The cognitive cell: bacterial behavior reconsidered

Finally, swarms are different from colonies. Swarming motility depends on growth—it ceases when growth ceases—but swarms expand mainly by movement, not growth (Kaiser and Warrick, 2011). Although cells in a swarm are in all stages of the growth cycle, only 10% of swarm expansion is shown to be due to growth, whereas in a colony 90% of expansion is growth-related (Kaiser and Warrick, 2011).

The remainder of the section, which is derived entirely from Kaiser and Warrick (2014), describes what time-lapse microscopy of cells with fluorescent-labeled proteins revealed about what goes on in myxococcal swarms.

First and foremost, swarming M. xanthus don’t just move, they build two types of structures, the function of which are unclear: planar rafts, which have flat tops, and rounded mounds, which have five layers. Periodically, the cells on the top layer of the mounds translocate in an explosive, synchronized descent to the peripheries of the fourth layer. The explosive descent appears to be the result of a synchronized and simultaneous inhibition of all motility engines, A and S. From their new position the translocated cells slowly and asynchronously begin the climb to the top again, using pili-driven S-motility. When they have reached the top layer, the cells align into a compact arrangement as a result of A-motility signal proteins. The entire process takes about 14 min.

The mechanism by which this occurs is unknown, but what appears to take place is this. At low cell density, the FACs assemble on the sides of the leading half of the cells. The FACs then vanish and relocate to the trailing half, which upon cell reversal becomes the leading half. At high cell density, cells transiently attach to one another, side-by-side, as they move in the opposite direction along their long axes, bringing the FACs into alignment.

Kaiser and Warrick (2014) hypothesize that pairing facilitates passage of a signal between cells involving CglB and other A-motility proteins (as per Nudleman et al., 2005; Pathak et al., 2012). Supporting earlier observations of a motility complex, CglB proteins are found in all of the compartment spaces of the bacterium from the outer member to the periplasm and peptoglycan succulus to the cytoplasm, where the Frz-cluster pacemakers are located (Kaiser and Warrick, 2014). If the mechanism operates similarly to Wnt signaling in eukaryotes, as the researchers propose, the four levels of layering could allow Cglb proteins to pass the signal sequentially from one compartment space to another, providing a high degree of specificity between the upstream and downstream binding partners as well as with the intracellular compartment.

The signaling synchronizes the Frz-cluster pacemakers in the cell pair, which brings them into phase, effectively ‘resetting’ both cells to the average value of their pre-connection phases, a mechanism reminiscent of the scaling of synaptic firing potentials in neurons to ensure that overstimulation of one synapse does not distort the functioning of the network (Turrigiano, 2008).

Pairing is followed by cell reversal, which allows the cells to separate and move in a new direction, thus enabling the signal (whatever it is) to be spread, and the cells to become synchronized, which is necessary for coordinating the explosive descent. Thus, as Kaiser and Warrick (2011, p. 1309) observe: “Every cell in a mound should be available for pairwise signaling, cells in each layer are in contact with one another, and cells in a mound are able to move from one layer to another.” CglB is critical to this dynamic process of construction: CglB mutants can swarm but they cannot build.

In short, the sensorimotor activity of M. xanthus is astonishingly complex, and understanding this behavior may throw light on the sensorimotor behavior of social animals, such as birds, fish, and insects.

Traditionally it was thought that, however, complex their signaling machinery, all bacteria can do is sense and respond to ensure that existential parameters stay within the range of tolerance for survival and reproduction (Freddolino and Tavazoie, 2012). A growing body of evidence in the past decade suggests that bacteria are not simply backward-looking responders to stimuli already past—an assumption long-discredited in animals (Cahill et al., 2001)—but are dynamic predictors actively oriented toward what comes next (Tagkopoulos et al., 2008; Mitchell et al., 2009; Goo et al., 2012).

