Dr. Blackwell received her VMD and PhD from the University of Pennsylvania and is an associate professor in the School of Computational Sciences, and the Krasnow Institute for Advanced Studies at George Mason University. Her research examines the synaptic and ionic currents and second-messenger pathways involved in associative learning, using a combination of experimental and computational techniques.
Learning is “a relatively long lasting and adaptive change in behavior resulting from experience” (Hall,1976). The breadth of species that exhibit learning is astounding, ranging from worms (Rankin,2004) to humans. Nonetheless, the capacity for learning varies greatly among species, being exceedingly large in primates as compared to most other mammals (Lefebvre et al.,2004). The long developmental period of primates is coupled to learning how to survive in the environment; the expense of a long developmental period is repaid in the form of extreme flexibility and adaptability.
Memory is intricately tied to learning, because memory is the storage of what we have learned. Thus, learning is analogous to the procedure producing the change in behavior, but memory is the physical change to the organism that allows the behavior to endure. A fascinating aspect of learning and memory is that there is a dichotomy in the types of learning, which is reflected in our language. Explicit memory is the memory for facts and events; implicit memory is the memory for skills (Squire and Zola,1996). When we learn facts and events, we have it in memory, or have memorized the information. When we learn a skill, that too is stored in memory, but as a capability, or enhancement of skill performance.
Another intriguing aspect of learning and memory is that most of the time, we are learning associations. From the earliest age, an infant learns to associate the smell of its mother with being fed, thus the infant calms when the mother picks it up, even prior to being fed. Such language as “that reminds me of …” and “looks like rain,” as well as observations that students learn information better when it is related to what they already know, is additional evidence of the associative nature of learning and memory.
A breakthrough in the study of learning came with the development of an animal model of an associative form of learning called classical conditioning, developed by the Russian physiologist Pavlov (Windholz,1989). Originally, Pavlov was studying glandular secretions involved in digestion and was inducing dogs to salivate by giving them food. The delivery of food was preceded by a sound such as a bell. The dogs learned that the bell predicted the food and began salivating in response to the bell. This early salivation response actually interfered with the original experiments, but Pavlov was clever enough to realize that this was an adaptive or learned behavior useful for studying the brain.
Continuous study during the last century has helped to define and delineate classical conditioning (Hall,1976). In Pavlov's paradigm, the bell, which initially does not evoke a response, is the conditioned stimulus (CS); the meat, which elicits a response prior to training, is the unconditioned stimulus (US; see Table 1 for a list of abbreviations). After repeated presentations of the CS (sound) followed by the US (food), the sound alone evokes a response, called the conditioned response (CR). One of the hallmarks of classical conditioning is that the timing between the CS and US presentation is critical. If the CS is presented after the US, the animal will never learn that the CS predicts the US. Thus, the CS must be presented before the US, and the time between CS and US presentation should not be too long (though the limiting length depends on the species and other characteristics of CS and US).
Table 1. Abbreviations
Extra-cellular signal related protein kinase
Mitogen activated protein kinase
Protein Kinase A
Protein Kinase C
Central Pattern Generator
Inhibitory post synaptic potential
Excitatory post synaptic potential
In-depth studies by psychologists of factors constraining classical conditioning were designed to determine the physical changes corresponding to memory storage. In other words, classical conditioning was considered a learning paradigm to determine how the brain controls learning behavior. Behavioral techniques coupled with neurophysiology techniques such as extracellular recording of neuronal activity or focal brain lesions revealed that multiple neuronal circuits participate in this simple form of learning. One set of sensory neuronal circuits transforms external conditioned and unconditioned stimuli into neuronal activity patterns. Another set of motor neuronal circuits transforms neuronal activity patterns into behavior. A third distinct set of neuronal circuits stores the memory: the activity patterns of these neurons is transformed when CS and US stimuli are presented with the correct temporal sequence (Berger et al.,1983). The most important question remains: how do neurons store memories? In other words, what are the mechanisms whereby activity patterns of the “memory storage” neurons are transformed?
