The formation of cooperative cell assemblies in the visual cortex

Higher mammals and humans, which have frontally positioned eyes with overlapping visual fields, can fuse the images of the two eyes and compute from their differences the distance of objects in space. The basis for this function are neurons in the visual cortex which possess two receptive fields, one in each eye, that are tuned to the same features and whose relative positions on the two retinas are precisely defined. Thus, during development the afferents arriving from each eye must be arranged in a highly selective way so that only those pairs of afferents which originate from retinal loci with similar interocular disparities converge onto individual cortical target cells (see Fig. 1). In principle this requires the establishment of a precise match between two receptor surfaces; i.e. both the neighborhood relationships of ganglion cells within the same retina and their relative interocular disparities have to be represented in a single map at the level of the striate cortex.

By analogy with other developmental processes in the brain, one might consider specific recognition molecules as a solution for this specification problem. However, in this case, there is a principal limiting factor to the degree of selectivity that can be achieved with chemical markers: there is no way to predict with any great precision which retinal loci will actually correspond in the mature visual system. Retinal correspondence depends on parameters such as the size of the eyeballs, the position of the eyeballs in the orbit and the interocular distance. Clearly, these parameters are strongly influenced by epigenetic factors. Moreover, they continue to change as long as the skull grows. In principle, it follows that positional markers alone, even if they were quantitatively sufficient, could not suffice to determine with the required precision the pattern of interocular connections. An elegant possibility exists, however, to identify fibers as coming from retinal loci of similar disparity by evaluating their responses to contours. When a two-dimensional pattern is fixated with both eyes, its contours stimulate corresponding retinal loci. When the target is a three-dimensional object, its outlines also stimulate retinal loci with different disparities, whereby any point on the object produces responses at sites which have similar disparities. Thus, responses of afferents from retinal loci with similar disparity are elicited by the same parts of the object and, therefore, are likely to be more similar than those of afferents coming from non-corresponding loci. What is required, then, is a developmental mechanism capable of selecting those retinal afferents for convergence onto common cortical target cells which convey correlated activation patterns.

Use-dependent selection of circuits also appears to be required at higher levels of cortical processing to allow for the development of network properties required for scene segmentation. Here intracortical connections have to be specified that link spatially distributed cortical neurons with each other. Again, the postulation is that those connections that convey correlated activity should be stabilized.

The retinal image of a visual scene consists of a two-dimensional continuous distribution of grey levels. To identify particular figures or objects it needs to be determined which of the local luminance gradients belong to particular objects and which to the embedding background. Some grouping must be performed to associate these luminance distributions with the contours of a single object and to segregate signals from objects with overlapping contours from each other and from the signals generated by the background. These operations are commonly known as scene segmentation or figure-ground segregation. Because most of them are usually carried out subconsciously and do not require the direction of selective attention to particular features of the scene, these operations are called ‘preatten-tive visual processes’ or ‘early visual processes’ (for a review and examples, see Julesz, 1971; Marr, 1976; Treisman, 1986; Ramachandran, 1988).

Important criteria for grouping are spatial contiguity and coherence in particular feature domains. The visual system interprets luminance gradients as originating from the same object or the same figure if they are closely spaced and especially if they are continuous.The same is true for spatially separate luminance gradients if they share similarities within particular feature domains. In the simplest case, objects are distinguished from other objects and background if they have a continuous outline. However, even when an object is partly obscured by another object, its spatially separate contours may still be interpreted as coming from one and the same object if they are sufficiently similar within one or more feature domains. Thus, if the visible parts of an object have the same spectral composition or the same retinal disparity (distance from the observer), they will still be interpreted as belonging to a single object.

A particularly powerful criterion according to which interrupted contrast borders are interpreted as belonging to the same object is colinearity. Discontinuous but colinear contrast borders tend to be interpreted as the interrupted outline of the same figure rather than as independent borders.

