Main

In the past decade, most research on neurodevelopmental disease has focused on identifying specific genetic defects that might lead to abnormal brain development. However, it is widely accepted that normal infant development also depends on sensory input that provides appropriate patterns of neural activity to shape the developing brain. Recent studies have revealed a similar requirement for endogenous neural activity generated by the nervous system itself, long before there is any sensory input. These patterns of sensory-driven and endogenously generated neural activity sculpt the precise circuits that are crucial to the many complex functions of the adult brain. In this article, we focus on activity-dependent development to emphasize the necessity of considering physiologic, activity-dependent mechanisms, rather than just single gene defects, when searching for clues to disorders of brain development. We review the principles of activity-dependent development as illustrated by the formation of precise connections in the mammalian visual system. Experiments that indicate that similar activity-dependent mechanisms are at work shaping many parts of the nervous system are also reviewed. We propose that disruptions in early neural activity, and thus disruptions in the formation of precise circuits, may contribute significantly to many common neurodevelopmental disorders.

ACTIVITY-DEPENDENT REMODELING OF EARLY NEURAL CONNECTIONS

As the human brain develops, billions of neurons make an average of one thousand synapses to become interconnected in precise neural circuits. How are these complex neural connections established? First, cells must be generated by successive cell divisions and their identity must be determined-as neurons, and then as particular classes of neurons. Second, neurons from one region must extend axons along specific pathways to appropriate target regions to form the linked pieces of a functional system. However, the initial pattern of connections is often imprecise. The third step, on which we focus here, is the refinement of these connections to form the specific patterns of connectivity that characterize the mature brain.

The first two steps, cell type determination and pathway formation, are often referred to as "activity-independent" processes because, in general, neuronal activity (i.e. transmission between presynaptic and postsynaptic partners via action potentials and neurotransmitter release) is not required for these processes to occur. Rather, signals for differentiation, cell type determination, and axonal guidance are given by genetically specified molecular cues that appear to provide for the reliable construction of a stereotyped framework (1,2; but compare3). In contrast, the elaboration, retraction, and remodeling of neural connections once they are within their targets is thought to be "activity-dependent" because the ultimate patterns of connections can be disrupted by blockade of neuronal activity(1,4).

Much laboratory and clinical work has focused on neurogenetics and the role of activity-independent factors in the neurologic disease of infants and children(5,6). Numerous "activity-independent" pediatric neurologic diseases have been described: disorders of neural induction (anencephaly)(7); disorders of neurogenesis (as seen in the "minibrain" mutation)(8,9); disorders of programmed cell death (spinal muscular atrophy)(1012); disorders of neural migration (focal dysplasias or heterotopias)(1315); and disorders of axonal pathfinding (agenesis of the corpus callosum)(5,16,17).

The list of "activity-dependent" disorders is much shorter. It is not that these disorders do not exist, rather they are under-recognized. Abnormal neuronal activity patterns have been implicated as the basis for widespread, diverse disorders including the progressive severity of some types of epilepsies (as in West syndrome in which hypsarrhythmia may promote the development of other epileptic circuits; for discussion see(18,19)), developmental delay(20), and autism(21). In addition, it is possible that common, more subtle learning deficits may result from abnormal neural circuitry(22) that could be produced by disturbances of activity-dependent processes.

The precise roles of abnormal cerebral circuitry and neuronal activity in specific disorders have been more difficult to identify because traditional histologic and radiologic studies in human subjects can visualize only the most profound disruptions in connectivity - those that result in major structural defects. Yet we know from animal studies that manipulations that disrupt normal neural activity can cause large changes in the detailed patterns of synaptic connections while leaving gross brain structure intact(23,24). The few detailed histologic studies that reveal dendritic morphology (i.e. by using the Golgi technique) in pathology specimens from children who have epilepsy or metal retardation can show striking changes in dendritic patterning(5,20,25), but such specimens are rarely obtained and these histologic methods are notoriously variable. In addition, the "activity-independent" disorders mentioned above often result in abnormal neuronal circuitry making primary and secondary causes of neurologic diseases difficult to untangle(5).

Newer methods of noninvasive neuroimaging are beginning to allow the exploration of functional changes in neural circuitry during development(22,2628). As the resolution of functional neuroimaging (positron emission tomography, functional magnetic resonance imaging, and spectroscopy) improves(2934), there is likely to be an explosion in the number of developmental neurologic disorders that are identified as "activity-dependent."

The difficulties in studying the role of activity-dependent developmental processes in humans emphasizes the need for good animal models. The development of the mammalian visual pathway is a model system for illustrating the principles of activity-dependent synaptic refinement: the precise anatomy of this pathway is well-defined; the need for neural activity in the formation of visual connectivity has been documented; specific patterns of neural activity present in this pathway have been described; and the mechanisms of synaptic reorganization underlying the fine-tuning of these connections are being investigated intensively. Understanding the general principles and mechanisms of activity-dependent neural development is, of course, the first step in understanding developmental brain disorders that may be caused by defects or interruptions of early neural activity.

