Cerebellar LTD: A Molecular Mechanism of Behavioral Learning?

Abstract

Changes in synaptic strength are almost certainly one mechanism of learning. Analysis of in vitro brain tissue has revealed two general kinds of changes in synaptic strength: long-term potentiation (LTP) and long-term depression (LTD) (7Linden D.J. Connor J.A. Annu. Rev. Neurosci. 1995; 18: 319-358Crossref PubMed Scopus (354) Google Scholar). Behavioral learning is dependent on experience and so are these two classes of molecular mechanisms. In many examples, LTP or LTD is caused by pairing either two inputs to a given site or presynaptic and postsynaptic electrical activity at the site. Thus, LTP and LTD provide a way to convert highly specific patterns of neural activity into changes in synaptic strength. One of the major challenges faced by neuroscience is to understand how specific examples of molecular mechanisms like LTP and LTD function in neural circuits that control behavior—in short, to find molecular mechanisms of behavioral learning. What does it mean to “find” the molecular mechanisms of learning? Systems neuroscience and molecular biology are currently collaborating in finding many examples of correlations between precise gene-targeting knockouts of known molecular mechanisms and deficits in specific behaviors. However, correlation doesn't establish cause and effect. Behaviors are generated by neural circuits. I will argue that it is essential to establish how a given knockout alters the operation of the neural circuits so that we can understand why the behavior is altered. I want to draw the distinction explicitly between molecular “mechanisms of plasticity” and “mechanisms of learning.” LTP and LTD are mechanisms of plasticity because they change how brain cells work. To my knowledge there are not yet any examples in vertebrates where a strong enough link has been made to the behavior to elevate a mechanism of plasticity to the causal role implied for a mechanism of learning. A number of recent papers have heralded a new era in the cooperation of systems and molecular neuroscience based on exploring the effect of molecular manipulations on behaviors that use well-understood neural circuits (1Aiba A. Kano M. Chen C. Stanton M.E. Fox G.D. Herrup K. Zwingman T.A. Tonegawa S. Cell. 1994; 79: 377-388Abstract Full Text PDF PubMed Scopus (368) Google Scholar, 14Shibuki K. Gomi H. Chen L. Bao S. Kim J.J. Wakatsuki H. Fugisaki T. Fujimoto K. Katoh A. Ikeda T. Chen C. Thompson R.F. Itohara S. Neuron. 1996; 16: 587-599Abstract Full Text Full Text PDF PubMed Scopus (292) Google Scholar, 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar). The cellular mechanism under investigation is LTD in the cerebellar cortex, and the behaviors being used are simple examples of learning in the motor system. Because the idea that the cerebellum could be a site of motor learning is supported by many different kinds of evidence, this seems like a propitious marriage of molecular and systems approaches. Two behaviors have been used extensively to study the role of the cerebellum in motor learning. One is classical conditioning of the eyelid response (4Kim J.J. Thompson R.F. Trends Neurosci. 1997; 20: 177-181Abstract Full Text Full Text PDF PubMed Scopus (276) Google Scholar), derived from Pavlov's dogs. If a noxious stimulus that always evokes an eye blink (e.g., a puff of air) is repeatedly paired with a neutral stimulus such as a tone, then the tone gradually acquires the power to evoke an eye blink even in the absence of the puff of air. One can think about this paradigm as using the puff of air as a teaching signal that informs the brain of the value of responding to the tone, which, because is precedes the puff, is always a good predictor of the need to blink the eye. The second behavior, and the focus of this minireview, is adaptive plasticity of the vestibulo-ocular reflex (VOR). But almost everything I say about the neural and molecular mechanisms of learning in the VOR is also true about eyelid conditioning (12Raymond J.L. Lisberger S.G. Mauk M.D. Science. 1996; 272: 1126-1131Crossref PubMed Scopus (471) Google Scholar, 4Kim J.J. Thompson R.F. Trends Neurosci. 1997; 20: 177-181Abstract Full Text Full Text PDF PubMed Scopus (276) Google Scholar, 9Mauk M.D. Neuron. 1997; 18: 343-346Abstract Full Text Full Text PDF PubMed Scopus (116) Google Scholar). In normal intact animals, head turns in one direction cause compensatory eye rotations in the opposite direction, even in darkness. The system has remarkable fidelity in monkeys: the amplitude of the compensatory eye rotation is nearly equal to the head turn. The fidelity is less convincing in wild-type mice, where the amplitude of the compensatory eye rotation is normally 20%–25% of the head turn, but the reflex is nonetheless present (5Koekkoek S.K.E. Alphen A.M.v. Burg J.v.d. Grosveld F. Galjart N. De Zeeuw C.I. Genes and Function. 