Another very powerful criterion is coherence of motion. Spatially distributed contrast borders are interpreted as belonging to the same object if they move with the same speed in the same direction. The following perception experiment exemplifies this. If one produces on a television display a cloud of dots of identical size and luminance which move with the same speed but in randomized directions, and then suddenly makes those dots that happen to be located on the perceived outlines of a triangle move in the same direction, one perceives these dots as the outline of a moving shape, in this case the triangle. Apparently, the visual system interprets those dots that share a common feature, in this case the same movement vector, as belonging together, and segregates the assembly of these coherently moving dots from the cloud of randomly moving dots. The former are interpreted as components of a figure, and the latter as components of the background.

These psychophysical data show that the visual system is capable of establishing relationships between the responses of spatially distributed feature detectors and can identify the responses to those features which have some coherent properties or a common fate. An important constraint of this evaluation process is that topological information must be strictly preserved. After the responses resulting from figures have been isolated, the figures need to be identified, which requires that the spatial relationships of their contours are accessible. A mechanism is needed, therefore, which allows one to establish relationships between spatially distributed feature detectors without losing positional information.

One solution for the detection of coherent features in a scene is to let feature detectors interact with one another and to provide a mode of interaction such that responses of detectors encoding coherent features become distinguishable. This can be achieved by selectively coupling neurons with similar feature preferences through reciprocal connections.

The effect of such connections is that they establish temporal correlations among the responses of feature detectors which respond to similar features in a scene. Thus, if an image contains a figure that is distinguished by the presence of similar features, the corresponding feature detectors in the visual cortex become activated and, owing to selective interactions, their responses are correlated and so become distinguishable from those of neurons encoding non-coherent features (Gray and Singer, 1989; Gray et al. 1989; Singer, 1990).

The postulation that feature detectors with particular preferences need to interact selectively with each other implies an extremely complex network of connections between spatially distributed neurons in the visual cortex. In order to extract the property of colinearity, for example, selective connections must be implemented between neurons which have the same orientation preference and whose receptive fields are aligned colinearly. Because of the columnar organization of the striate cortex, such neurons are distributed in spatially separate columns. Thus, it is not possible simply to connect nearest neighbors. Since the arrangement of iso-orientation domains in the striate cortex is not very regular, and in some species even shows marked anisotropies, it is also not possible to define a simple rule for these selective connections, such as that only neurons that are a fixed distance apart should be interconnected (Fig. 2). The same holds true for connections between neurons that code for the same direction of motion or for the same color, etc. This raises the problem of how such selective connections can develop.

Selection according to functional criteria provides one solution. This is particularly attractive in the present case, since the feature detectors which must be selectively coupled can be identified with high probability on the basis of then- responses. If, for example, an object moves across the visual field, movement detectors with similar preferences for speed and direction will become activated simultaneously by the moving contours of the object. The same holds true for the assembly of orientation detectors whose receptive fields are aligned colinearly if a straight contrast border is present, or for detectors of a particular disparity if contours are present at a certain distance. It would thus be sufficient to provide initially redundant sets of connections which link in an unselective way the various subpopulations of feature detectors and then to stabilize selectively connections between neurons which are often activated simultaneously. Such a process would be advantageous for two reasons. First, it would greatly economize on the genetic instructions. Second, it would ensure that selective coupling is established preferentially between feature detectors whose combined properties are matched by actual and frequent constellations of features present in the physical environment.

In the following paragraphs developmental results will be summarized which demonstrate that there are mechanisms that allow for activity-dependent selective stabilization of connections in the striate cortex. The development of binocular connections will be reviewed to illustrate the principles, and then the extent to which these use-dependent selection processes can be generalized to intracortical connections will be examined.

As D. H. Hubel and T. N. Wiesel have demonstrated in their pioneering studies, the connections between the two eyes and the cells in the striate cortex are malleable during early postnatal development and are subject to use-dependent modifications. For example, if one eye is deprived of contour vision while the other is allowed to view normally, the large majority of cortical cells rapidly lose the ability to respond to the deprived eye (Wiesel and Hubei, 1965; for a review, see Frégnac and Imbert, 1984).