Invertebrate Classical Conditioning
A breakthrough for studying the neuronal basis of learning emerged when invertebrates such as Hermissenda crassicornis were shown to exhibit such behavior (Crow and Alkon,1978). Hermissenda is a small nudibranch mollusk, also known as a seaslug, that lives in the coastal waters of the Pacific Ocean, e.g., off the coast of Monterey, California. Hermissenda learns to associate light, the CS, with turbulence, the US. Prior to learning, naive animals move forward in response to light and contract their foot in response to turbulence; after learning, conditioned animals contract their foot in response to light and delay forward movement toward the light. It is important to point out that these are stimuli that appear in Hermissenda's natural environment; thus, light is a surrogate for sunlight, and turbulence is the surrogate for wave action. This behavior is adaptive in that during storms, when the water is extremely turbulent, Hermissenda learns to avoid moving toward the surface of the water and instead contracts its foot in order to cling to a rock or other structure.
Many of the behavioral characteristics of Hermissenda classical conditioning are similar to those found in mammalian classical conditioning (Crow and Alkon,1978; Schreurs and Alkon,2001). Hermissenda do not learn if presented with light alone, or turbulence alone. Learning requires that the light occurs prior to turbulence, and that turbulence follows within a few seconds. Other behavioral properties include contingency sensitivity (both Hermissenda and other animals learn more slowly if unpaired stimuli are interspersed with paired stimuli) and savings (Hermissenda that have previously learned and forgotten an association learn more quickly the second time). Other behavioral characteristics of classical conditioning, both those shared with mammals and higher-order mammalian characteristics lacking in Hermissenda, are reviewed elsewhere (Matzel et al.,1998; Crow,2004).
NEURONAL CORRELATES OF CLASSICAL CONDITIONING
For neuroscientists, classical conditioning is a paradigm used to probe deeper questions of how the brain stores memories. Thus, the unquestionable value in studying classical conditioning in Hermissenda is attributed to its simple nervous system and the ability to measure neuronal changes corresponding to the memory trace. The central nervous system consists of paired cerebropleural ganglia, pedal ganglia, eyes, and statocysts, the vestibular organs composed of 13 hair cells that sense acceleration (Fig. 1). Compared to the billions of neurons in the mammalian brain, the thousands of neurons in the Hermissenda brain and the stereotypical configuration of many of these neurons allows for precise investigation of neuronal properties and neuronal networks participating in memory storage and expression. Furthermore, it is possible to remove the entire central nervous system and apply the same paired stimuli of classical conditioning to this in vitro nervous system. Thus, the in vitro preparation allows scientists to probe the inner workings of neurons while memories are being stored.
Photoreceptors Comprise Sensory Neurons for CS
Each of the Hermissenda eyes has three type B and two type A photoreceptors (Fig. 2A) (Stensaas et al.,1969). The photoreceptors are sensory cells for the CS stimulus and transduce light energy into depolarization. The type A and type B photoreceptors exhibit characteristic differences reminiscent of the difference between mammalian rods and cones: type B photoreceptors are sensitive to dimmer lights than are type A photoreceptors (Alkon and Fuortes,1972; Mo and Blackwell,2003). The temporal response to light stimuli also differs (Fig. 3A). Type A photoreceptors respond to an increase in illumination with a rapid increase in firing frequency, followed by rapid light adaptation (Crow,1985; Farley et al.,1990; Yamoah et al.,1998; Mo and Blackwell,2003). When the light stimulus is removed, type A photoreceptors quickly cease firing, a process known as dark adaptation. In contrast, type B photoreceptors fire less strongly in response to an increase in illumination, and they also light-adapt less strongly then type A photoreceptors (Crow,1985; Farley et al.,1990; Mo and Blackwell,2003). Thus, after the initial light response, both type A and type B photoreceptors respond at near equal firing frequencies. When returned to darkness, type B photoreceptors continue to fire for prolonged periods due to their slow dark adaptation. These different response properties suggest different functional roles: type A photoreceptors signal changes in illumination, and type B photoreceptors signal background illumination.
Photoreceptors Play Additional Role as Memory Storage Neurons
Decades of work by Alkon and colleagues (Crow and Alkon,1978) have revealed that the photoreceptors are the first site of convergence of the CS and US stimuli and a key locus of memory storage. The sensory cells for the US are the statocyst hair cells, which transduce turbulence or gravitational energy into depolarization (Detwiler and Fuortes,1975). Statocyst hair cells release the inhibitory neurotransmitter GABA (Alkon et al.,1993) onto the terminal branches of the photoreceptors. Thus, during classical conditioning, paired CS and US stimuli produce depolarization and GABA receptor activation of the photoreceptors. When light and turbulence are presented in the correct temporal sequence, intracellular processes are activated and produce intrinsic changes to the photoreceptors. Thus, the memory of the association is stored as a change in response properties of these photoreceptors. Interestingly, type A and B photoreceptors are modulated differently by classical conditioning paradigms.