THE DEVELOPING VISUAL PATHWAY AS A MODEL SYSTEM

The anatomy of the system. In most adult mammalian visual systems, from mouse to human, RGC axons from each eye project to the LGN on both sides of the brain, with the axons from nasally placed ganglion cells projecting to the contralateral LGN, and axons from the temporally placed ganglion cells projecting to the ipsilateral LGN (Fig. 1A). Within the LGN, the axons from the two eyes terminate in a set of separate, alternating layers in which LGN neurons receive purely monocular inputs. In addition to the ocular segregation pattern of the LGN, there is a topographic patterning to the connections such that the nearest neighbor relations of the RGC are preserved, projecting a retinotopic map of one hemisphere of the visual world onto the opposite LGN(35,36). In turn, LGN neurons project to layer IV of the primary visual cortex, where the axons are sorted into OCD, in which alternating ½-mm-wide columns of cortical neurons are driven only by input from one eye, and into orientation columns (or "pinwheels"), and the retinotopic map is again preserved(37,38). RGC also project to the superior colliculus, and the axon terminals are again grouped in an eye-specific and cell class-specific manner(3941).

Figure 1
figure 1

A, Specificity of the adult mammalian visual pathway. RGC axons project to the LGN topographically. The nasal axons cross over at the optic chiasm to the contralateral side of the brain, whereas the temporal axons project to the LGN on their own side of the brain. In the LGN, the RGC axons terminate in discrete eye-specific layers (only two are shown schematically here). The LGN neurons in turn project to the primary visual cortex, maintaining the retinotopic map. The LGN axons terminate in layer IV of the primary visual cortex where they are segregated into alternating eye-specific regions that form the basis of the ODC. B, Developing visual pathway. During development, the projection from the retina to the LGN is not segregated into eye-specific layers. The retinal axons from the two eyes initially overlap substantially (only the binocular portion of the LGN is shown). Refinement of this projection into the precise, mature pattern of connections requires neural activity(23,24). During this process of segregation waves of spontaneous, correlated activity that vary in location and direction spread through the developing retina (shown schematically in the eyes as colored patches representing the activity of groups of RGC; the gradation in intensity from dark to light indicates the direction of propagation over time)(83). After the layers in the LGN have formed, the LGN neuron axons will segregate from an initially diffuse projection into the ODC of the visual cortex (not shown). C and D, Disruption or imbalance of retinal activity blocks the development of the visual pathway. C, When the RGC waves of activity are blocked during RGC axon segregation in the LGN, then the layers in the binocular region of the LGN fail to form. When the RGC activity in both eyes is blocked then segregation does not occur; both sets of axons expand the LGN territory that they fill. D, When the balance of competition is disrupted by blocking the activity in only one eye, then the active eye (red) gains most of the binocular territory in the LGN and many of the axons from the inactive eye (blue) are driven out(86).

In the initial development of the visual system, neither the layers in the LGN nor the cortical ODC are present, and the retinotopic map is crude(4250). When RGC axons from each eye first innervate the LGN, they are intermingled in most of the nucleus (Fig. 1B)(43,51). The LGN neurons receive binocular innervation during this period(52), but, through a process of axon retraction and elaboration, connections are refined and the adult pattern emerges. Similarly, the LGN projections to layer IV are initially intermingled. LGN axons representing either the right or left eye then segregate to define OCD in the primary visual cortex (47,53,54; but compare55). In all mammals, the segregation of axons into eye-specific layers in the LGN precedes the formation of the cortical ODC.

GENERAL PRINCIPLES OF ACTIVITY-DEPENDENT DEVELOPMENT

Competition shapes the patterns of connections. Many experiments, beginning with those of Hubel and Wiesel(56,57), have indicated that visual experience plays a key role in segregation of LGN axons to form ODC in layer IV of the primary visual cortex(1). Segregation is thought to be achieved by an activity-mediated competitive process involving the formation, and elimination, of specific synapses. The process is said to be competitive in the sense that unequal levels of activity, or use, results in the dominance of connections from the more active eye at the expense of those from the less active eye. Segregation of LGN axons into ODC in the cortex is thought to occur from the initially intermixed state by a process in which interactions between LGN axons lead to the strengthening of those inputs that are simultaneously active, and activate the postsynaptic neuron, at the expense of those synapses that are not active synchronously and are consequently weakened and eliminated (Fig. 2A). These synapses are likely to have characteristics of Hebbian synapses-defined as synapses that are strengthened by the synchronous firing of both presynaptic and postsynaptic neurons(58). In other words "cells that fire together, wire together" and "those that don't, won't." When there is a disruption in this process caused by unequal activity levels in the two sets of inputs, the active inputs have a competitive advantage and gain more connections.