In all animals and plants (Trewavas, 2014), the ability to predict is predicated on memory and learning, which can take place within the life of an organism or across generations through epigenetic inheritance (Jablonka and Raz, 2009), very likely the basis of many evolved traits. Circadian entrainment of physiological processes to the planet’s most regular large-scale change—the daily cycle of light and dark—is a well-known class of evolved eukaryotic mechanisms that displays this anticipatory property. Thus far the only prokaryotes found to possess a proper circadian clock are photosynthetic cyanobacteria (Ouyang et al., 1998; Dvornyk et al., 2003), in which the phosphorylation state of kai proteins, not genetic transcription and translation, constitute the memory of elapsed time (Nakajima et al., 2005). The hemi-methylated state of DNA just after replication is an example of a genetic substrate for memory4. The key question, then, is whether and how bacteria remember and learn, and to what extent these capacities are relevantly like memory and learning in more complex animals.

Memory, typically the capacity to retain information acquired during the organism’s lifetime for a period sufficient to affect behavior, is part of the bedrock of cognition (Hilgard, 1980). For centuries memory, especially in its long-term modes, has been intrinsic to the anglophone concept of ‘mind’ (Earle, 1881). As stated earlier, the idea that bacteria exhibit some form of memory, or ‘memory effects,’ is now four decades old (Macnab and Koshland, 1972; Adler and Tso, 1974; Brown and Berg, 1974).

In his seminal explication of the RR model, Koshland (1977) describes the duration of protein phosphorylation as a bacterium’s memory because the period of modification effectively holds information about the state of the environment at a particular time, and the capacity to retain information, however, fleeting, is “no less real or useful to the bacterium than the memory of humans is to their behavior.” On the model proposed by Koshland (1980a), bacterial memory is “encoded in the enzymes that form and remove the RR.” Importantly from the standpoint of comparison with animals, the temporal duration of bacterial memory is not fixed, but is constituted (in CT proteins at least) by the difference between “a fast excitation process and a slow adaptation process,” which depends crucially on the strength and salience of the stimulus and the pathway involved (Morimoto and Koshland, 1991). The duration of response memory thus varies greatly among different chemotactic proteins, from seconds to more than 5 min (Yamamoto et al., 2005), the observational limit in the cited study. Moreover, chemotactic proteins may have multiple modifiable sites, suggesting the possibility of signalling and memory at different time-scales even within a single pathway (DeFranco and Koshland, 1980).

Morimoto and Koshland (1991) proposed a physiological distinction between short-term and long-term cellular memory, using the RR concept as a memory parameter. In the model short-term memory occurs when “stimuli evoke a transient elevation of the [amount of the] RR” in the cell, which “gradually returns to its initial value” when the stimulus is removed. Long-term memory occurs when “presentation of a second stimulus occurs before the initial response has returned to the baseline,” and “[s]uch stimulation results in a successive increase in the RR, which takes longer to return to baseline” (p. 2061). In Kandel’s work with Aplysia, the neuronal comparison that inspired Koshland, small-molecule neurotransmitters act rapidly because they are degraded and removed from the synaptic cleft by neuronal reuptake mechanisms, whereas neuropeptides passively diffuse away, and have longer-lasting effects (Kandel and Abel, 1995).

While protein dynamics are clearly important in memory processes in vertebrates and invertebrates, long-term memory in mammals, including humans, requires stable changes in neuronal structure. DNA methylation is now recognized as the mechanism by which such changes occur (Day and Sweatt, 2010). DNA methylation in general and chromatin remodeling in particular are also the mechanisms by which epigenetic learning takes place in vertebrates and invertebrates, allowing an organism’s experience in one generation to be incorporated into the adaptive repertoire of subsequent generations (Jablonka and Raz, 2009), particularly in relation to stress (Franklin et al., 2010). Epigenetic inheritance has been shown to modify stress responsiveness of bacteria (Veening et al., 2008; Chen et al., 2014).

Significantly, because stressful circumstances require immediate remedial action in a way attractive circumstances do not, in humans and other mammals memories for adverse events tend to be much stronger than those for positive events (Baumeister et al., 2001), which is why aversive conditioning is such a powerful tool.