Intracellular recordings from animals that have been classically conditioned, or from applying a classical conditioning procedure to the in vitro nervous system, have revealed several changes in photoreceptor properties that are correlated with learning behavior, or the presentation of paired stimuli, respectively. Type B photoreceptors show an increase in excitability due to the conditioning procedure, as compared with type B photoreceptors, which received the unpaired training procedure (Fig. 3C and D). The increase in excitability is exhibited as an increase in input resistance, an increase in the firing frequency due to light stimulation or current injection (Crow and Alkon,1980; Matzel and Rogers,1993; Blackwell and Alkon,1999), and an enhanced long-lasting depolarization following light stimulation (Alkon and Grossman,1978). The increased excitability is accompanied by a reduction of the transient potassium current, the calcium-dependent potassium current, and the persistent calcium current (Alkon et al.,1985; Collin et al.,1988). In contrast, type A photoreceptors exhibit a decrease in excitability (Farley et al.,1990; Frysztak and Crow,1993). This includes a decrease in the generator potential, and a decrease in input resistance, which is accompanied by an increase in the delayed rectifier potassium current. The effect of classical conditioning on firing frequency of type A photoreceptors is less clear, with reports of increase (Frysztak and Crow,1993) and decrease (Farley et al.,1990; Farley and Han,1997).
The most exciting aspect of these discoveries is that they may represent general mechanisms of memory storage (Matzel et al.,1998; Schreurs and Alkon,2001; Daoudal and Debanne,2003). Long-term changes in excitability have been demonstrated in rabbits consequent to classical conditioning (Moyer et al.,1996; Schreurs et al.,1998). An accompanying reduction in afterhyperpolarization suggests that the increase in excitability is mediated by a reduction in potassium currents (Coulter et al.,1989; Schreurs et al.,1998). Additional evidence for the role of potassium currents comes from genetic studies of Drosophila. Alteration of shaker transient potassium channels or eag potassium channels produces deficits in conditioning (Cowan and Siegel,1986). Thus, though memory storage in sensory neurons may be unusual, the changes in potassium currents and excitability in neurons receiving convergent inputs is likely to be a general mechanism of memory storage.
Photoreceptor Synaptic Interactions Contribute to Memory Trace
Another mechanism underlying classical conditioning in Hermissenda is changes in synaptic interactions among photoreceptors. As illustrated in Figure 2A, the type B photoreceptors are mutually inhibitory, and the type A photoreceptors are inhibited by the type B photoreceptors (Fig. 2A) (Alkon and Fuortes,1972; Goh and Alkon,1984). Classical conditioning enhances the inhibition of medial type A photoreceptors by medial type B photoreceptors (the IPSP is enhanced), but there is no change in inhibition of lateral type A by lateral type B photoreceptors (Frysztak and Crow,1997). In contrast, in vitro conditioning produces an enhancement of IPSPs from type B to type A photoreceptors, both medial and lateral (Fig. 2B) (Schuman and Clark,1994). There have been no reports of changes in synaptic strength of inhibitory connections between pairs of type B photoreceptors, but the increased input resistance and firing frequency would produce enhanced inhibition regardless of whether the synapse itself was modulated. Furthermore, action potential width in type B photoreceptors is increased (Gandhi and Matzel,2000), which may be accompanied by an increase in calcium influx and consequent increase in release probability. In summary, classical conditioning potentiates synaptic interactions not only by synaptic mechanisms, but also secondary to the modulation of voltage-dependent channels and membrane properties.
Long-term synaptic potentiation, specifically that produced by synaptic mechanisms, has long been studied as a mechanism of memory storage in mammals (Bliss and Collingridge,1993; Malinow and Malenka,2002). Despite decades of study, the connection between memory and hippocampal LTP is tenuous (Teyler and DiScenna,1984; Shors and Matzel,1997). Perhaps the strongest evidence for synaptic plasticity as a mechanism of memory storage is observed in the basolateral nucleus of the amygdala. In vivo recordings in rats reveal synaptic potentiation after fear conditioning but not in unpaired controls (Blair et al.,2001). In mammals, addressing how synaptic plasticity produces a change in behavior is inaccessible due to the large number of neurons and synapses involved in producing behavior. In contrast, further studies of Hermissenda may reveal scalable mechanisms whereby changes in synaptic plasticity produce changes in activity patterns that control behavior.