These use-dependent modifications of ocular dominance do not solely depend on the level of activity in the afferents from the two eyes (Cynader and Mitchell, 1977; Singer et al. 1977; Wilson et al. 1977; Greuel et al. 1987). Other critical variables include the state of activation of the postsynaptic neuron and, in particular, the degree of temporal correlation between pre- and postsynaptic activation (Rauschecker and Singer, 1979, 1981; Frégnac et al. 1988; Greuel et al. 1988). The use-dependent modifications of excitatory transmission seem to follow rules which closely resemble those postulated by Hebb (1949) and Stent (1973) for adaptive neuronal connections (Rauschecker and Singer, 1981). As summarized in Fig. 3, the direction of the change - towards an increase or decrease of efficacy - depends on the correlation between pre- and postsynaptic activation. Connections stabilize if the probability is high that the presynaptic afferents and the postsynaptic cell are active in temporal contiguity and they destabilize when the postsynaptic target is strongly activated while the presynaptic terminal is silent. These rules, when applied to circuits when two or more afferent pathways converge onto a common postsynaptic target cell, have the effect of selectively stabilizing, and hence associating, pathways that convey correlated activity. Likewise, these modification rules lead to competition between converging pathways if these convey uncorrelated activity. Eventually, the subset of afferents which has the highest probability of being active in temporal contiguity with the postsynaptic target cell will ‘win’.

Thus, according to these modification rules the converging pathways arriving from the two eyes will only remain connected to a common cortical cell if their activities are sufficiently correlated. Direct proof for this has been provided by Stryker and Harris (1986), who stimulated the optic nerves electrically and showed that only synchronous activation led to maintenance of binocular connections. If the responses from the two eyes are too asynchronous, the corresponding afferents compete with one another. In that case only one pathway consolidates and the other becomes repressed. The maximum interval of asynchrony still compatible with the maintenance of binocular connections was found to be of the order of 200–400 ms (Altmann et al. 1987).

The data reviewed above indicate that there are developmental mechanisms which are, in principle, capable of selectively stabilizing pathways conveying coherent activity. However, this selection can only be successful in optimizing binocular connectivity if it is restricted to moments when the animal actually fixates a non-ambiguous target with both eyes. Pruning must not take place when the visual axes of the two eyes are not properly aligned. If the images on the two retinas are too different and cannot be fused, all signals from the two eyes, including those originating from retinal loci with similar disparity, are uncorrelated. If selection occurred under such conditions, all afferents from the two eyes would compete with one another and the consequence would be complete disruption of binocular connections. The same would be the case if the spontaneously produced bursts of activity that occur, for example, in the geniculate afferents during certain sleep stages, were capable of inducing changes in circuitry. Moreover, prior to selection, the two eyes must be aligned to ensure an optimal match between the images in the two eyes. Thus, the direction of gaze must be adjusted until the activity patterns arriving from the two eyes are maximally correlated within the coarsely prespecified retinotopic representation. Only when these adjustments have been made can selection occur. To achieve its goal, selection must therefore be gated by non-retinal control systems which enable use-dependent modifications only when conditions are appropriate.

In agreement with this postulation it has been found that a number of non-retinal afferents to the striate cortex play a crucial role in gating ocular dominance plasticity. If these projections are inactivated or destroyed, retinal signals are no longer capable of inducing changes in binocular connectivity. The following projections have been identified as having a permissive role in ocular dominance plasticity: the proprioceptive afferents from extraocular muscles (Buisseret and Singer, 1983), the noradrenergic afferents from locus coeruleus (Kasamatsu and Pettigrew, 1979; Bear and Singer, 1986) and the cholinergic projection from the basal forebrain (Bear and Singer, 1986). Results of lesion studies (Singer, 1982) have further suggested that retinal signals only influence the development of cortical functions when the animal uses them for the control of behavior. Correspondingly, retinal signals do not lead to changes of cortical functions when kittens are paralyzed and/or anaesthetized while exposed to visual patterns (Buisseret et al. 1978; Freeman and Bonds, 1979; Singer, 1979; Singer and Rauschecker, 1982) or when the visual signals are manipulated in a way that renders them inappropriate for the control of behavior (Singer et al. 1982b).