1997; 1: 175-190Crossref PubMed Scopus (59) Google Scholar). In both species, pairing of visual and vestibular stimuli causes learning. If a visual stimulus moves exactly with head turns, requiring no VOR, then subsequent tests in darkness reveal that the amplitude of the VOR has decreased. If a visual stimulus moves opposite to head turns, requiring a VOR that is twice head amplitude, then the amplitude of the VOR in darkness increases. Learning starts within 30 min but takes several days to complete. One can think of the VOR as using visual inputs as a teaching signal, to inform the brain that the world appears to be moving because the performance of the VOR is inadequate. The molecular end result is increases or decreases in the strength of synapses in pathways from the vestibular input to the eye motor outputs. Why should the readers of Cell be interested in the VOR? Even though—or perhaps because—it is a lowly reflex instead of a higher cognitive function, the VOR is an ideal system for analysis of learning at all levels from molecules to computational models. The neural circuit is simple and largely known. The behavioral system for studying learning is straight-forward. At least in the monkey, recordings during and after learning have revealed both the neural signals that are available to guide learning and how the discharge of neurons throughout the circuit are altered in association with learning (12Raymond J.L. Lisberger S.G. Mauk M.D. Science. 1996; 272: 1126-1131Crossref PubMed Scopus (471) Google Scholar). In short, learning is understood, to a first approximation, at the level of neurons and neural circuits. The VOR is an opportunity to fit the molecular nuts and bolts of plasticity into the framework of learning in a neural circuit. I'm a VOR chauvinist, so I think this is the best opportunity neuroscience has to understand how learning works in the brain. Questions of how the brain learns must be posed first as questions of where the brain learns. For the VOR (and the eyelid response), the where question has been answered. Learning almost certainly occurs both in the cerebellar cortex and the deep cerebellar nuclei. It is plausible that the first learning is primarily in the cerebellar cortex and that the deep cerebellar nuclei take a larger role as hours and days pass. Figure 1 explains how cerebellar learning is envisaged to work at the level of neural circuits. The cerebellum has a highly stereotyped structure. One kind of input enters the cerebellum as “mossy fibers” and terminates both on granule cells in the cerebellar cortex and on neurons in the deep cerebellar nuclei. Granule cells project to the molecular layer where they bifurcate into “parallel fibers” and make contacts with Purkinje cells, the only output neurons from the cerebellar cortex. Purkinje cells inhibit neurons in the deep cerebellar nuclei. A second kind of input arises in the “inferior olive” and enters the cerebellum as “climbing fibers.” This input makes an unconventional and powerful synapse on Purkinje cells and also gives collaterals to the deep cerebellar nuclei. In normal wild-type animals, each Purkinje cell receives inputs from only one climbing fiber and (indirectly) from many, many mossy fibers. The standard techniques of systems neuroscience—ablation, stimulation, recording, and computer simulation—have suggested that learning in the VOR and the eyelid response occur both in the mossy fiber to parallel fiber to Purkinje cell pathway and in the mossy fiber to deep cerebellar nucleus pathway (e.g.,12Raymond J.L. Lisberger S.G. Mauk M.D. Science. 1996; 272: 1126-1131Crossref PubMed Scopus (471) Google Scholar, 9Mauk M.D. Neuron. 1997; 18: 343-346Abstract Full Text Full Text PDF PubMed Scopus (116) Google Scholar). Though far from proven, the site for learning in the cerebellar cortex is usually thought to be at the synapse from parallel fibers to Purkinje cells. Part of the basis for this consensus is the demonstration of LTD at this synapse in a variety of preparations (e.g.,13Sakurai M. J. Physiol. 1987; 394: 463-480Crossref PubMed Scopus (255) Google Scholar, 7Linden D.J. Connor J.A. Annu. Rev. Neurosci. 1995; 18: 319-358Crossref PubMed Scopus (354) Google Scholar, 6Lev-Ram V. Jiang T. Wood J. Lawrence D.S. Tsien R.Y. Neuron. 1997; 18: 1025-1038Abstract Full Text Full Text PDF PubMed Scopus (200) Google Scholar). If the consensus proves true, this would show remarkable prescience in 3Ito M. Brain Res. 1972; 40: 80-84Crossref Scopus (356) Google Scholar initial suggestion that one site of motor learning for the VOR would be in the parallel fiber to Purkinje cell synapse and that the mechanism would be LTD-driven by a teaching signal from visual climbing fiber inputs. The second generally accepted site for learning is in the deep cerebellar nucleus, which is in the brainstem for the VOR. The molecular mechanisms at this site have not yet been identified. Molecular genetics now has provided the best evidence that LTD plays a causal role in VOR learning. By selectively expressing an inhibitor of PKCγ in Purkinje cells2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar have concurrently abolished cerebellar LTD and adaptive plasticity of the VOR in mice. This approach obviates three potential criticisms of earlier papers that abolished LTD and weakened eyelid conditioning by gene targeting knockouts of mGluR1 (1Aiba A. Kano M. Chen C. Stanton M.E. Fox G.D. Herrup K. Zwingman T.A. Tonegawa S. Cell. 1994; 79: 377-388Abstract Full Text PDF PubMed Scopus (368) Google Scholar) or GFAP (14Shibuki K. Gomi H. Chen L. Bao S. Kim J.J. Wakatsuki H. Fugisaki T. Fujimoto K. Katoh A. Ikeda T. Chen C. Thompson R.F. Itohara S. Neuron. 