Discussion

The purpose of this study was to investigate what responses Purkinje cells learn when they are trained with a uniform CS and a US that is presented after mixed temporal intervals. In eyeblink conditioning, such training protocols are known to produce complex response sequences and long-duration responses (15, 16).

The first question we wanted to answer was whether a Purkinje cell can learn more than one ISI and respond with temporal sequences that could generate double-peaked or long-duration overt blink CRs. The results show that Purkinje cells were indeed capable of learning more than one interval. Most cells learned a sequential response pattern with pauses timed to each of the two ISIs (13 of 27 cells), or a long-duration pause that extended over both ISIs (seven of 27 cells). In the remaining seven cases, cells acquired pause responses that were timed to either one of the two ISIs used during training.

To promote learning of clearly separable response components, we wanted to use ISIs that were sufficiently different that separate pause responses would not merge into one. This meant the second ISI had to be relatively long compared with the first. However, conditioning with long ISIs requires more time (3, 13, 24), and the acute setup in these experiments provides only a limited window for studying learning effects. In this trade-off, we settled on a paradigm with a 300-ms difference between the two ISIs (150 and 450 ms). This factor is a plausible cause of the observed long-duration pauses.

Importantly, the particular ISIs used here are not the only ones that can produce adaptively timed sequential PcCRs. This is illustrated by a case in a previous study (13), in which a cell had first been conditioned with a mossy fiber CS to an interval of 200 ms that was later shifted to 600 ms, after which it learned a response sequence that was adaptively timed to both intervals (compare Fig. 4c in ref. 13). Similarly, Fig. 2H in ref. 12 illustrates a case with well-timed response sequences to a parallel fiber CS after an ISI shift from 200 and 350 ms. In addition, double Purkinje cell pauses in two conditioned rabbits after training with ISIs of 200 and 700 ms have been reported by Halverson et al. (18).

These observations of sequential PcCRs, together with the observation of dual responses in the anterior interpositus nucleus reported by Choi and Moore (17), strongly suggest it is sequential PcCRs that produce the complex overt response patterns seen in eyeblink conditioning with mixed ISIs and after ISI shifts.

The second question we wanted to answer was how response components are related to each other. Let us first consider the possibility that two pauses occur because the CS elicits the same pause response twice. This interpretation can be rejected because the different pauses were not acquired simultaneously during training: The early pause was always learned before the late pause (see raster plots in Fig. 2 C and D for examples). The latter was thus not merely a repetition of the first, as it should have been elicited twice by the continuous CS once it had been learned. Instead, we believe the response pattern should be interpreted as a sequence with two different pause responses, separated by a brief period of simple spike firing. Further evidence in support of this interpretation comes from previous studies in which, when training with a single ISI, we have applied long-duration CSs that outlast the ISI by several hundred milliseconds (12, 13, 19, 25), but almost never observed more than one pause except when two ISIs have been used during training.

If the second pause in the sequential PcCR was not a repetition of the first pause, could the first pause be elicited independent from the second? We have previously found, after training with single ISIs, that a mossy fiber CS lasting only a few milliseconds or tens of milliseconds can elicit a full-blown, accurately timed PcCR (19), as well as a complete and accurately timed blink CR (20). If the sequential PcCR is a single response unit, then a CS that elicits the early components could potentially also elicit the later ones. However, short CSs that elicited the first pause, and the peak in spiking activity afterward, did not elicit the second pause. Because the second pause was not elicited by the early part of the CS, or as a consequence of the early pause or the middle peak, the second pause component was most likely elicited by signals arriving at the Purkinje cell later than 80 ms after CS onset.

What could be the source of these late signals? In an awake animal that has been conditioned to blink twice, the first blink will elicit sensory feedback that could potentially act as a CS for a second blink. However, with regard to the PcCRs in our experiments, such sensory feedback was excluded as a second CS, as the animals were immobilized and could not blink. Nevertheless, one could imagine that the first Purkinje cell pause produced neural activity downstream in the anterior interposed nucleus, the red nucleus or the facial nucleus, that eventually was fed back up to the Purkinje cell via mossy fibers or nucleocortical projections (26), as an efference copy that could act as a CS for the second pause. The trials with short CSs are a test of this idea as well. The short test CS elicited the first pause and the following excitation, but did not elicit a second pause. An efference copy or feedback signal was thus unlikely to play any role in eliciting the second pause, as it should have been present also in the short test CS case. Rather, it suggests additional (i.e., later) stimulus pulses in the mossy fiber CS were necessary to elicit the second pause component. The second pause was thus a separate response, and it was in response to a particular part (i.e., a later part) of the CS. A behavioral phenomenon that may be related to this PcCR property is the observation that a short (50 ms) CS could elicit blink responses after training with a 250-ms ISI, but could not elicit long latency responses after training with a 500-ms ISI (27).