The results reported above suggest that use-dependent modifications of synaptic transmission require a certain amount of cooperativity between retinal input and internally generated gating signals. This is supported by a number of studies indicating that the process mediating competitive disconnection has a threshold which is reached only when retinal signals are coincident with additional facilitatory input (Singer and Rauschecker, 1982; Greuel et al. 1988; Frégnac et al. 1988). Recently, a molecular mechanism has been identified that is likely to be responsible for this threshold. It was found that the activation of the N-methyl-D-aspartate (NMDA) receptor is a necessary prerequisite for the activity-dependent disconnection of the deprived pathway (Kleinschmidt et al. 1987; Gu et al. 1989; Bear et al. 1990).

A special property of this receptor-gated channel is that it is also voltage-dependent. It becomes permeable to Ca²⁺ only when the receptor is occupied by the synaptic transmitter and when the membrane of the postsynaptic neuron is sufficiently depolarized (Mayer et al. 1984). Thus, this synaptic mechanism is ideally suited to ‘evaluate’ the degree of contiguity of pre- and postsynaptic activation. Furthermore, this mechanism accounts for the evidence that the activation threshold for the induction of long-term modifications is high and is reached only if there is sufficient cooperativity between converging inputs. The fact that this channel, when activated, becomes permeable for Ca²⁺ makes it particularly well-adapted for the mediation of changes in synaptic transmission, because Ca²⁺ serves as a second messenger for the initiation of a variety of biochemical processes in nerve cells. Another interesting aspect is that the NMDA receptor mechanism is also involved in use-dependent synaptic plasticity in cortical structures of the mature nervous system such as the hippocampus (Collingridge and Bliss, 1987) and the visual cortex (Artola and Singer, 1987). Together with recent data that are reviewed below, this suggests that the use-dependent modifications of synaptic transmission, which presumably mediate learning in the adult brain, depend on similar mechanisms as experience-dependent self-organization of neuronal connectivity during development.

The mechanisms mediating use-dependent modifications of ocular dominance possess many of the features that have been postulated above for the development of assemblies of selectively coupled feature detectors. In this paragraph evidence is reviewed that is compatible with the hypothesis that similar self-organization processes occur at the level of feature representation. It has been known for a long time that one of the prominent features of cortical organization is the presence of an extremely dense network of far-reaching connections which are tangential to the cortical lamination (Szentagothai, 1973; Fisken et al. 1973). Electron microscopic (McGuire et al. 1985; LeVay, 1988) and electrophysiological (Luhmann et al. 1990a) data indicate that these connections are excitatory, originate predominantly from pyramidal cells and terminate preferentially on the apical dendrites of other pyramidal cells. These pathways are thus capable of mediating interactions between cortical neurons that are located in different columns. Furthermore, there are indications that these connections are selective, linking in a reciprocal way neuron clusters that tend to be spaced periodically (Rockland and Lund, 1983) and that share certain functional properties such as the same orientation preference and/or the same eye dominance (T’so et al. 1986; Gilbert and Wiesel, 1989; Gray et al. 1989; but see also Matsubara et al. 1985).

Developmental studies in the cat have shown that these tangential connections essentially appear postnatally, pass through a phase of exuberant proliferation during which they are particularly numerous and far-reaching, and subsequently become pruned (Price and Blakemore, 1985a, b;,Luhmann et al. 1990c; Callaway and Katz, 1990). This pruning occurs at a time when visual signals are readily available and appears to be influenced by retinal activity. If visual experience is unrestricted, subpopulations of these pathways are stabilized; if vision is prevented by dark rearing or binocular deprivation, elimination of tangential connections is initially retarded (Callaway and Katz, 1990) but subsequently enhanced so that eventually only a rudimentary network of horizontal connections is maintained (Luhmann et al. 1986, 1990c).

