From Molecules to Memory in the Cerebellum

See allHide authors and affiliations

Science  19 Sep 2003:
Vol. 301, Issue 5640, pp. 1682-1685
DOI: 10.1126/science.1090462

This article has a correction. Please see:

The “holy grail” of memory researchers is to produce a comprehensive model of memory storage that flows from molecules to behavior with all of the intermediate steps defined. This level of understanding does not yet exist for any form of memory in any model organism. However, the report by Koekkoek et al. (1) on page 1736 of this issue brings us a step closer to a model for a simple form of motor memory required for a task known as associative eyelid conditioning. The authors analyzed learning in a transgenic mouse that lacks a form of persistent synaptic modification called cerebellar long-term depression.

Pioneering neuroscientists such as Sechenov, Cajal, and Hebb hypothesized that memory could be stored by experience-dependent changes in synaptic strength. This idea gained momentum in 1973, when Terje Lomo and Tim Bliss discovered that brief high-frequency stimulation of excitatory synapses in the hippocampus of the brain could produce a long-lasting increase in the strength of synaptic transmission, a phenomenon called long-term synaptic potentiation (LTP). Some years later, a use-dependent synaptic weakening called long-term synaptic depression (LTD) was also found in the hippocampus. The idea that LTP and LTD could underlie memory storage became popular, in part, because damage to the hippocampus was known to produce impairments in memory for facts and events, so-called declarative memory. As the molecular underpinnings of hippocampal LTP and LTD became more defined, neuroscientists sought to understand memory storage in terms of these changes in synaptic strength.

Realization of this promise for the hippocampal system has not been straightforward. The current state of the art among aficionados of hippocampal memory is to inject a drug or introduce a mutation in mice that will interfere with LTP or LTD. Sometimes this will produce a deficit in a declarative memory task (2), but this approach has yielded little understanding of the intermediate processes by which LTP and LTD might change hippocampal circuit and regional function to constitute a behavioral memory trace.

Fortunately, this approach is somewhat more tractable for some simple, nondeclarative forms of memory such as associative eyelid conditioning. In this task, a weak periorbital shock (the unconditioned stimulus, US) is delivered to the eye, eliciting a reflexive blink (the unconditioned response, UR). When a neutral conditioned stimulus (CS) such as a tone is repeatedly paired with a periorbital shock so that the two terminate simultaneously, the animal learns to blink its eye in a carefully timed manner (the conditioned response, CR) such that the eyelid is lowered when the shock arrives. This simple form of motor learning is blocked if the cerebellum is damaged or temporarily inactivated with drugs during training. Furthermore, in well-trained animals, populations of cells in a region called the deep cerebellar nuclei begin to fire during the interval between CS onset and US onset. This firing is predictive of the performance of the CR, suggesting that the memory trace for associative eyelid conditioning is expressed in the firing rate and pattern of deep nuclear neurons. Importantly, during training, artificial electrical stimulation of cerebellar mossy and climbing fibers can substitute for the tone CS and shock US, respectively. These observations constrain the potential site of the memory trace: It must be in a location where the streams of CS and US information converge, and it must result in a carefully timed increase in firing of the deep cerebellar nuclei.

One favored hypothesis is that simultaneous activation of US-encoding climbing fibers and a CS-encoding mossy fiber-parallel fiber disynaptic relay results in LTD of excitatory parallel fiber-Purkinje cell synapses (35), particularly those that are active shortly before the US signal arrives (6). This results in reduced Purkinje cell firing, which attenuates inhibitory drive from the Purkinje cell to the deep cerebellar nuclei. This drives the increase in deep nuclear activity that underlies the timed eyeblink CR (see the first figure). Thus, for associative eyelid conditioning, an actual circuit-based model exists to relate a synaptic phenomenon (parallel fiber LTD) to a behavior (acquisition of the CR).

A cerebellar circuit for associative eyelid conditioning.