Figure 2
figure 2

A, Synaptic modifications based on patterns of activity. When co-innervating axons are synchronously active (axons 1 and 2; illustrated by the pattern of action potentials drawn schematically above each axon) with the postsynaptic neuron, they will both be maintained and strengthened. However, when a co-innervating axon is asynchronously active (axon 3), then this axonal connection will be lost. B, Predicted signals for the activity-dependent refinement of connections. An enlargement of the schematic shown in (A) illustrates the basic mechanisms that are necessary for activity-dependent competition. When axons 1 and 2 fire action potentials synchronously and depolarize the postsynaptic neuron, the postsynaptic neuron must have a mechanism for detecting this coincident firing (a coincidence detector [CD], such as an NMDA receptor). When presynaptic axons are simultaneously active, a "retraction" signal is produced (black arrows). The active axons (1 and 2) are protected from elimination, perhaps by a retrograde messenger that specifically stabilizes only active synapses (open arrows). When the inactive axon receives the retraction signal (axon 3), it is not protected and will therefore be eliminated. When all of the axons are inactive (not shown), no signals will be sent and the connection will remain (although they may be weaker than normal because of lack of support from retrograde messengers, such as neurotrophins; see text). Note that the messengers that govern protection and retraction may be presynaptically or postsynaptically generated (they are shown as postsynaptic here only for simplicity).

As Hubel and Wiesel first observed in the 1960s(56), closing one eye in a kitten during its early postnatal development profoundly disrupted the pattern of ODC; the eye that had visual input dominated the cortex, whereas the eye that had been closed lost its connections. Cortical connections to the opened or closed eye as assessed both physiologically(57,59) and anatomically(60) can be changed in less than a single week of deprivation. These experiments show that not only does vision drive the refinement of these cortical connections, but also that the balance of activity is critical, because these profound shifts occur with monocular visual deprivation.

When both eyes are closed and consequently there is no imbalance in ocular activity, the ODC can still form, but much more slowly(6164). Even with binocular eye closure, columns can still form as a result of competition because ganglion cells are spontaneously active in the dark(65,66). Visual deprivation does not abolish this spontaneous activity. Moreover, anatomic studies show that in monkeys, the ODC are already segregating in utero(46,64), most likely through competitive interactions driven by spontaneous activity coming from the retina (see below).

ODC segregation can be blocked by eliminating all ascending activity in the visual pathway, achieved by silencing RGC activity intraocularly with the sodium-channel blocker, TTX(67). If the optic nerves are then stimulated electrically, the normal shift from initial binocular innervation of cortical neurons to the emergence of monocularly driven cells can be achieved (as assessed physiologically), indicating that neural activity drives this process(61). However, electrical stimulation only induces ocular segregation at the cortical level when it is applied to the two optic nerves in an alternating (asynchronous) manner. If the optic nerves are stimulated synchronously, cortical neurons remain binocularly innervated and segregation does not occur. This stimulation regimen does not alter the eye-specific LGN layers that have already formed. Similarly, recent experiments demonstrate that synchronous stimulation of RGC axons can disrupt the development of precise orientation tuning in the visual cortex(68). Because orientation tuning also depends on precisely patterned cortical connections, this experiment suggests a role for competitive, activity-dependent interactions in the development of yet another set of specific functional circuits.

There is also a well-studied requirement for action potentials and synaptic transmission in the development of the visual system of cold-blooded vertebrates. Neural activity is required for both the formation of the topographic map in the tectum (the equivalent of the superior colliculus in mammals) and the segregation of eye-specific stripes that can be experimentally induced in frog optic tectum. [The "3-eyed" frog model is produced by grafting a third eye that sends its RGC axons to the tectum, normally monocular in frogs, such that the axons from the two eyes now innervating a single tectum segregate into eye-specific stripes(6971)]. Blockade of activity prevents refinement of the retinotectal map from an initially coarse projection and desegregates the eye-specific stripes in the tectum of the 3-eyed frog, which require activity to be maintained even after they have formed because the frog tectum continues to grow throughout life(71,72). Again, experiments show that it is the temporal pattern of neural activity that is critical for the process of axonal refinement. Rearing goldfish or 3-eyed frogs under strobe-lights, which synchronizes the activity of all ganglion cells, prevents topographic map refinement in the fish and eye-specific stripe formation in frogs(73,74). Taken together, these observations demonstrate that not just activity per se, but specific aspects of the timing of electrical activity, is necessary for this axonal segregation. They suggest that the inputs from both eyes compete for connections in the cortex on the basis of correlations in their firing that allow axons coming from the same eye to gain territory together. There is increasing evidence that synapse formation on the basis of spatiotemporally correlated activity may be a general mechanism used elsewhere in the nervous system to establish precise connections(1,4,75).

Critical periods for competition. The time when normal patterns of neural activity are necessary for the formation of the adult pattern of connections is called the "critical period" for the development of those connections. In humans, one example of activity-driven rearrangements occurring during a critical period is the dramatic loss of functional vision (e.g. amblyopia) in one eye that occurs when a child younger than 18 mo with a cataract goes untreated for more than 1 to 2 wk(76,77). Even temporary disruption of appropriate sensory input can permanently affect cortical patterns of connections if it occurs during the appropriate critical period, as Hubel and Wiesel's studies of monocular visual deprivation in kittens demonstrated(56,57). The plasticity of connections inherent in the critical period is followed by relative stability in the mature pattern of connections. This stability of adult neural connections is seen when a cataract forms in an adult. In contrast to the infant with a cataract, normal vision is restored in the adult who has the cataract removed, even if it has been present for many years, because the neural connections in the mature visual pathway do not change(76).