The anatomical indications that excitatory tangential connections are initially exuberant and imprecise and assume their selectivity through pruning is supported by physiological data. In kitten visual cortex excitatory interactions occur over much larger tangential distances and the receptive fields of individual cortical neurons are significantly larger than in adult cats (Luhmann et al. 1990a,b). Moreover, in kittens about 20% of the cells have additional ectopic receptive fields, that are excitatory, and can be located as far as 20° away from the center of the conventional receptive field. These ectopic fields occur mainly in cells located in supragranular layers where tangential connections are also densest and most far reaching (Luhmann et al. 1990b; see also Singer and Tretter, 1976). This age-dependent decrease in receptive field size, the laminar distribution of cells having an ectopic receptive field and the numerical reduction of such cells with age correlate well with the organization and postnatal pruning of tangential projections, suggesting a causal relationship. More direct evidence for this possibility comes from a study of kittens whose visual experience had been restricted to vertically oriented gratings of constant spatial frequency (Singer and Tretter, 1976). These kittens developed neurons with ectopic receptive fields that matched precisely the orientation and spacing of the grating. Such a result is expected from selective stabilization of connections between columns that are often activated simultaneously by the regularly spaced bars of the grating.

The development of the network of intrinsic tangential connections thus resembles in a number of aspects that of the connections between the eyes and their target structures in the visual cortex. Both continue to develop postnatally and achieve topological selectivity through a pruning process. For the connections from the eyes to the visual cortex, there is direct electrophysiological evidence that this pruning is guided by visual experience. For the intracortical connections this proof is still lacking, but there are indications from both in vivo and in vitro experiments (see below) that they are susceptible to use-dependent modifications which follow the same rules as the selection of binocular connections. This predicts selective strengthening of interactions between columns that are often activated together. Since cortical columns respond selectively to particular features, the architecture of the selected intra-cortical connections is expected to reflect the frequency with which certain constellations of features have occurred during early development.

Selective stabilization of tangential intrinsic connections would thus generate a non-topographically organized map which matches the coherent properties of ‘feature constellations’ in physical reality. The effect is similar to that of use-dependent pruning of binocular connections. The selective connections between the two eyes and common cortical target cells have the effect that the cortical cells become ‘detectors’ of a particular interocular disparity. As an analogy, selective connections between distributed clusters of feature detectors with related preferences have the effect that these, as an assembly, become a detector of coherent constellations of features.

Thus, by iteration of the very same processes of self-organization that, at peripheral levels of the visual system, increase the precision of topographic maps, it is possible to generate non-topographic maps which represent relationships in feature space. Once such maps are established, they can be used to detect coherences among features of visual scenes, as illustrated with the psychophysical experiment described above.

Use-dependent modifications of receptive field properties have been demonstrated in the visual cortex of adult animals, but only under rather extreme experimental conditions involving pharmacological treatment (Kasamatsu et al. 1979; Greuel et al. 1988), aversive sensory experience (Singer et al. 1982a) and grafting of glial cells (Müller and Best, 1989).

Recently, it has also become possible to induce use-dependent modifications of synaptic transmission in slices of the mature visual cortex. This has enabled a more detailed investigation of the properties of adaptivity in the adult cortex than was possible in vivo. Interestingly, in the adult, the synaptic modifications and the conditions for their induction resemble in many respects those of the experience-dependent changes that occur during development in vivo. As in the latter, they involve activation of NMDA-receptor-dependent conductances (Artola and Singer, 1987; Kimura et al. 1989) and their occurrence is facilitated by the neuromodulators noradrenaline and acetylcholine (Brocher et al. 1989). Because the activation conditions of pre- and postsynaptic elements can be controlled more readily in vitro than in vivo, the dependence of synaptic gain changes on the activation state of pre- and postsynaptic elements can be analysed in greater detail.