Modulation of Ionic and Synaptic Channels Produce Change in Spatiotemporal Firing Patterns
How is modulation of ionic and synaptic channels translated into a change in behavior? A novel approach to this question involves considering the five photoreceptors as a small neuronal network (Fost and Clark,1996b). The light response of this network constitutes a spatiotemporal firing pattern that is shaped by inhibitory interactions among photoreceptors. Modulation of both synaptic and ionic currents transforms the spatiotemporal pattern in a nonlinear manner due to the inhibitory synaptic interactions. In this manner, the increase in variance of the interspike interval due to classical conditioning (Crow,1985) may be due to the nonlinear transformation from tonically firing to periodic burst firing (Fig. 3B) (Fig. 8 of Mo and Blackwell,2003).
Periodic burst firing patterns are seen in many systems, such as networks with mutually inhibitory interactions and sustained excitatory inputs (Popescu and Frost,2002), and networks of coupled excitation and inhibition (Warren et al.,1994). Both of these network configurations are present in the visual system of Hermissenda. Mutual inhibition with excitation is present in the three mutually inhibitory type B photoreceptors; tonic excitatory input is provided by the light-induced long-lasting depolarization (Alkon and Grossman,1978). Coupled excitation and inhibition is seen in connections between type B cells and optic ganglion cells: the photoreceptors inhibit optic ganglion cells, and optic ganglion cells excite type B photoreceptors (Alkon,1980a). Both strength and persistence of neuronal synaptic interactions influence the occurrence and characteristics of periodic activity (White et al.,2000). Thus, periodic bursting may develop after classical conditioning due to the increased strength of inhibitory interactions among type B photoreceptors.
Computational Models Investigate Relationship Between Ionic Channels and Neuronal Activity
The Hermissenda community has a strong tradition of computational modeling to integrate and synthesize results from myriad experiments into a coherent view of neuronal memory storage. The ultimate goal with such models is to demonstrate how paired stimuli produce changes in neuronal properties, neuronal activity, and motor output. An additional advantage of computational models is to explicate and evaluate the assumptions that always accompany conceptual models. Presently, computational models of Hermissenda photoreceptors have been used primarily to investigate two issues. The first is the second-messenger mechanisms underlying temporal sensitivity (discussed in the next section), and the second is the mechanisms (ionic currents) underlying the various changes in excitability and synaptic efficacy.
Several models were created to test hypotheses regarding the role of ionic currents in producing the enhanced generator potential, increased spike frequency, synaptic potentiation, and increase in input resistance. These models include voltage-dependent, calcium-dependent, and light-induced currents. The first model (Sakakibara et al.,1993) supported the hypothesis that a reduction in the transient and calcium-dependent potassium current produces an increased generator potential and enhanced long-lasting depolarization. Later models confirmed these results and in addition showed that the reduction in the calcium-dependent potassium current produced a greater influence on the plateau potential and light-induced firing frequency than did the transient potassium current (Fost and Clark,1996a). A reduction in the persistent calcium current and an enhancement in the hyperpolarization-activated current, two other ionic current changes correlated with classical conditioning, tended to oppose each other in terms of effect on plateau potential and firing frequency (Cai et al.,2003).
The enhancement of synaptic connections is more difficult to explain and may be produced by mechanisms distinct from the changes in voltage-dependent currents. The models above demonstrate that the reduction in the transient potassium current produces spike broadening (Fost and Clark,1996a; Flynn et al.,2003). Since neurotransmitter release depends on calcium influx, which depends on spike width, this result suggests that a reduction in the transient potassium current produces synaptic facilitation. Nonetheless, when the reduction in transient potassium current is coupled with a reduction in calcium current, not only is the enhanced calcium influx eliminated, but calcium influx is depressed below the control (preconditioned) value. Furthermore, an enhancement in the hyperpolarization activated current does not compensate for the reduced calcium influx, most likely due to inactivation of this current at depolarized potentials. Thus, these four current changes result in a reduction in calcium influx, suggesting that the synaptic facilitation is produced by a mechanism independent of spike broadening, such as modulation of postsynaptic channels or presynaptic release probability.