The synaptic plasticity present in the developing nervous system endows it with the ability to adapt to the many variations of the external world, as in language acquisition, or the ability to recover from early damage when one region of cortex subsumes the functions of a damaged area. However, this plasticity also leaves the developing nervous system uniquely vulnerable to injury from abnormal patterns of activity (see below). In the human, this period of neuronal vulnerability most likely extends from the end of the second trimester of gestation well into childhood(5,17,78).

NEURONAL ACTIVITY: SPONTANEOUS AND SENSORY-DRIVEN

Requirement for spontaneously generated activity. How early in development is neural activity needed for the refinement of connections? As mentioned above, in the binocular eye-closure experiments, ODC still form, and the source of the activity that drives this process is presumed to be the retina, which is spontaneously active in the dark(65,66). There is now strong evidence that even before the development of vision (before retinal photoreceptors develop), the RGC are spontaneously active (see below; 79-83), and this activity may be transmitted through the entire developing visual pathway(84). This observation can explain the fact that in primates, before visual experience, orientation columns and ODC are already forming in utero(46,64,85). The hypothesis that this spontaneous activity drives the initial establishment of ODC and orientation columns is currently being investigated intensively.

In addition, this spontaneous activity is required for the earlier formation of patterned visual system connections. Recall that before ODC form, RGC axons segregate to form the eye-specific layers in the LGN (Fig. 1B). This process occurs entirely during the period before photoreceptors mature, yet it depends on neural activity. Infusion of TTX into the LGN, which blocks both presynaptic and postsynaptic sodium action potentials, blocks this layer formation(23). Individual axons in TTX-treated LGN branch widely, rather than retracting inappropriate branches and growing selectively into their appropriate eye-specific layers(24). In recent experiments, we selectively blocked only the RGC activity(86). When activity was blocked binocularly, layers also failed to form, indicating a requirement for the presynaptic action potentials (Fig. 1C). But when the spontaneous retinal activity was blocked only in one eye, the projection from the active retina expanded greatly into territory normally belonging to the other eye, and the projection from the inactive retina was substantially reduced (Fig. 1D). Thus, as in the later occurring formation of the ODC, interocular competition-this time driven exclusively by endogenous retinal activity-determines the pattern of connections. These experiments demonstrate that spontaneous activity can produce highly stereotyped patterns of connections long before the onset of visual experience. They also imply that disruptions in the competitive mechanisms of axonal segregation, either prenatally or postnatally, may result in profound disruptions in the appropriate patterning of neural connections and lead to neurologic dysfunction.

Patterns of spontaneous activity. It is not only neuronal activity that is required to drive the refinement of connections, but also information contained in the specific patterns of this activity-the spatiotemporal correlations-that appears to be important(1,87,88). The nervous system generates complex patterns of activity long before there is any patterned sensory input(75). For example, ex utero microelectrode recordings from fetal rat retinas indicated that cells in the ganglion cell layer are spontaneously active and fire together long before there is any visual input(79,80). When hundreds of neurons in developing retinas were then recorded simultaneously in vitro, either on a multielectrode array or using optical recording techniques(8183), a particular spatiotemporal pattern of firing was revealed. Individual ganglion cells fire bursts of action potentials of 2-8 s in duration, separated by extended periods of quiescence 40-90 s long(82). Measurements from groups of cells show that neighboring cells fire action potentials together and undergo increases in levels of intracellular calcium synchronously(8183,89). On a larger spatial scale, the pattern of activity resembles a "wave" that travels across local regions of the retina at about 100-300 µm/s, and involves cells situated within approximately 300 µm of each other(82,89). These waves "tile" the retina with highly restricted domains of activity whose borders change over time(83,89). Significantly, this highly correlated spontaneous activity is only present during the period of retinogeniculate axon terminal segregation (first, into eye-specific layers and second, into even more refined sublayers, such as the ON- and OFF-system connections). The waves are gone by the time photoreceptors are present and visual input is available. After this early period, RGC are spontaneously active in the dark, but do not participate in waves of activity(65,82).

The early pattern of spontaneous retinal activity is well-suited to drive segregation not only because it is present during the appropriate period, but also because of these particular spatiotemporal characteristics. The short duration of the activity compared with the long intervening periods of silence makes it unlikely that presynaptic cells from the two eyes will be active simultaneously (Fig. 2A). Spurious correlations are unlikely because the location, timing, and direction of wave spread are highly variable within a single piece of retina(8183,89). Within a retina, correlations in the timing of the bursts are stronger between neighboring retinal neurons than for distant neurons. As mentioned earlier, lasting increases in synaptic strength are thought to require that presynaptic inputs be sufficiently correlated so that there is an overlap in postsynaptic response. For synaptic weakening, the inputs from the two eyes onto a single postsynaptic neuron should be significantly uncorrelated, resulting in the weakening of one of the two inputs (Fig. 2A). Thus, the correlations created by the retinal waves could underlie cooperative synaptic strengthening thought to be necessary to group monocularly driven cells into eye-specific layers in the LGN and, at slightly later ages, into ODC.

How are these correlated patterns of activity generated? In the retina, recent experiments point to a role for synaptic transmission(83). During the period when the waves are generated, the retina contains differentiated RGC and amacrine cells (a class of retinal interneurons). Both of these cell types participate in the waves(83,90). A synaptic circuit between RGC and the cholinergic amacrine cells is necessary for the generation and the propagation of the waves; blockade of cholinergic transmission blocks the waves(83,86). These results emphasize the extent to which even very immature circuits, in which few classic synaptic specializations are visible(91,92), can produce sophisticated patterns of spontaneous activity.