The modification rules resulting from this analysis are summarized in Fig. 4 Briefly, there are two different thresholds for synaptic modifications, both of which appear to depend upon the activation state of the postsynaptic target cell. If an input fails to reach either threshold there is no change in synaptic gain. If the first, lower threshold is reached by an active input, synaptic gain decreases at the activated synapses but there is no change at other, inactive synapses. The threshold for this homosynaptic depression mechanism is lower than that of the activation threshold of NMDA-receptor-dependent conductances. Accordingly, depression occurs even if NMD A receptors are blocked at the active synapses. If activation increases further, a second threshold is reached and the active input is no longer depressed but now becomes potentiated (A. Artola, S. Brocher and W. Singer, in preparation). This second threshold is related to NMDA-receptor-dependent conductances and cannot be reached if NMDA receptors are blocked. If this threshold is reached, however, the active input induces, in addition to its own potentiation, heterosynaptic depression of other inputs, i.e. other synapses on the same neuron become weakened if they are inactive. These inputs can, in turn, protect themselves against depression if they are also active and capable of activating NMDA-receptor-dependent conductances at their respective synapses. Thus, synapses capable of activating the NMD A receptor mechanism have a double competitive advantage: first, they increase their gain and are protected against heterosynaptic depression and, second, they are capable of repressing other synapses if these are not sufficiently active.

This brief survey shows that the modifications obtained in vitro resemble in detail those observed during development in vivo. They can account for the changes of synaptic transmission that occur in response to monocular deprivation and reverse suture. Moreover, the results on homosynaptic depression provide an explanation for the paradoxical observation that the afferents from the open eye can become weakened rather than strengthened if cortical neurons are exposed to high concentrations of 5-amino-phosphono-valeriate (APV) (Bear et al. 1990) or the γ-aminobutyric acid (GABA) receptor agonist muscimol (Reiter and Stryker, 1988). Such a modification is expected if the afferents from the open eye reach the threshold for homosynaptic depression but fail to activate NMDA-receptor-gated channels.

The persistence into adulthood of adaptive mechanisms closely resembling those active during development raises a number of important questions. The synaptic modifications observed in the mature visual cortex have all the properties postulated for mechanisms serving associative learning, suggesting the possibility that the primary visual cortex participates in storage of visual information. This is compatible with recent theoretical arguments predicting use-dependent gain changes of connections between neurons in the striate cortex which follow Hebbian (von der Malsburg and Bienenstock, 1986; von der Malsburg and Schneider, 1986) and anti-Hebbian (Barlow and Foldiak, 1989) rules.

The striking similarities between the adaptive processes occurring during development and those observed in the adult suggest the interesting possibility that the latter are actually the same as those which initiate the modifications of connectivity during development. If this is true the question remains why, in the adult, use-dependent modifications of synaptic gain are no longer associated with irreversible and physical removal of connections. One possibility is that the synapses that become fixed in the adult lose their adaptivity altogether while those that are still plastic can undergo the same physical modifications as during development. Alternatively, the only mechanisms lost may be those that are required for the physical removal and/or reconnection of pathways whereas those that persist underlie the reversible gain changes of synapses. These questions are directly related to the search for processes that limit circuit modifications to a critical period of early development. Therefore, in the next paragraph age-dependent changes will be reviewed that show some correlation with the time course of the decline of cortical malleability.

A characteristic feature of developmental plasticity is its rapid decline with age. Thus, deprivation-induced modification of ocular dominance is strongly reduced beyond 2–3 months of age and shrinkage of territories innervated by the deprived eye no longer occurs (Mower et al. 1985). The same holds true for other properties of cortical neurons, such as their orientation and direction preference (for a review see Frégnac and Imbert, 1984). Thus, it is of interest to define the factors that limit circuit modifications to early development.