Computational Models Suggest a Role for Light-Induced Potassium Current
In all of these models, the enhanced long-lasting depolarization not only requires a reduction in the transient and calcium-dependent potassium currents, but also involves the light-induced potassium current. Specifically, the baseline long-lasting depolarization, observed in untrained Hermissenda after light offset (Alkon and Grossman,1978), is produced by the light-induced potassium current (Detwiler,1976; Alkon et al.,1984; Blackwell,2002b,2004), which has properties similar to leak potassium currents (Buckler,1999; Donnelly,1999). This baseline depolarization activates both the transient and calcium-dependent potassium currents, which are inactivated at resting potential due to their voltage dependence. A reduction in the transient and calcium-dependent potassium currents does not produce an enhanced depolarization in a model lacking the light-induced potassium current (Blackwell,2004).
The light-induced potassium current also may be required to explain the increase in input resistance measured with hyperpolarizing current injection. A decrease in the transient and calcium-dependent potassium currents is not sufficient because these currents are not active at resting potential. Similarly, an enhancement in the hyperpolarization-activated current, which is active at resting potential, actually decreases the input resistance at hyperpolarized potentials. One possibility, which needs to be evaluated experimentally, is that the light-induced potassium current is modified by classical conditioning. Since this current is normally open and conducting at rest, a reduction in this current (either a decrease in total conductance or a decrease in open probability or duration) would decrease total membrane conductance and increase input resistance at all membrane potentials (Fig. 3C and D).
SUBCELLULAR MECHANISMS FOR TEMPORAL SENSITIVITY
This research in Hermissenda, together with overwhelming evidence from other model systems (Matzel et al.,1998; Kandel and Pittenger,1999), demonstrates that long-term memory storage consists of changes in ionic and synaptic currents in individual neurons. This naturally leads to a more intriguing question: What are the subcellular events that lead to modulation of ionic currents when paired stimuli are presented in correct temporal sequence? In other words, what are the subcellular mechanisms that are sensitive to temporal pattern in the photoreceptors of Hermissenda? The requirement for paired stimuli in the behavioral paradigm naturally translates into a requirement for either two different intracellular messengers or two different sources of a single intracellular messenger molecule.
Elevation in Calcium Required for Memory Storage
In Hermissenda photoreceptors, one of the first intracellular messengers identified as meeting this requirement was calcium. Both light stimulation and depolarization produced elevations in calcium via voltage-dependent calcium channels (Connor and Alkon,1984) and release from intracellular stores (Berridge and Galione,1988; Muzzio et al.,1998). The rebound depolarization observed after relief of hair cell inhibition, along with the progressively enhanced depolarization and calcium elevation noted during in vitro training, led to the hypothesis that temporal sensitivity was attributed to the requirement for two sources of calcium: one in response to light, and the second due to depolarization (Alkon,1980b). This hypothesis was supported by the observation that a calcium elevation was required for the enhanced excitability of type B photoreceptors. Blocking the elevation in calcium (Matzel and Rogers,1993), or even just one of the calcium sources (Talk and Matzel,1996; Blackwell and Alkon,1999), prevented these correlates of memory storage.
Activation of PKC Is Involved in Memory Storage
Protein kinase C (PKC) was the second intracellular molecule identified as playing a critical role in classical conditioning. Classical conditioning produces a translocation of PKC from the cytosol to the membrane (Muzzio et al.,1997) and PKC phosphorylates the transient and calcium-dependent potassium channels, decreasing their maximum conductance and producing an increased input resistance and evoked spike frequency (Neary et al.,1981; Farley and Auerbach,1986). More importantly, PKC requires two or more intracellular messengers for activation: transient activation requires an elevation in calcium and diacylglycerol (Asaoka et al.,1988; Oancea and Meyer,1998), both of which are produced by light stimulation (Talk and Matzel,1996; Talk et al.,1997; Sakakibara et al.,1998). Persistent activation of PKC requires not only calcium and diacylglycerol, but also arachidonic acid (Lester et al.,1991; Shinomura et al.,1991), which is produced by phospholipase A2. Prevention of in vitro classical conditioning with an inhibitor of phospholipase A2 (Talk et al.,1997) supports the role of arachidonic acid in classical conditioning.
Since light alone produces transient activation of PKC (but does not produce classical conditioning), then turbulence-induced hair cell activity must somehow contribute to activation of PLA2 and persistent activation of PKC (Fig. 4). Either GABAB receptors are directly coupled to PLA2, or GABAB stimulation produces an elevation in calcium, which combines with the light-induced calcium, to produce a larger or more prolonged calcium elevation that activates PLA2 (Hirabayashi et al.,1999). The possibility that GABAB stimulation contributes to the calcium elevation in the soma (independent of the effect on membrane potential) was suggested by the observation of a calcium wave propagating from the terminal branches to the soma (Ito et al.,1994). Moreover, dantrolene, which prevents the propagation of calcium waves, prevents in vitro classical conditioning of Hermissenda (Blackwell and Alkon,1999).