In humans short-term and ‘working’ memory ranges from seconds to minutes, long-term memory from hours to decades (Terry, 2005). Long-term memory is often indistinguishable from non-associative learning, a type of learning in which “presentation of a particular stimulus alters the strength or probability of a response according to the strength and temporal spacing of the stimulus” (Rosenzweig et al., 1996, p. G-17). Non-associative learning includes sensitization, the amplification of a response following presentation of a stimulus, and habituation, the attenuation or extinction of a response to a stimulus upon repeated presentations (Shettleworth, 1998). Habituation and sensitization have both been demonstrated in bacterial CT (Koshland et al., 1982; Stock, 1999; Porter et al., 2011)—a discovery that “gave some neurophysiologists apoplexy, because they believed that a nervous system” was required (Taylor, 2004, p. 3761).

The operation of artificial and real neural networks depends on memory and non-associative learning effects, and the idea began to circulate that the ST systems of a bacterial cell might function as a neural network (Bray, 1990, 1995, 2009; Hellingwerf et al., 1995). Signal amplification is required for learning by neural networks, so a group of Dutch researchers decided to test the idea that autoamplification of genes in certain 2CSs, which further stimulates ST component production, might result in learning effects (Hoffer et al., 2001). Memory storage in animals with nervous systems was long thought to involve a mechanism involving an autophosphorylating protein kinase and paired phosphatase that operate together as a bistable switch (Lisman, 1985).

The pho regulon of E. coli, which operates to detect the metabolically critical nutrient Pi through the canonical PhoR/PhoB 2CS, was selected as the signaling pathway for experimentation (Hoffer et al., 2001). Upon stimulation the histidine kinase PhoR autophosphorylates and transfers phosphoryl groups to the cognate RR PhoR, which targets the pho operon, which then amplifies production of the two components. Thus signaling through PhoR/PhoB boosts production of the transduction components.

Cells from an exponentially growing population of phoA mutants were incubated first in a Pi limited medium (for 45 min at 42°), then in a high Pi medium (for 1 h at 30°). Later, the cells were transferred to Pi limited medium and incubated at 30°. The speed at which alkaline phosphatase production was induced, enabling the cell to scavenge traces of Pi or phosphorylated compounds from the environment, was then measured.

As hypothesized, the cells previously incubated in the Pi limited medium responded to the new limiting conditions faster than the control cells incubated exclusively in the high Pi medium. The faster response time correlated with the accumulation of ST components, a good probability that the response was the result of operon autoamplification rather than “a consequence of unspecific physiological effects.” Moreover, the learning behavior was “mechanistically and effectively different from the adaptation effects observed in CT,” and appeared to resemble immune system learning (Hoffer et al., 2001).

Another series of experiments involving B. subtilis are similarly indicative of non-associative learning. Wolf et al. (2008) selected B. subtilis for its sensitivity to environmental conditions and the well-known mechanisms governing sporulation. These mechanisms, triggered as part of a stress response, exhibit switch-like bistability, the basis of memory effects in computers, some physical compounds (i.e., magnetized iron), and neuronal activity in animals. Bistability is believed to have a role in human memory, particularly working memory (Durstewitz et al., 2000), but it has also been implicated in memory-associated processes in cellular plasticity and differentiation, e.g., in bone marrow (Wang et al., 2009).

Sporulation is an irreversible, energetically costly process of morphological and functional differentiation initiated as a last resort under starvation conditions (Dworkin and Losick, 2005) or a ‘bet-hedging’ strategy when nutrient supplies deteriorate (de Jong et al., 2011). The developmental sequence unfolds over 8 h, and results in the formation of specialized cells resistant to heat, chemical and UV stress. This involves the death of a significant proportion of the population, so the existential ramifications of movement toward sporulation are profound (for the survival of individual cells, at least), even before commitment becomes irreversible (Gonzalez-Pastor et al., 2003).