These spontaneous retinal waves are present in mammals (mouse, ferret, and cat;(81,82,93,94), as well as in turtle and chick(95,96)), emphasizing that this pattern of activity may be a well-conserved mechanism for refining neural connectivity before vision. Moreover, the generation of spontaneous, correlated activity is not limited to the visual system. Spontaneously generated activity has been described in a variety of locations in the developing CNS, including cortex, hippocampus, cerebellum, thalamus, superior colliculus, locus ceruleus, spinal cord, and cochlea(75,97). It varies in pattern, described as "waves, domains, and oscillations"(75), but each of these patterns provides strong spatiotemporal correlations that may shape synaptogenesis. For example, in immature cortical slices, domains of neurons that tangentially span the thickness of the cortical layers and horizontally are within a 50- to 200-mm circular diameter undergo synchronous calcium increases(98100). Unlike the spontaneous activity in the retina, this pattern of activity appears to be coordinated not by synaptic connections, but rather by gap-junctional networks that transmit a chemical signal (inositol 1,4,5-trisphosphate), instead of creating classic electrotonic coupling between the neurons(99101). It has been suggested that these cortical domains of activity help to shape the columnar structure of cortical connections(98,102). Another neural tissue that is endogenously active is the developing spinal cord, which appears to combine the use of synaptic and gap-junctional connections to generate patterned activity. Synaptic connections appear to underlie the spontaneous, oscillating bursts of activity that occur in dispersed groups of motoneurons that will innervate either flexor or extensor muscles(103). There also is extensive gap-junctional coupling in the spinal cord between motoneurons that innervate homonymous muscles(104). The spontaneous, coordinated oscillations of activity seen in developing motoneurons may drive the activity-dependent competition known to refine synaptic connections at the NMJ(105,106). The developing retina, cortex, and spinal cord illustrate the variety of neural circuitry-based on immature synapses, gap junctions, or a combination of the two-that can generate the correlated, spontaneous activity that may drive the activity-dependent refinement of connections throughout the developing nervous system.

Because spontaneous, highly patterned activity occurs in many areas of the nervous system, disruption of particular patterns of endogenous activity could result in abnormalities in neural circuitry. For example, damage could result from prepartum use of drugs that interfere with synaptic transmission (i.e. nicotine) or change gap-junctional connections (i.e. alcohols). Large-scale structural brain defects would not be expected to result from these exposures; rather, changes in the patterns of connections might result in subtle deficits, such as the mild cognitive impairments documented in prenatal nicotine or alcohol exposure(107112). Although it is difficult to determine whether such deficits are simply the result of global insults from these agents (i.e. reduced oxygen delivery to the fetus or carbon monoxide poisoning), it is worth considering the specific changes in neural connections that may occur when the patterns of endogenous, correlated activity are disrupted.

Requirement for sensory-driven activity. Once sensory input becomes available, features of the external environment drive the activity that shapes connections. The shapes and forms present in our visual world provide multiple sources of correlated activity; just the edge of this page of paper can correlate the firing of many RGC across which the linear image falls simultaneously. Disruptions in these correlations, as discussed earlier, can lead to profound anatomic and functional changes in connectivity within the visual cortex during the appropriate critical periods(37,56,57). In addition, extreme manipulations of the visual environment, for example, raising kittens in an environment containing only vertical lines, preferentially selects for cortical neurons responding to vertical stimuli and leaves the rest of the cortex unresponsive(113,114). In many other developing systems, neuronal receptive field properties and the underlying connections are also sensitive to alterations of sensory input (in the somatosensory system see(115), in the auditory system see(116,117). The ability of the developing brain to incorporate information from the external world into its precise circuitry allows it to adapt to a myriad of changing environments.

The use of these two sources of activity-endogenously produced and sensory-driven-should not be viewed as occurring in mutually exclusive periods. For example, studies of language acquisition demonstrate that infants are born with preferences for particular sounds present in their mother's language(118). This observation suggests that before birth (when spontaneous activity plays a dominant role in shaping connections) circuitry may also be modulated by auditory input the filters into the uterus. After birth, endogenous activity may still play a role in shaping connections. After the onset of vision, the local circuitry between the LGN and the surrounding reticular nucleus generates spontaneous, correlated waves of action potentials in the form of thalamic spindle oscillations that travel in waves of spike activity across the LGN(119121). These oscillations occur during sleep, and therefore do not normally disrupt the processing of sensory experience. In the normal brain, this activity could help shape thalamocortical connections because it creates highly synchronized inputs from neighboring LGN neurons. On the other hand, the mechanism that generates thalamic spindles during sleep also appears to underlie the classic 3-Hz spike-and-wave seizure discharges seen in children with absence epilepsy(121). Because the mechanisms that shape the normal connectivity are tuned to be sensitive to highly correlated activity, it is possible that these highly synchronous epileptic discharges could generate abnormal neuronal connectivity by interrupting the normal pattern of sensory input. More generally, the highly synchronous activity of seizures may, in part, be responsible for the progressive nature of some epileptic syndromes (e.g. Lennox-Gastaut;(122123)). The interplay between these two forms of activity-spontaneously generated and sensory-driven-is most likely the result of shared mechanisms underlying activity-dependent synaptic competition.