The notion that cortical plasticity is substantially influenced by cholinergic mechanisms suggested the possibility that the age-dependent decline of malleability is the result of selective changes in the cholinergic projections to the striate cortex. In this case one would expect a withdrawal of cholinergic afferents from particular subsets of neurons and/or removal of muscarinic receptors. These changes ought to be particularly pronounced in layer IV, because the circuit modifications that lead to ocularity changes occur mainly within this layer. Experimental data support this hypothesis. The density and laminar distribution of high-affinity binding sites for muscarinic agonists change during development in a way that suggests a relationship to the time course of the critical period (Shaw et al. 1984). However, this redistribution of receptors is not paralleled by changes in the laminar pattern of cholinergic afferents. The density of fibers with choline acetyl transferase (ChAT)-like immunoreactivity increases with age, and there are no indications of a selective reduction of cholinergic fibers or terminal boutons from layer IV (Stichel and Singer, 1987a,b). One reason for this could be that in layer IV acetylcholine acts mainly on the nicotinic receptors, which are prominent in this layer and most probably located on the thalamic afferents (Prusky et al. 1987; Parkinson et al. 1988). The morphological data thus suggest that developmental changes of the cholinergic input to the striate cortex are not responsible for the temporal limitation of ocular dominance plasticity. The interpretation that developmental changes in modulatory afferents are not a limiting factor of plasticity is also supported by electrophysiological evidence. Combining ionophoretic application of noradrenaline and acetylcholine with patterned light stimuli has been found to be rather ineffective in producing substantial modifications of response properties in neurons of the visual cortex of adult cats while this procedure has been effective in kittens (Greuel et al. 1988). Recent evidence indicates, however, that a redistribution of calcium-dependent mechanisms might account for reduced plasticity in the adult. A first indication for this possibility came from the observation that it is more difficult to induce stimulation-dependent calcium fluxes from extra- to intracellular compartments in the visual cortex of adult cats than it is in kittens (Geiger and Singer, 1986). This result has recently been confirmed in slices of the visual cortex. In slices from kittens, white matter stimulation led to large Ca²⁺ sinks that were most prominent in the middle cortical layers, whereas in slices from adult cats, Ca²⁺ sinks were essentially confined to superficial layers (K. M. Bode-Greuel and W. Singer, in preparation). This suggests a laminar redistribution of Ca²⁺-dependent mechanisms during development.

Complementary support for an age-dependent redistribution of calcium dependent mechanisms comes from investigations of calcium-binding proteins and of both voltage- and receptor-gated calcium channels. An analysis of developmental changes in the distribution of the calcium-binding proteins parvalbumin and vitamin-D-dependent calcium-binding protein revealed that these calcium-buffering systems are particularly abundant in neurons of the developing visual cortex (Stichel et al. 1987). More recently, we obtained evidence that binding sites for organic calcium channel blockers of the dihydropyridine class are much more abundant in the visual cortex of kittens than in that of adults. The highest concentrations of these binding sites have been observed in layer IV at the peak of the critical period for ocular dominance plasticity. In good correlation with the age-dependent decline in ocular dominance plasticity, these binding sites disappeared from layer IV and became restricted essentially to supragranular layers. Changes in absolute density were also found for APV-sensitive, glutamate-binding sites (Bode-Greuel and Singer, 1989). If one assumes that the APV-sensitive binding sites of glutamate reflect NMDA receptors this would indicate that transmitter-activated calcium conductances also redistribute during development and become reduced with age. The presumptive reduction of both voltage-sensitive calcium channels and of transmitter-dependent calcium channels is thus compatible with the reduced calcium fluxes in the adult. It also correlates well with the in vivo (Tsumoto et al. 1987; Fox et al. 1989) and the in vitro (Kato et al. 1988) observation that in the visual cortex of young animals NMDA-receptor-gated conductances contribute more to the excitatory synaptic responses than they do in adult animals. Taken together, these results suggest that certain neuronal profiles lose their calcium conductances and hence are no longer exposed to activity-dependent calcium influx. Since the latter appears to serve as a prerequisite for activity-dependent long-term modifications of neuronal response properties, removal of calcium conductances could be one cause of the decline in plasticity with age.

