acrid smell of the recording paper mixed with the stench of clucking pigeons." (Gabbay and Stern, 2012, p. 445).

Amidst this cacophony, while simultaneously pursuing her thesis research on muscular contraction, Gelber somehow found time to train Paramecia. In this effort, she benefitted from interactions with Tracy Sonneborn’s laboratory in the Department of Biology. Sonneborn was trained by Jennings in the study of ciliates, specializing in the study of Paramecia and becoming one of the leading geneticists of the era before molecular biology (Sapp, 1987). He made pioneering contributions to the study of non-Mendelian forms of inheritance (Preer, 1996). In addition to Skinner and Sonneborn, Indiana University was inhabited at that time by a number of other important figures in biology and psychology: Salvador Luria (whose collaboration with Max Delbrück, showing that bacteria and phage had genes, earned them Nobel Prizes), Luria’s first graduate student James Watson (co-discoverer of the structure of DNA), and William Estes (a former Skinner student who became faculty, making important contributions to learning theory). It is conspicuous that Gelber’s interests occupied the epicenter of this intellectual cauldron. Yet she appeared to be sui generis. Davis’ son, Chris Davis, who followed his father into psychophysiology, worked with Gelber as an undergraduate student and recalls her being a demanding experimentalist as well as a pleasant, if formal, supervisor, who occupied a place within the department that was apart from both her fellow students and the faculty (personal communication). Gelber seems to have had considerable independence in her research while still a PhD student, reflecting perhaps both her own maturity and her supervisor’s encouraging mentorship (Gabbay and Stern, 2012).

Taking advantage of the fact that Paramecia feed upon bacteria, Gelber asked the following question: would Paramecia learn to approach a wire that was repeatedly coated with bacteria? In her first experiment, Gelber, 1952 compared one group of Paramecia who received reinforced training with another group that received no training (Figure 2, top). Each reinforced trial consisted of swabbing the wire with a bacterial suspension and then dipping it into the Paramecia culture for 15 s. Both groups were given preliminary and final tests, which consisted of dipping the wire without any bacteria. Gelber reported that the reinforced group clung to the wire in greater numbers on the test trials compared to the preliminary tests, and also compared to the no training group.

One problem with Gelber’s first experiment was that it did not control for repeated exposure to the wire itself as a driver of behavioral change. Accordingly, in her second experiment (Figure 2, bottom) she tested a non-reinforced training group that received wire-alone trials; this group did not show any evidence of conditioned responding. Thus, the combination of wire and bacteria was necessary to produce the learning effect.

It is worth noting here that Gelber was a very careful experimentalist. She knew that reorganization of nuclear material occurs during conjugation (sexual reproduction) and autogamy (self-fertilization) and was keen to avoid this because of her initial hypothesis that the memory trace (engram) was stored in the macronucleus (see Gelber and Rasch, 1956; Gelber, 1958). To accomplish her goal, she used a single mating type (eliminating conjugation) and small amounts of bacteria (suppressing autogamy). She also tracked the number of fissions, knowing that autogamy would occur after a certain number of fissions.

Having established the basic learning phenomenon, Gelber moved to the University of Chicago. She appears to have been a postdoctoral researcher in the Department of Psychology, judging by the affiliation given in her papers (Gelber, 1957), and again to have had substantial independence in formulating her own research. The Chicago Maroon describes her on 3 June 1955 as a NIH Postdoctoral Fellow in the Department of Physiology (which we presume to be a misprint for the Department of Psychology) and a recipient of NIH grant funding, in the same company as the psychologist Bruno Bettelheim and the biophysicist Aaron Novick.

Gelber proceeded to examine how a number of different factors influenced both learning and retention. In one of her only collaborative papers, she worked with Ellen Rasch to determine Paramecium DNA content and carefully analyzed the interplay between feeding, fission and autogamy (Gelber and Rasch, 1956). In her solo work, she first demonstrated that Paramecia would not approach the wire when tested in the dark (Gelber, 1956), indicating that they relied on photosensing to detect the conditioned stimulus in the absence of chemical gradients from the bacteria. Second, she demonstrated that Paramecia retained their memory 3 hr after spaced (long inter-trial interval) training but not after massed training (Gelber, 1958). This finding agrees with spacing effects observed in many animal species, ranging from bees (Menzel et al., 2001) to humans (Ebbinghaus, 1885; Cepeda et al., 2008). She acknowledged that this was a rather short retention interval compared to animal studies, but also noted that Paramecia can go through two reproductive cycles during this time, so it could be argued that 12 hr is a long period of time for a Paramecium. In later work, she showed that retention could last up to 12 hr (Gelber, 1962b), and other work using a different paradigm has shown retention up to 24 hr (Hennessey et al., 1979). Third, she observed a phenomenon known as reminiscence (Payne, 1987), where performance increases with moderate retention intervals before declining again, a more complex picture than the one classically given by monotonic forgetting functions. A recurring theme in these investigations was the parallel between learning in animals and Paramecia.

Gelber was repeatedly confronted by critics who denied that Paramecia were capable of sophisticated behavior. Donald Jensen, another psychologist working with Paramecia, argued that the increased rate of attachment to the wire as a consequence of training could be a form of ‘thigmotropism’ (attraction to a haptic stimulus) induced by the bacteria, possibly due to release of metabolic products like carbon dioxide or carbonic acid. He showed that addition of bacteria to a culture increased attachment to the bottom of the slide, but he did not demonstrate that bacteria increase attachment to the wire (which was not used at all in this experiment). Indeed, by the same logic we might expect thigmotropism to inhibit the learning effect, since the wire contacted the bottom of the depression slide, and therefore the Paramecia would be attracted to both the slide surface and the wire. In any case, Gelber, 1957 failed to replicate Jensen’s findings.

Jensen further argued that Gelber’s findings might be an artifact of changes to the culture induced by training, rather than any persistent change in the organism itself (Jensen, 1957a). In support of this claim, Jensen showed that samples taken from the vicinity of the wire contained large numbers of bacteria after training, in contrast to samples taken far away from the wire. In response, Gelber, 1957 pointed out that she obtained a learning effect despite routinely stirring the slide between trials, presumably diluting any residual concentration of bacteria around the wire. Jensen, 1957b retorted that this stirring method was not sufficient to disperse the bacteria. She also showed that inserting the wire with food after training produced a greater response than food alone, indicating that the wire was contributing something beyond the effect of the bacteria. Finally, she pointed out that Jensen deviated from her experimental procedures in a number of other ways, including using bacterial concentrations several magnitudes higher than she had used (a point brushed aside by Jensen without any justification), and introducing the bacteria into distilled water instead of culture fluid (“the amount of steak found in a dish placed on the floor in an empty room would be very different from the contents of the same dish with a hungry dog present.”). The difference in bacterial concentration may have resulted in a greater amount of residual bacteria after removal of the wire, thereby degrading the contingency between the wire and bacteria (see further discussion below). It is now understood, as it was not at the time of this controversy, that when the background frequency or concentration of US is as frequent as that in the presence of the CS, no conditioning occurs (Rescorla, 1988).

"The use of this analogy symbolizes what is perhaps the most basic difference of opinion between Gelber and me. Gelber freely applies to Protozoa concepts (reinforcement and approach response) and situations (food presentation) developed with higher metazoan animals. I feel that such application overestimates the sensory and motor capabilities of this organism… If analogies are necessary, a more apt one might be that of an earthworm which crawls and eats its way through the earth, blundering onto food-rich soil and avoiding light, heat, and dryness. Gelber’s assertion loses its force when the blind, filter-feeding mode of life of Paramecia is considered." (Jensen, 1957b, p. 1341).