MECHANISMS OF ACTIVITY-DEPENDENT COMPETITION

Physiologic and Structural Synaptic Changes

What mechanisms operate so that correlated activity can drive the refinement of synaptic connections? The process is competitive in the sense that direct or indirect interactions between incoming axons for common postsynaptic targets drive the retraction of all but one (or one class of) axon and allows expansion and stabilization of the remaining axonal terminals. Ideas about how this process occurs on a cellular and molecular level in the developing CNS are drawn from primarily two sources: studies on synapse formation and elimination at the NMJ, and studies on long-term changes in synaptic strength in the hippocampus(105,106,124126). Although the details of these systems differ, they share a common theme: synaptic refinement depends on a mechanism in which synapses that are synchronously active with the postsynaptic cell are reinforced ("Hebbian" synapses) whereas those synapses that are not synchronously active are eliminated (Fig. 2A)(105,125,127).

The individual steps of synaptic refinement have been best studied at the NMJ(105). As in the visual system, axonal projections-in this case from mononeurons-are initially diffuse, with individual muscle fibers receiving innervation from multiple motoneurons before birth. These connections are then refined through an activity-dependent process(105) so that each muscle fiber receives input from only one axon [possibly driven initially by the spontaneous patterns of activity generated by the developing spinal cord as mentioned above(128,129)]. The process of synaptic refinement at the NMJ has been monitored elegantly in vivo(130132). Neurotransmitter receptor sites on the postsynaptic membrane (i.e. the muscle's acetylcholine receptors) first decrease, the presynaptic axon's terminals are eliminated from the muscle, and finally the nerve terminal resorbs into the main axonal trunk(131). Asynchronous receptor activation seems to be the signal that allows the muscle to destabilize particular synaptic sites: when focal postsynaptic blockade is produced by application of α-bungarotoxin (which blocks only the postsynaptic acetylcholine receptors, and not presynaptic transmitter release) then these blocked synapses are selectively eliminated(130). Elimination of the blocked synapse only occurs when there is activity at the rest of the junction, suggesting that active synaptic sites are stabilized and that this activity must somehow destabilize inactive sites. Studies of neurons and myocytes in tissue culture show similar results. Stimulation of one innervating neuron can suppress transmission from the synapse of a second, inactive neuron(133,134), whereas simultaneous stimulation of both neurons either results in strengthening of transmission at both synapses or no change(135). These results led to the proposal of the following scenario: when synapses are active they are somehow protected from elimination, but when they are inactive they receive a "withdrawal" signal(105,136). When neither input is active, both remain stationary because no withdrawal signals are present; when both axons are simultaneously active, they are both protected from elimination (Fig. 2B). This dependence on the balance of activity, not on activity per se, is strikingly similar to the requirement for patterned activity discussed above for the developing visual system.

Cellular and Molecular Synaptic Changes

Coincidence detection. Although the precise molecular mechanism that allows detection of coincident activity at the NMJ remains obscure, the cellular and molecular correlates of changes in synaptic strength have been studied in detail in the mature hippocampus. In area CA1 of the hippocampus, the application of high-frequency electrical stimulation leads to lasting increases in synaptic strength, known as LTP. In contrast, lower frequency stimulation can cause a lasting decrease in synaptic strength, LTD. LTP is an appealing mechanism for synaptic modification during development because it requires cooperativity (a certain number of synapses must be active to induce LTP), associativity (the coincident firing of a weak synapse along with stronger synapses on a convergent input will potentiate both sets of synapses), and input-specificity (only synchronously active inputs are potentiated). All of these criteria are thought to operate at developing synapses as well(125,126,137,138). Additionally, LTP and LTD can occur homosynaptically (at the same synapse that is stimulated) and heterosynaptically (at other synapses on the same postsynaptic cell that is being depolarized by the active synapse), allowing for the strengthening and weakening of multiple synapses, as is necessary in the refinement of developing connections (Fig. 2)(125,126,139141). These attributes have led to a number of experiments examining the possible role of LTP and LTD as mechanisms underlying synaptic refinement during development.

In the developing visual system, both retinogeniculate and geniculocortical synapses demonstrate forms of synaptic enhancement similar to hippocampal LTP(94,137,142). Developing retinogeniculate synapses can undergo long-term increases in transmission with high-frequency stimulation or pairing of presynaptic and postsynaptic activity(94). In the developing neocortex, LTP has been demonstrated directly at the geniculocortical synapse(142), as well as in other cortical layers(143146). LTD has also been described in the neocortex(147,148). Both LTP and LTD are easier to elicit at synapses during the period of synaptic refinement in the cortex(142,147,149). The link between activity-dependent synaptic refinement and the mechanisms of LTP and LTD currently remains one of correlation rather than causation; however, the molecular mechanisms underlying LTP and LTD are presently the best-understood cellular correlates of activity-dependent synaptic changes.