PLA2 Mediates Contribution of GABA Stimulation to PKC Activation
These contributions to the calcium elevation were tested both experimentally and using model simulations. Model simulations demonstrated that the rebound depolarization produced an enhanced calcium elevation and a cumulative depolarization only when light and hair cell stimulation occur in the correct temporal relationship (Werness et al.,1993). However, subsequent experiments showed that a cumulative depolarization was not necessary to produce the enhanced calcium (Matzel and Rogers,1993). The alternative, that GABA stimulation contributes to the calcium elevation in the soma independent of the effect on membrane potential, was evaluated subsequently. Model simulations demonstrated that, if metabotropic GABAB receptors are coupled to phospholipase C, then the GABA stimulation alone initiates a calcium elevation that propagates as a wave (via release through ryanodine receptors) toward the soma (Blackwell,2002a). However, inactivation of the ryanodine receptor by the light-induced calcium wave propagating from the rhabdomere toward the terminal branches produces a refractory period and prevents the GABA-induced calcium wave from propagating past the light-induced calcium wave when the stimuli are paired (Blackwell,2004). Thus, the CS and US waves destructively interfered with each other, preventing the US-induced calcium from adding to the CS-induced calcium. This result suggested that a different mechanism activated by hair cell activity combines with the light-induced calcium elevation to initiate memory storage.
A more likely possibility is that GABAB receptors are directly coupled to PLA2 and that GABA released onto the terminal branches leads to production of arachidonic acid. Application of GABA produces an increase in arachidonic acid, comparable to that produced by direct stimulation of PLA2 (Muzzio et al.,2001). Application of an arachidonic acid analogue plus light produces an increase in excitability similar to that seen with classical conditioning, and it is blocked by the PKC inhibitor chelerythrine. Presently, this hypothesis has the most support, but confirmation requires delineation of the G-proteins that couple GABAB receptors to PLA2.
Serotonin-Activated Pathways Contribute to Memory Storage
Yet another possibility is that hair cells activate interneurons that release serotonin onto the photoreceptor soma (Land and Crow,1985). Serotonin is synthesized and released in the Hermissenda nervous system (Auerbach et al.,1989), and type B photoreceptors depolarize when serotonin is applied to the soma (Rogers and Matzel,1995). Light paired with serotonin has been shown to be an effective in vivo training paradigm, producing a reduction in phototaxis similar to that observed with standard classical conditioning paradigms (Crow and Forrester,1986). Light paired with serotonin leads to an enhanced light response (Crow and Bridge,1985) and a PKC-dependent increase in both input resistance and excitability of the type B photoreceptor (Crow et al.,1991). The modulation of ionic channels accompanying these changes also are similar to that observed with paired light and hair cell stimulation (Crow and Bridge,1985; Farley and Wu,1989; Acosta-Urquidi and Crow,1993). Despite all of this evidence, no one has demonstrated serotonergic projections onto the photoreceptors. Thus, support for serotonin as a critical neurotransmitter requires identifying the neurons that release serotonin due to hair cell stimulation.
Whereas GABAB stimulation produces arachidonic acid that is required for persistent activation of PKC, the precise role of serotonin in the activation of second-messenger systems is poorly understood. Conditioning using light paired with rotation or serotonin induces the activation of ERK1 and ERK2, members of the mitogen-activated protein kinase (MAPK) family (Crow et al.,1998), by ERK activating kinase (MEK1; Fig. 4), via both PKC-dependent and PKC-independent pathways (Crow et al.,2001). In mammals (Sweatt,2001), MAPK kinase is activated by an MAPK kinase kinase such as Raf-1 or B-Raf, which are activated (directly or via intermediaries) by PKC, protein kinase A (PKA), or growth factor tyrosine kinase receptors. In Aplysia, serotonin receptors are coupled to PKA activation and subsequent ERK activation (Barbas et al.,2003). Whether PKA or tyrosine kinase is involved in ERK activation in Hermissenda photoreceptors and the role of ERK in temporal sensitivity await further study.