1996; 16: 587-599Abstract Full Text Full Text PDF PubMed Scopus (292) Google Scholar). First, the loss of VOR learning appears to be complete rather than partial. Second, the inhibition of PKC was highly selective for Purkinje cells, making it likely that this is the site of the molecular lesion. Third, because the manipulation is not a gene-targeting knockout, it presumably does not cause up-regulation of compensatory genes that could confound the interpretation. I imagine that the VOR aficionados will find a number of reasons to take exception with 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar. Until there are transgenic monkeys with all of the advantages of primates for behavioral analysis, however, this is the best we've got for bridging from molecular mechanisms to behavioral learning. It is good enough to think about making the next step. Because of the tight relationship among the behavior, neural circuits, synapses, and cellular function, the analysis of the VOR has brought us much closer to the promised land of “understanding learning” than even the lauded analysis of the molecular basis of spatial learning in the hippocampus. With this in mind, I'll devote the rest of my minireview to setting the bar yet several notches higher, by outlining the general classes of problems that have to be solved to establish cerebellar LTD as a molecular mechanism of behavioral learning in the VOR. I am being careful to talk about LTD as a putative mechanism of learning rather than as the putative mechanism. Given the large amount of systems neuroscience data that cannot be explained without postulating two anatomical sites that are equal partners in causing learning (8Lisberger S.G. J. Neurophysiol. 1994; 72: 974-997PubMed Google Scholar), it would seem shortsighted to focus entirely on the site where a molecular mechanism is known. Recall that sites of plasticity are discovered by molecular approaches, sites of learning by systems approaches. Though LTD has been under the street lamp, it seems inevitable that molecular mechanisms of plasticity will exist in virtually every neuron in the brain. The presence of a well understood mechanism at one site and the absence of a mechanism at another site, perhaps for lack of looking, does not provide much information about which site(s) and what mechanism(s) actually cause learning. To make the desired causal link from molecular mechanisms to behavior, it will be important to overcome three classes of problems with current knowledge. These problems have to do with the temporal and spatial specificity of the molecular manipulations, the relevance of cellular analysis in vitro to cellular function in vivo, and the strength or weakness of the essential link from behavioral analysis in mice to the wealth of information about the neural circuit basis for the behavior in monkeys. In a sense, these questions all boil down to one issue. The VORs of transgenic and wild-type mice are black boxes. Little is known about the neural signals that give rise to the performance of the black box, either during development or in the adult mice used to analyze the behavioral phenotype of each mutation. Neural circuits are not static, and they do not emerge as a result of the genetic code without environmental shaping. Everyone knows that cerebellar circuits might not function correctly in adults if a molecular mechanism such as LTD were inhibited starting from the end of Purkinje cell development. This causes a fundamental problem in interpretation of all experiments that use the logic implicit in 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar. Neural signals guide learning. They are the afferent input to molecular mechanisms of plasticity. If the neural signals are abnormal in the mutant, then any and all molecular mechanisms of learning will be deafferented. One reason for thinking that inhibition of PKCγ might affect neural signals is that nearly all ion channels appear to be regulated by kinases. Since some of the kinase-regulated ion channels control interspike intervals, inhibiting a kinase in Purkinje cells seems likely to alter their spike frequencies. Molecular and behavioral studies alone cannot determine whether the loss of motor learning in the VOR might be due to a fundamental circuit abnormality that itself is caused directly by the inhibition of LTD. LTD itself might not be involved in learning at all. In many systems, this criticism is the last resort of a chronic skeptic. For the VOR, the situation can and should be analyzed by recording the neural signals in the relevant part of the cerebellum in the wild-type mouse and the mutant with PKCγ inhibited. If the signals are proven normal in their mutants, then 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar will have strengthened the evidence for a causal relationship between LTD and VOR learning. Cellular analysis of LTD has been highly controversial. Even from a strictly cellular perspective, the signaling chain seems to be quite different in culture (7Linden D.J. Connor J.A. Annu. Rev. Neurosci. 