One molecular requirement that is necessary for LTP and LTD is physiologic "coincidence detector," a mechanism by which coincident activity of the presynaptic and postsynaptic neurons is detected (Fig. 2B). The glutamatergic NMDA receptor is a molecular coincidence detector that has been studied in great detail in the CA1 region of the adult hippocampus(125,150152). This receptor can detect coincident activity because it fluxes Ca2+ ions only when glutamate, released by the active presynaptic neuron, binds to the receptor and the postsynaptic neuron is depolarized simultaneously, relieving a voltage-dependent Mg2+ gating of the channel pore. When NMDA receptors are blocked, LTP and LTD cannot be elicited at the CA1 synapse(125,150). The level of Ca2+ influx into the postsynaptic neurons appears to determine whether LTP (high Ca2+) or LTD (low Ca2+) occurs(153155).

Like the hippocampal synapse, the retinogeniculate and thalamocortical synapses use glutamate as their neurotransmitter, which has lead to an investigation of which glutamate receptors are critical for synaptic refinement during visual system development. At the retinogeniculate synapse, physiologic synaptic enhancement appears to depend on NMDA receptor activation at least in some neurons(94), especially at later stages of development. In vivo infusion of NMDA receptor antagonists (such as MK801) into the ferret thalamus at appropriate ages can prevent retinogeniculate axons from segregating into the ON- and OFF-sublayers of the LGN(156), although other NMDA receptor antagonists (such as 2-amino-5-phosphonovaleric acid) do not block the earlier occurring formation of eye-specific layers(157). In the developing cortex, in vivo infusions of NMDA receptor antagonists block the physiologic shifts in ODC produced by monocular blockade [although the specificity of NMDA receptor blockade in the cortex has been called into question because these antagonists also block most activity in developing cortical neurons(158)]. The requirement for NMDA receptor-mediated plasticity is perhaps best defined for the optic tectum of lower vertebrates. In the 3-eyed frog model, the eye-specific in the tectum not only desegregate when all activity is blocked with TTX, as discussed earlier, they also desegregate when NMDA receptor antagonists are applied(159). However, although NMDA receptor-mediated plasticity has been most thoroughly studied, there are many examples of plasticity mechanisms that do not require NMDA receptors, e.g. LTP in the hippocampal mossy fibers in CA3 and LTD in the cerebellum(125,160). Indeed, it has recently been found that some forms of AMPA receptors (the other major class of ion-fluxing glutamate receptors), can flux not only Na+, but Ca2+ as well, which may allow them to participate in changing synaptic strength in some locations(161163). Thus, the lack of dependence on NMDA receptors in some areas of the visual system that do undergo activity-dependent refinement(84) points to the likely presence of other mechanisms of coincidence detection that will be critical for activity-dependent synaptic refinement.

The functioning of particular mechanisms for coincidence detection, such as a reliance on glutamate receptors, may increase the susceptibility of the developing brain to particular insults. In addition, recent experiments suggest that the earliest synapses that form during the development or rearrangement of glutamatergic circuits may contain solely NMDA receptors (optic tectum:(164,165); hippocampus:(166,167)). In the hippocampus, such synapses have been dubbed "silent synapses" because they are undetectable at resting membrane potentials, but the NMDA receptor-dependent excitatory postsynaptic potentials are revealed when the postsynaptic cell is depolarized(166). After LTP induction protocols, AMPA receptors appear at these synapses so that they are no longer functionally silent. It is possible that the severe disruption caused by glutamate or glycine toxicity in the developing brain is caused not only by acute cell death, but additionally by the disruption of the first step in the refinement of synaptic connections. Exacerbating this disruption may be the role that presynaptic activity can play in the regulation of postsynaptic ion receptor subtypes, which control ion flow and neuronal excitability(168,169). Changes in the properties of the glutamate receptor subunit composition during development also allows for increased calcium flux through these receptors(162,170,171). Neurons with glutamatergic synapses may be prone to death by toxicity because of excessive Ca2+ entry after the release of large amounts of glutamate during a hypoxic-ischemic episode, through activation of either NMDA receptors or AMPA receptors(162,163,172174). Similar damage can also follow the release of glycine (which at high levels can stimulate NMDA receptors) as seen in genetic disorders such as nonketotic hyperglycemia(175). The specific physiology of the developing circuitry is presumably intimately related to morphologic changes that are later seen in the brain, both in normal development and after injury.

Retrograde signals. Coincidence detection allows for anterograde information flow about correlated activity between the presynaptic and postsynaptic neuron (Fig. 2B). For this activity to result in a structural change in presynaptic connections during synaptic refinement, it seems there must also be a retrograde signal that allows communication back from the postsynaptic neuron to the active presynaptic terminals (Fig. 2B). Such a signal would need to be regulated by neuronal activity and cause selective stabilization and growth at synapses that are simultaneously active with the postsynaptic neuron. Again, studies of hippocampal LTP and NMJ synaptic refinement have provided possible candidates for retrograde messengers that could act in the activity-dependent refinement in the developing brain. At the NMJ, possible retrograde messengers that have been implicated in synaptic refinement include neurotrophins (particularly BDNF and NT-3)(176), NO(177), ATP(178), and proteases (for detailed reviews see(105,179)). In the hippocampus, again both neurotrophins (BDNF and NT-3) and NO can potentiate CA1 synapses(180182); antagonists of the neurotrophins and of NO can block LTP induction, and transgenic mice that lack BDNF or the ability to make NO have a diminished capacity for hippocampal LTP(180,181,183). These experiments have lead to the investigation of the role of these compounds in activity-dependent development of the mammalian brain.