Classical conditioning leads to a change in the spatiotemporal firing pattern of the set of five photoreceptors, which is further transformed into a qualitative change in motorneuron activation by interactions within interneuron layers. Which neurons are involved, and how do they interact to transform the spatiotemporal firing pattern into a change in behavior? Many of the relevant interneurons in the cerebropleural ganglia and motorneurons in the pedal ganglia have been characterized.
Light-Sensitive Interneurons Receive Mono- and Polysynaptic Input From Photoreceptors
One layer of interneurons in the cerebropleural ganglia, which are spontaneously active, are called central visual neurons or type I interneurons (Crow and Tian,2002b). These interneurons receive either EPSPs (type α or Ie) or IPSPs (type β or Ii) through a direct synaptic connection from ipsilateral type B photoreceptors. These interneurons also receive hair cell inputs of the same sign as type B photoreceptor inputs, i.e., interneurons that received EPSPs from type B photoreceptors also received EPSPs from hair cells (Fig. 5). Note that the connections from photoreceptors to type I interneurons is not indiscriminate and convergent, but highly targeted and divergent. Each photoreceptor makes connections to several type I interneurons, but each type I interneuron receives inputs from only a single photoreceptor. Thus, lateral and medial photoreceptors do not project to the same interneurons, nor do type A and type B photoreceptors project to the same interneurons.
A second layer of interneurons are light-sensitive but do not receive monosynaptic projections from photoreceptors. These spontaneously active neurons are called type II interneurons. The type IIe interneurons receive polysynaptic excitatory input from photoreceptors and project primarily to the ipsilateral pedal ganglion. The type IIi interneurons receive polysynaptic inhibitory input from photoreceptors and project to unidentified neurons in the contralateral cerebropleural ganglion (Crow and Tian,2002b).
Role of Dorsal Motorneurons in Delayed Phototaxis
The first set of light-responsive motorneurons studied are located on the dorsal surface of the pedal ganglia. These motorneurons are spontaneously active in the dark. Some of the motorneurons (MN1, P9, some MN4) increase their activity to light; others (P7, some MN4) decrease their activity in response to light (Richards and Farley,1987; Hodgson and Crow,1991). Of these, MN1 is thought to be the most significant, because stimulation causes movement of the tail, and because it receives polysynaptic excitatory input (via a type 1 interneuron) from type A photoreceptors (Goh and Alkon,1984).
The first hypothesis regarding the expression of classical conditioning was based on the differences in these motorneurons between paired and randomly trained Hermissenda. The light-induced increase in activity of MN1 and P9 is smaller in paired animals (Goh et al.,1985; Hodgson and Crow,1992). Similarly, the light-induced decrease in activity of P7 is smaller in paired animals. Prior to conditioning, type A photoreceptors indirectly excite motor neurons, such as MN1, responsible for positive phototaxis (Fig. 6, left: gray connections). After conditioning, an increase in excitability of type B photoreceptors leads to an increase in type B to type A photoreceptor inhibition, which leads to a decrease in excitation from type A photoreceptors to MN1 (and P9) motorneurons (Fig. 6, right: gray connections), which results in less movement toward the light and delayed phototaxis.
This hypothesis is consistent with the effect of classical conditioning on pedal nerve activity and delayed phototaxis. Locomotion is controlled by four of the six pedal nerves (Richards and Farley,1987) through which all motorneuron axons travel. The multiunit pedal nerve activity of control animals increases by about 10–20% during light stimulation; in paired animals, the multi-unit activity increases less (Rogers and Matzel,1996) or decreases below the dark activity level (Richards and Farley,1987). These observations are consistent with the hypothesis that type B photoreceptors suppress pedal nerve activity via suppression of type A photoreceptors, but do not explain foot contraction that precedes delayed phototaxis.