The long-duration pause responses that were sometimes observed could plausibly be interpreted as double-pause response sequences lacking the intermediate simple spike firing. They were thus probably two differently timed pause responses that had merged into one, meaning the long-duration response profile illustrated in Fig. 3C could be a summation of short-latency and long-latency responses with profiles resembling those illustrated in Fig. 3 C and D, respectively. As was the case for the Purkinje cells that learned sequential responses, the cells we observed as they learned long-duration pauses also acquired the short latency pause onset before the longer one. Again, the first response property that was learned was the latency to onset.

The excitatory simple spike components that followed pauses were observed in all cells except two, regardless of whether the response profile was sequential, a long-duration pause, or timed to only one of the intervals. The offset of the CS and the offset of pause often overlapped, making it hard to tell which was the cause. However, the short CS tests also elicited an excitatory component after the pause in all cases, suggesting rebound excitation after pauses was the main cause of increased activity, although a delayed and adaptively timed excitatory response to the initial parts of the CS may also have been involved. In previous investigations of single PcCRs (12, 13, 19, 28), we sometimes observed excitation after the pause when the CS was delivered to the mossy fibers or parallel fibers and extended beyond the CS duration. Similar excitation was observed here. This suggests a direct mossy/parallel fiber CS may recruit more parallel fibers than a peripheral CS. In addition, in cases in which PcCR offset and CS offset did not coincide, the increased activity coupled to CS offset and not PcCR offset suggests tonic inhibition from molecular layer interneurons may also have been elicited by the CS.

One characteristic property of PcCRs is that they can be acquired to CSs consisting of repetitive trains of electrical stimuli delivered for several hundreds of milliseconds. In all cases, the PcCRs are timed to the ISI (7, 13), even though the CS is uniform. Importantly, Purkinje cells acquire the same PcCR regardless of where in the signal pathway the CS is applied. The CS can be a peripheral sensory stimulus consisting of subcutaneous electrical stimulation of the forelimb, or it can be direct electrical stimulation of mossy fibers projecting to the cerebellum, or even direct stimulation of the parallel fibers (12). In all cases, the PcCR displays a characteristic adaptively timed temporal profile that is shaped by the ISIs used during training (7). The most parsimonious explanation for this, in our opinion, is that the different CSs elicit similar parallel fiber activation in response to each electrical pulse in the CS, regardless of where it is delivered.

Earlier theories of cerebellar response timing have generally assumed that temporal information in the CS is transmitted to the Purkinje cell by numerous parallel fibers with different time-varying activity patterns that span several hundreds of milliseconds (29⇓–31). The mechanism by which a Purkinje cell learns when to respond was supposed to be realized by selectively modifying the strength of those parallel fiber synapses that are active around the time of the US.

We have previously shown that a Purkinje cell could learn an adaptively timed PcCR to a uniform parallel fiber CS where no temporal information in the CS signal was present (12). Furthermore, expression of the PcCR was independent of GABA-ergic transmission from inhibitory molecular layer interneurons. Instead, the PcCR seems to be elicited by glutamate from parallel fibers acting on type 7 metabotropic glutamate receptors (32). The temporal memory is therefore unlikely to be encoded exclusively in synaptic connection strength, but probably also involves a timing mechanism within the Purkinje cell (14).

The question now arises whether a Purkinje cell that has acquired a sequential PcCR has memorized more than one time interval. We think the observations presented here suggest this is the case, and that they constitute additional evidence challenging the adequacy of the standard theory of temporal learning as relying exclusively on strengthening and weakening of synaptic connections (see e.g., ref. 33 for a discussion). The results are in better agreement with theories suggesting the existence of additional intracellular mechanisms for storage of temporal relations in associative learning and for response timing (34⇓–36). However, the inadequacy of synaptic weight changes to account for temporal memory traces does not mean the learning is not synapse specific. Most likely, the memory trace is coupled to a specific subset of synapses, or dendritic compartments (37), but it is not exclusively reliant on a change in the strength of those synapses. The memory trace also probably includes a postsynaptic intrinsic mechanism for response timing that can produce timed pauses in simple spike activity by delayed hyperpolarization.

In conclusion, individual Purkinje cells are able to learn sequential response patterns that include both pauses and increased simple spike activity. Such sequential Purkinje cell responses may underlie learned motor sequences that have been shaped by sensory feedback signals arriving with different delays, such as, for example, the musical tones that are produced by finger movements while playing the piano. The ability of single Purkinje cells to learn response sequences suggests cerebellar control and coordination of motor behaviors may rely more on intracellular mechanisms and less on neuronal network properties than previously thought. It also suggests the capacity for information storage in the individual neuron is vastly greater and of a very different nature than suggested by the dominant paradigm.