Neural activity evoked by the periorbital shock unconditioned stimulus (US) and the tone conditioned stimulus (CS) are conveyed by the climbing fiber system and the mossy/parallel fiber disynaptic system, respectively. These streams of information converge on cerebellar Purkinje cells. Repeated tone/shock pairings result in LTD of those excitatory (+) synapses between parallel fibers and Purkinje cells in which activity immediately precedes climbing fiber activation. This may reduce tone-evoked Purkinje cell activity, thereby attenuating inhibitory (-) synaptic drive from Purkinje cells to the deep nuclei. Disinhibition of the deep nuclei leads to increased activity in this structure, which appears to drive the carefully timed, learned eyeblink (the conditioned response, CR). [Source (15)]

Recently, parallel fiber LTD has yielded many of its molecular details (7). Its induction requires activation of the glutamate receptor mGluR1 by parallel fibers and the simultaneous activation of voltage-gated calcium channels by climbing fiber-evoked depolarization. These signals converge to activate protein kinase C (PKC), which phosphorylates serine 880 in the carboxyl-terminal tail of the AMPA receptor GluR2 (8). Ultimately, GluR2 is internalized via clathrin-mediated endocytosis, leading to synaptic depression. Previous experiments have tried to use these insights to test the hypothesis that parallel fiber LTD is required for associative eyelid conditioning. For example, mGluR1-null mice have deficits in acquisition of associative eyelid conditioning, but this has been difficult to interpret because these learning deficits are accompanied by defects in both motor coordination (the mice are clumsy before training starts) and wiring of the cerebellar circuit (9). Furthermore, there has been a recurring technical problem with the analysis of mouse eyelid conditioning in which false-positive eyeblinks have been recorded due to twitching of nearby facial muscles (10).

In the new work, Koekkoek et al. solve the first problem by using a transgenic mouse called L7-PKCi, which has a Purkinje cell-specific promoter to drive the expression of a peptide that inhibits PKC. This mouse has normal motor coordination, normal cerebellar wiring, and normal basal and dynamic Purkinje cell firing properties but lacks parallel fiber LTD (1113). The authors solve the second problem by replacing the traditional electromyographic recording of eye muscles with a clever magnetic field-based measure of eyelid distance (10). The simple expectation from classic models is that L7-PKCi mice would show CRs that are rare and small. What Koekkoek et al. find is actually much more interesting: In well-trained PKC-inhibited mice there is only a small deficit in the initiation of the eyelid CR, which begins with the typical delay, proceeds at a lower velocity, and achieves a somewhat smaller peak amplitude (see the second figure). The robust effect they see is that the eyelid CR is not sustained, such that the eyelid is almost completely open when the shock US arrives. This timing deficit differs from that found in LTD-deficient mutant mice. In studies using knockout mice and electromyographic eyelid recording, the emergence of CRs was not ill-timed but rather suppressed (9). Thus, the entire memory trace of associative eyelid conditioning is not abolished in L7-PKCi, LTD-deficient mice, but rather information critical to the timing of the CR is lost.

A preemptive blink.

(Left) In an untrained mouse, the periorbital shock US gives rise to a reflexive eyeblink (the unconditioned response, UR). The tone conditioned stimulus (CS) produces no response at all. (Middle) In a trained animal, after many repeated pairings in which the CS and US terminate simultaneously, the mouse produces a carefully timed eyeblink (the conditioned response, CR) such that the eyelid is lowered when the US arrives. This correlates with a burst of activity in the cerebellar deep nuclei that appears to drive the CR. (Right) In L7-PKCi mice deficient in parallel-fiber LTD, the CR is smaller and, more important, is poorly timed so that it ends before the US arrives.

Does this result prove the hypothesis that parallel fiber LTD underlies adaptive timing of associative eyelid conditioning? Although it is a big step in the right direction, the answer would have to be no. PKC phosphorylates hundreds of proteins in Purkinje cells, not just GluR2 serine 880, which appears to be the critical substrate for parallel fiber LTD (8). Furthermore, recent work has shown that the climbing fiber-Purkinje cell synapse also exhibits LTD, which is PKC dependent and is therefore likely to be blocked in this mouse (14). Therefore, the causal link between parallel fiber LTD and learned timing of eyelid responses is not absolute. The next-generation test of this hypothesis will involve a subtle mutation of the critical PKC substrate for LTD so that it cannot be phosphorylated, while the normal phosphorylation status of other PKC substrates in Purkinje cells is maintained. Future work will also explore whether other motor-learning tasks not involving eye (11) or eyelid movements also require parallel fiber LTD.


  1. 1.
  2. 2.
  3. 3.
  4. 4.
  5. 5.
  6. 6.
  7. 7.
  8. 8.
  9. 9.
  10. 10.
  11. 11.
  12. 12.
  13. 13.
  14. 14.
  15. 15.
View Abstract

Navigate This Article