Ventral Motorneurons Are Involved in Classical Conditioning Behavior
Recently, Crow and colleagues have discovered additional circuits of information flow involved in classical conditioning behavior. They have characterized a second set of motorneurons on the ventral surface of the pedal ganglia, as well as additional interneurons receiving input from statocyst hair cells. VP1 and VP2 are motor neurons located on the ventral surface of the pedal ganglia and are responsible for ciliary movement mediating phototaxis (Crow and Tian,2003a). VCMN is a foot contraction motorneuron located on the ventral surface of the pedal ganglia. Several interneurons directly and indirectly project to these ventral motorneurons. Firing of both Ii and IIi interneurons is accompanied by an increase in IPSPs in VP1 and VP2; thus, these neurons have an indirect inhibitory projection to the ciliary motorneurons. In contrast, firing of the interneuron IIe produces a decrease in IPSPs to VP1 and VP2, implying that IIe is inhibiting a spontaneously active neuron that provides tonic inhibition to VP1 and VP2. Such a neuron may be the IIIi interneuron, which inhibits VP1 and VP2 (Fig. 5). All three of these pathways (Fig. 6, red connections) produce polysynaptic disinhibition of VP1 and VP2 motorneurons in response to light. Alternatively, the pathway through interneuron Ie is polysynaptic inhibition (Fig. 6, blue connections), because firing of interneuron Ie produces an increase in IPSPs to VP1 and VP2. The US pathway is mediated by VCMN, which does not receive input from interneurons that receive polysynaptic inputs from photoreceptors (Crow and Tian,2004). It receives excitatory input from the spontaneously active Ib interneuron (Fig. 5), which is indirectly excited by statocyst hair cells (Fig. 6, black connections). Thus, via this pathway, vestibular signals are converted into foot contraction.
Changes in excitability and synaptic connection strength in this network can further explain the change in phototaxis caused by classical conditioning. First, paired training causes facilitation of the monosynaptic PSP measured in type I interneurons and an increase in light-evoked spike activity (Crow and Tian,2002a). The enhancement of complex EPSPs to interneuron Ie and complex IPSPs to interneuron Ii is partly due to the increase in type B photoreceptor spike frequency and partly due to the enhancement in intrinsic properties of the interneuron. This enhancement in the Ie light response is propagated through the circuit as an increase in inhibition of VP1 (Crow and Tian,2003b). Presumably, this enhancement of inhibition is greater than the enhanced disinhibition (caused by the increased type B activity) mediated by the IIe and IIi interneurons. The net result is that prior to pairing, light causes VP1 excitation via disinhibition, whereas after pairing light causes VP1 inhibition. In summary, the change in VP1 response consequent to classical conditioning can explain both the change in pedal nerve multi-unit activity and the change in phototaxis; however, the neural circuitry and information flow that explains the foot contraction in response to light after classical conditioning remains concealed.
Do Central Pattern Generators Play a Role in Hermissenda?
In other systems, many motor behaviors are controlled by neuronal activity patterns produced by an interconnected sets of interneurons called the central pattern generator (CPG). The CPG activates motorneurons to produce motor behavior; an alteration in the neuronal activity pattern of the CPG transforms the motor behavior (e.g., Combes et al.,1999; Nargeot et al.,1999). The pattern may be altered by a change in synaptic inputs, e.g., due to sensory inputs, or neuromodulators, which modify synaptic or ionic channels in CPG neurons (Marder and Calabrese,1996). For example, in mollusks such as Tritonia and Pleurobranchia, both crawling and escape swimming are produced by central pattern generators (Jing and Gillette,1999; Popescu and Frost,2002). The rhythmic motor pattern is switched from crawling to escape swimming by patterns of synaptic inputs from command neurons (Jing and Gillette,2000).
The periodic burst firing of pedal nerves suggests that Hermissenda locomotion is controlled by central pattern generators homologous to those in Tritonia and Pleurobranchia. The change in photoreceptor spatiotemporal firing pattern caused by classical conditioning may produce a nonlinear change in the rhythmic motor pattern from that producing phototaxis to that producing foot contraction. Experiments to fill in the missing connections in Figure 5 may discover such a central pattern generator and demonstrate for the first time the complete circuitry and activity patterns responsible for classical conditioning behavior.
The study of Hermissenda classical conditioning was initiated decades ago in the belief that uncovering the secret to learning and memory had the best chance of success by studying simple learning behaviors in simple animals. Indeed, a body of research convincingly demonstrates that ionic and synaptic channels in photoreceptors are modified during memory storage. Similarly, many of the second-messenger pathways activated by CS and US have been delineated. Nonetheless, this review argues that significant gaps in knowledge exist, most notably (but not exclusively) regarding the neural circuitry translating changes in photoreceptor excitability to changes in behavior. The complexity of memory storage mechanisms revealed thus far, as well as the parallels with mammalian systems, reinforces the imperative to investigate invertebrate learning and memory.
The author thanks Harold Morowitz for comments on an earlier version of the manuscript. The author apologizes to the many authors whose work was not cited; the references had to be limited due to editorial discretion. Parts of the research reported in the manuscript was supported by the National Science Foundation (NSF) and the National Institute of Mental Health (NIMH).