1995; 18: 319-358Crossref PubMed Scopus (354) Google Scholar) versus slice (6Lev-Ram V. Jiang T. Wood J. Lawrence D.S. Tsien R.Y. Neuron. 1997; 18: 1025-1038Abstract Full Text Full Text PDF PubMed Scopus (200) Google Scholar) preparations. 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar demonstrated a loss of LTD in cultures. It will be important to determine whether inhibition of PKCγ abolishes LTD in slices and in vivo. The need to have LTD work in the neural environment of the intact functioning animal adds two very sticky issues. One is the well-recognized need for timing contingencies in LTD so that it compares each climbing fiber input with the parallel fiber inputs from 100 ms earlier. Without this timing contingency, analysis of the neural signals present at the Purkinje cell during learning predicts learning when it doesn't occur and even predicts increases in the amplitude of the VOR when decreases are observed (11Raymond J.L. Lisberger S.G. Learning and Memory. 1997; 3: 503-518Crossref PubMed Scopus (30) Google Scholar). Cellular analyses have focused on coincidence of climbing fiber and parallel fiber inputs. Although there have been a number of studies of the timing contingencies, no convincing consensus has emerged yet. Finally, there is the annoying and contentious problem that the changes in response of Purkinje cells recorded after VOR learning in monkeys are consistent with changes in synaptic strength opposite in direction to those predicted by LTD (10Miles F.A. Fuller J.H. Braitman D.J. Dow B.M. J. Neurophysiol. 1980; 43: 1437-1476PubMed Google Scholar). These problems don't disprove the idea that cerebellar LTD is a molecular mechanism of VOR learning. But they do present issues and paradoxes that need to be resolved. The final issue concerns possible species differences in the normal VOR and in learning in the VOR between mice and monkeys. I'm not suggesting that species differences are a safe refuge for the chronic skeptic. Indeed, the VOR is evolutionarily old, and it is hard to imagine that there are fundamental species differences in how the brain solves this relatively simple problem. Still, one would be reassured if the behavior were qualitatively similar across species. It is disconcerting that the VOR in wild-type B6C3F1 littermates of the PKCγ-inhibited mice are highly abnormal. Compensatory eye movements should be out-of-phase with head movements, but they are as close to in-phase as to out-of-phase. In contrast, compensatory eye movements in wild-type B6CBACa and 129/C57B16 mice (5Koekkoek S.K.E. Alphen A.M.v. Burg J.v.d. Grosveld F. Galjart N. De Zeeuw C.I. Genes and Function. 1997; 1: 175-190Crossref PubMed Scopus (59) Google Scholar) and monkeys are almost exactly out-of-phase with head movement. Although controversial, it also seems that there may be species differences in the neural signals in the relevant part of the cerebellum. Those recorded in the rabbit, for example, seem to be quite different from those in the monkey. Are the neural signals in the relevant part of the mouse cerebellum like those in the monkey or the rabbit? Asked a different way: is my initial assertion valid, that enough is known about the operation of the neural circuits for the mouse's VOR so that molecular manipulations on mice can be used to elucidate the molecular mechanisms of behavioral learning? In my opinion, not yet. I will be reassured once the standard tools of systems neuroscience have been used to understand the detailed workings of the neural circuits for the VOR, in both wild-type and mutant mice. If we can understand how neural circuits work, or why they don't work, after deletion or inhibition of specific molecular mechanisms of plasticity at specific anatomical sites, then we can draw strong conclusions about the role of those mechanisms in behavior. I am suggesting that the next step, to get over the elevated bar, is to analyze the very promising mouse generated by 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar in detail. I'd like to see this done instead of, or at least concurrently with, a search for more mutations that reveal yet another correlation between deficits in LTD and behavioral learning. Because of the feasibility, in the VOR, of experiments to resolve the issues I've raised, the finding of a concurrent loss of LTD and VOR learning by 2De Zeeuw C.I. Hansel C. Bian F. Koekkoek S.K.E. Morpurgo M.M. Linden D.J. Oberdick J. Neuron, in press. 1998; Google Scholar may have brought us to the brink of achieving an unprecedented understanding of how learning occurs in at least one behavioral system and structure in the brain.