Recently, data have become available which suggest the interesting possibility that maturational changes of glial cells might also be relevant for the decline of cortical malleability. The maturation of astroglial cells, i.e. their transformation from the protoplasmatic type to the glial fibrillary acidic protein (GFAP)-positive fibrous type, coincides in time with the reduction of ocular dominance plasticity (C. Müller, in preparation), suggesting the possibility of a causal relationship. This possibility has received strong support from the demonstration that injection of immature, cultured astrocytes into the striate cortex of adult cats promotes ocular dominance changes in response to several weeks of monocular deprivation (Müller and Best, 1989). An involvement of processes related to glial activity is also suggested by the finding that the ectoenzyme 5 ′ -nucleotidase, commonly attributed to glial cells (Kreutzberg et al. 1986), occurs in high concentrations in layer IV at the time when competition between the afferents from the two eyes leads to the segregation of ocular dominance columns. For a short period, the enzyme is even distributed in a discontinuous, patchy pattern related to ocular dominance columns (Schon et al. 1990). Interestingly, the cellular distribution of this enzyme changes during development and in relation to the decline of synaptic malleability. At a time when the connections are malleable the enzyme is associated with neuronal membranes, in particular with synaptic structures, and it only redistributes to its common glial location when malleability declines. This suggests a relationship between glial maturation and synaptic plasticity but gives no indication of the actual role of glial cells in ocular dominance plasticity. It is well established that glial elements play an important role in development and regeneration by providing contact cues and growth factors on the one hand and by acting as growth barriers and as phagocytes of cellular debris on the other. The early changes of ocular dominance that lead to rearrangements of axonal territories occur at a developmental stage when cellular differentiation, axonal sprouting and synapse formation are still active. Thus, as proposed by Müller and Best (1989), glial cells could promote ocular dominance plasticity in as much as they promote these developmental processes. At this stage of development, glial cells are not fully differentiated and contacts between neuronal processes are frequent, providing ideal conditions for the formation of new synapses. Later, the maturation and structural differentiation of glial cells lead to complete coverage of synapse-free neuronal membranes, so that contacts between the membranes of different nerve cells become rare. It is conceivable that this, together with the arrest of the proliferative processes, effectively prevents any reorganization of connectivity, confining use-dependent changes of synaptic connections in the adult to reversible modifications of synaptic gain. Whether transplantation of immature astrocytes actually reinstates the developmental processes required for the remodelling of circuitry or whether it only reinduces expression of neuronal mechanisms that facilitate use-dependent changes of synaptic gain remains to be clarified. Both possibilities need to be considered, since ocular dominance changes have been induced in the adult with manipulations that are probably not reactivating developmental processes. Thus, ocular dominance changes have been observed in the adult after intracortical infusion of noradrenaline (Kasamatsu et al. 1979), after pairing retinal stimulation with ionophoretic application of noradrenaline, acetylcholine and NMDA (Greuel et al. 1988) or potassium ions (Frégnac et al. 1988) and after manipulating experience in a way that made visual signals highly aversive (Singer et al. 1982a).

In conclusion, there are numerous developmental changes in several systems that could all be causally related to the age-dependent decline of cortical malleability, but we are still ignorant of the relative importance of these various factors. One of the reasons for this uncertainty is our present inability to decide whether the ocular dominance changes observed in the adult, as well as the use-dependent changes of synaptic gain that are observed in vitro, depend on the same or different mechanisms from the changes occurring in the young animal. The fact that large-scale rearrangements of afferent connections no longer occur in layer IV of the adult visual cortex does not necessarily exclude the possibility that similar processes occur in other layers, perhaps at a more microscopic level and with higher activation thresholds.