Three different strains of B. subtilis were grown under 10 different conditions, which varied according to the nutrient richness of the growth medium (rich and less-rich) and the density to which the population was allowed to grow, of which there were five classes (Wolf et al., 2008). In addition to a wild-type control, two fluorescent ‘reporter’ strains were created to measure three variables: (1) commitment to sporulation, a bistable process indicated by SpoIIE expression; (2) synthesis of AprE, a degradative enzyme that breaks down food sources, functions as a nutrient locator in deprived conditions, and shares many regulators with sporulation, but has a different expression pattern and is not believed to be bistable; and (3) growth, an indicator of general fitness.

After the growth period, the 30 different cell populations were sorted into identical densities and exposed to a common stressor, a starvation salt medium known to induce sporulation. Every 15 min for 24 h the 30 samples were measured in relation to growth, AprE synthesis, and commitment to sporulation (SpoIIE expression). Although Wolf et al. (2008) defined short-term memory effects as those measured within the first 11 h following exposure to the stress medium, in humans this timescale already would be said to involve long-term memory.

The study found that differences in life history gave rise to differences of varying strength in patterns of behavior relative to three variables (SpoIIE expression, AprE synthesis and growth) following a common stressor. Some of these effects lasted up to 36 h, far beyond the measurement period, which could either indicate ‘lifelong’ memory or the beginning of epigenetic memory across generations.

The data also suggest that memory strength and persistence are highly correlated with the energetics of survival. Growth medium was a better predictor of subsequent behavior than the density to which the original population was grown. Cells grown on the less-rich medium prior to exposure to the starvation salts pursued a more aggressive strategy. On average they grew “very fast” after exposure, induced AprE synthesis “after a brief lag,” and continued to grow even as the sporulation machinery switched on. By contrast, cells grown in the richer, Luria broth medium “seem[ed] to adopt a wait-and-see strategy,” did not grow and delayed sporulation and AprE synthesis “for many hours.”

As noted earlier, the finding that sporulation dynamics triggered by worsening circumstances exhibited “the most memory” (Wolf et al., 2008), despite involving the most energetically costly and far-reaching changes in the population, is consistent with results in human psychology and animal behavior (Lyon, 2014). Similarly, recent studies demonstrating anticipatory behavior in bacteria involve responses to predictable stressors.

The ability to anticipate and pre-emptively respond to regular changes in the environment confers a considerable fitness advantage, and has been observed in bacteria quite apart from circadian periodicity. Fruiting body formation in M. xanthus in the presence of nutrients from prey displays a predictive quality (Berleman and Kirby, 2007). Fruiting body formation was long assumed to be a response to the stimulus of starvation, because that is how it is commonly induced for experiment, but Berleman and Kirby (2007) found that a mere ‘step down’ in nutrients, not depletion, is sufficient to initiate the process, which ceases if nutrient availability increases. Nevertheless, only recently has anticipatory behavior been explicitly tested in bacteria using as cues predictable environmental changes across the life cycle.

Escherichia coli inhabits several ecological niches during its life cycle, from water, sediment and soil to the mammalian gastrointestinal tract (Mitchell et al., 2009). To determine whether E. coli signaling networks are capable of predictive behavior “in a fashion similar to metazoan nervous systems,” Tagkopoulos et al. (2008, p. 1313) tested strains under conditions mimicking the transition from the outside world to the gastrointestinal tract of a mammalian host. As they enter the oral cavity cells immediately experience rising temperatures up to 37°. As they transition to the gut available oxygen drops precipitously to anaerobic conditions. If the homeostatic (sense-respond) framework is correct, E. coli should not repress respiration until a drop in oxygen is detected. On the other hand, if enteric bacteria are capable of dynamic predictive behavior, rising temperatures should induce respiratory repression.

This is precisely what the studies showed. Exposure to temperature upshift, from 25° to 37°, not only induced the heat shock response regulon, but also strongly repressed genes encoding components of the molecular machinery for aerobic respiration, rapidly reprogramming to anaerobic mode. Similarly, a downshift in temperature (mimicking the organism’s excretion from the host) initiated the return to aerobic respiration. The two responses are so tightly coupled that strains were evolved in the laboratory to reverse the sequence of expression; a downshift in oxygen activated the heat shock response in the evolved strain (Tagkopoulos et al., 2008).