There is evidence accumulating that neurotrophins are good candidates for molecules that modulate activity-dependent developmental plasticity(184,185). The receptors for BDNF and NT-4 (TrkB receptors) are expressed in mammalian LGN and cortex during the appropriate developmental period (TrkC receptors, which bind NT-3, are also expressed;(186,187)). Infusion of BDNF and NT-4 (but not other neurotrophins such as nerve growth factor or NT-3) into the visual cortex during the period of ODC development blocked ODC in the region of the infusion(188); blockade of the TrkB receptor also blocked ODC formation(189). Putting NT-4, but not BDNF or NT-3, into the visual cortex can also rescue LGN neurons from the shrinkage of their cell bodies that occurs with monocular deprivation(190). These experiments suggest that the competitive interactions that normally underlie synaptic remodeling are based on a limited supply of neurotrophins, and infusion of excess neurotrophins acted on the LGN axon terminals, removing this competition. When there is no activity, as in monocular deprivation, then the application of NT-4 can maintain the silenced LGN neurons, although it would be predicted that their terminals would still not have been able to segregate into ODC. Whether BDNF or NT-4 function endogenously during development has yet to be determined.

Although it seems likely that the neurotrophins may play a permissive role, allowing activity-dependent changes to occur, they are unlikely to function on the 10- to 100-ms scale on which coincidence detection must occur because the Trk receptor is likely to remain activated for a prolonged period once neurotrophin binds to it(184,191). NO is a more rapidly acting candidate that could function rapidly to change synaptic strength during development. Experiments in chick tectum(192) and ferret LGN(193) indicate that systemic blockade of NO synthesis can block refinement of some connections, but these effects are quite restricted.

It is likely that many molecular messengers play critical roles in activity-dependent development and are yet to be identified. In the future, it will be necessary to balance potential therapeutic effects of molecular messengers, such as the increased neuronal survival seen when BDNF is administered after an hypoxic-ischemic insult(194), against the potential risk of creating abnormal neural circuitry.

CONCLUDING REMARKS

Recent studies are beginning to forge a link between normally occurring activity-dependent synaptic modifications and specific neurologic disorders. Two examples are the reorganization of hippocampal connections after prolonged epilepsy(195)[x] and the activity-dependent translation of fragile X mental retardation protein at synapses(196,197). In the hippocampus, rearrangement and sprouting of mossy fiber connections is a striking feature of temporal lobe epilepsy; current work is focusing on the similarity of the mechanisms that appear to underlie activity-dependent formation of hippocampal connections during development and the pathologic sprouting seen after seizures(195). The link between the expression of the fragile X mental retardation protein and its role in normal development remains unknown, but its expression appears at least to be regulated by neural activity and could play a role in the initial formation of neural circuits that might in turn be disrupted when a pathologic form of the protein is produced(196199). Information about the role of activity-dependent developmental changes in neurologic disease is just starting to emerge, but understanding normal activity-dependent developmental processes at the cellular and molecular levels should provide clues to human neurologic illness that may result from changes in the circuitry of the developing nervous system.

Activity-dependent development can be disrupted at many points. We have indicated a critical role for correlated activity, both spontaneous and sensory-derived, in driving neuronal competition that shapes precise connectivity. Either subtle disruptions in the spatiotemporal patterns of activity or major disruptions in the overall balance of activity may result in abnormal connectivity. Such disruption of patterned neural activity may result from interference with synaptic transmission, with coincidence detection, or with retrograde signals, as might happen with prepartum use of drugs (such as nicotine, benzodiazepines, or narcotics) or postnatal exposure to a variety of medications. In addition, as different sensory systems become active at particular developmental stages in premature infants, it is worth considering what effect the sensory environment might have on neural activity patterns needed to drive synaptic remodeling. During critical periods, specific areas of the developing nervous system may be particularly susceptible to these types of disruption.

Disruptions of activity would affect the fine-tuning of neural circuits, a process that requires neural activity to shape the adult patterns of connectivity, not the large-scale wiring of the brain, which generally occurs independent of neural activity. It is all too common for a child to have neurologic deficits when no structural abnormalities can be identified. The current techniques in neuropathology and neuroradiology are just beginning to be able to detect the critical changes that we know must be occurring. Consequently, the major challenge for the clinician and medical scientist is to develop techniques with high resolution-such as the newer methods of noninvasive functional imaging-that can monitor the normal or abnormal functioning of neural circuits on a fine-scale during development. Once disorders resulting from disruptions in activity-dependent development can be more accurately identified, we should be able to design therapies aimed at correcting the abnormal circuits (for example, by developing treatments to extend critical periods) or optimizing the use of the abnormal circuitry (as is being tried for specific language impairments;(200,201)). Recognizing and understanding in detail the role of activity-dependent development in the formation of precise neural circuits should expand greatly our ability to identify, treat, and prevent many neurologic disorders.