- Email: [email protected]
Feedforward, feedback and inhibitory connections in primate visual cortex
Neural Networks 17 (2004) 625–632 www.elsevier.com/locate/neunet
Feedforward, feedback and inhibitory connections in primate visual cortex Edward M. Callaway* Systems Neurobiology Laboratories, The Salk Institute for Biological Studies, 10010 North Torrey Pines Road, La Jolla, CA 92037, USA Received 20 April 2004; accepted 20 April 2004
Abstract Visual cortical circuits are organized at multiple levels of complexity including cortical areas, layers and columns, and specific cell types within these modules. Making sense of the functions of these circuits from anatomical observations requires linking these circuits to function at each of these levels of complexity. Observations of these relationships have become increasingly sophisticated over the last several decades, beginning with correlations between the connectivities and functions of various visual cortical areas and progressing toward cell type-specificity. These studies have informed current views about the functional interactions between cortical areas and modules and the mechanisms by which fine scale microcircuits influence interactions at more coarse levels of organization. q 2004 Elsevier Ltd. All rights reserved. Keywords: Local circuit; Driving; Modulatory; Gating; Inhibition; Excitation
1. Introduction and summary Anatomical studies of cortical circuits have revealed enormous complexity. At most levels of analysis the neurons in the cortex appear to form widespread distributed networks. For example, the primate primary visual cortex (V1) receives connections from at least 4 other cortical areas and from numerous subcortical structures and also provides output to most of the same structures (Felleman & Van Essen, 1991). But relating the various components of these circuits to their function reveals that connections differ in their influence on the behavior of the network (Salin & Bullier, 1995). Some connections appear to have a dominant, driving influence while others are modulatory (Callaway, 1998; Crick & Koch, 1998; Sherman & Guillery, 1998). Maunsell and van Essen (1983) and Van Essen and Maunsell (1983) used inferences about the functional relationships between different components of cortical circuits to derive anatomical rules that could be used to define hierarchical relationships between visual cortical areas. These rules led to the famous visual hierarchy described by Felleman and Van Essen. More recently, the connectivity of cortical neurons and their relationships to in vivo function have been studied with increasingly fine resolution. This has allowed feedforward (driving) and * Tel.: þ1-858-453-4100x1158. E-mail address: [email protected] (E.M. Callaway). 0893-6080/$ - see front matter q 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.neunet.2004.04.004
feedback (modulatory) connections to be identified within V1 local cortical circuits (Callaway, 1998) and also revealed how multiple diverging and converging streams (Van Essen & Deyoe, 1994) are embedded within this framework. There appear to be multiple interacting pathways for processing visual information, and the relative influences and interactions between pathways are likely to be dynamically regulated. Thus, a key problem which emerges for future studies is to understand whether and how neural circuits regulate these interactions and how such dynamic changes relate to visual perception. Inhibitory neurons are likely to play important and cell-type specific roles in these functions.
2. Classical feedforward and feedback connections between visual cortical areas Because the great majority of visual information reaching the cerebral cortex via the lateral geniculate nucleus of the thalamus (LGN) terminates in V1 (Benevento & Standage, 1982), and most (but not all) extrastriate cortical areas are dependent on V1 for their activation (Salin & Bullier, 1995), this area can be assigned to the lowest level in a hierarchy of visual cortical areas (Felleman & Van Essen, 1991). In contrast, other areas receiving input from V1 represent higher processing levels and therefore corticocortical connections out of V1 are defined as feedforward
626
E.M. Callaway / Neural Networks 17 (2004) 625–632
Fig. 1. Schematic diagram illustrating the anatomical features of cortico-cortical connections used by Felleman and Van Essen (1991) to assign hierarchical relationships between visual cortical areas. Forward connections terminate in layer 4 and originate from either superficial layers or from both superficial and deep layers. Feedback connections terminate outside layer 4 and originate either from deep layers or from both superficial and deep layers. From Felleman and Van Essen (1991).
and those to V1 are feedback. Connections from V1 terminate most densely in layer 4, as does the strongest input to V1 from the LGN. Thus termination of connections in layer 4 is a hallmark of feedforward cortico-cortical connections (Fig. 1; Felleman & Van Essen). With this rule as a starting point, it can then be seen that the feedforward connections that terminate in layer 4 originate from neurons in superficial cortical layers. And it can be shown that feedback connections terminate predominantly outside of layer 4 and originate from cells in deep layers (Fig. 1). (For more detailed reviews of these relationships see (Felleman & Van Essen, 1991; Salin & Bullier, 1995)).
3. Local feedforward and feedback connections in V1 The hierarchy of visual areas rests on a strong foundation. It is relatively straightforward to infer that V1 is the primary sensory area and most higher visual areas depend on this structure for their proper function. Understanding the flow of information within V1 is more
circuitous, but much can be learned from the rules that emerged from studies of cortico-cortical connections. Because feedforward cortico-cortical connections terminate predominantly in layer 4, it can be inferred that this layer represents an initial stage for processing by local circuits. And because feedforward cortico-cortical connections originate from superficial cortical layers this can be considered as the output stage. Finally because the layer 4 excitatory neurons that receive feedforward connections project axons most densely to more superficial layers (Yabuta & Callaway, 1998), it is natural to describe the local flow of information within a cortical area as being composed of two steps. The first step is the input to layer 4 and the second step from layer 4 to superficial layers (Fig. 2). The presence of numerous other local connections both to and from these layers illustrates that local circuits are not truly this simple (see Callaway, 1998 for review). Nevertheless, the presence of a basic, two-step framework for local cortical circuits is well-supported by correlations between connectivity and functional architecture in primate V1 (Callaway). These correlations suggest that LGN input
E.M. Callaway / Neural Networks 17 (2004) 625–632
Fig. 2. A two-level model of local cortical circuitry based on relationships between anatomical and functional properties in monkey V1 (Callaway, 1998). (See text for details.) Each processing level has a feedforward module (larger circles to the left) and a feedback module (smaller circles to the right). Feedforward modules receive driving excitatory input from the next lower level and make a similar driving connection to the next higher level (thick arrows). Feedback modules receive weaker input from the next lower level and from the feedforward module at the same level (dashed arrows). Finally, the feedback modules provide modulatory feedback connections to the feedforward module at the same level (thin arrows). Thus, feedforward modules relay information directly to the next level, while feedback modules combine information about the input to and output from the level and send it back to modulate the activity of the output neurons. From Callaway (1998).
to layer 4C, connections from layer 4 to more superficial layers, and connections from superficial layers to layer 4 of the next higher area are driving connections and thus define a dominant pathway for the feedforward processing of visual information (Fig. 2). The LGN input to layer 4C of V1 is organized into both sublayers and columns. LGN neurons driven by the ipsi- or contralateral eye terminate in distinct ocular dominance columns and the recipient layer 4C neurons are also driven only by visual stimulation of the corresponding eye (Hubel & Wiesel, 1968, 1972; Wiesel, Hubel, & Lam, 1974). Similarly, functionally distinct magno-(M) versus parvocellular (P) neurons in the LGN connect specifically to layers 4Ca and 4Cb, respectively (Blasdel & Lund, 1983; Hendrickson, Wilson, & Ogren, 1978; Hubel & Wiesel), and the recipient neurons have functional properties that strongly reflect these streamspecific inputs (Blasdel & Fitzpatrick, 1984). The relationships between functional properties and feedforward connections from layer 4C to more superficial layers of primate V1, suggest that these are also strong, driving connections. M pathway-recipient neurons in the upper part of layer 4Ca connect most strongly to layer 4B and within layer 3 have a bias toward cytochrome oxidase (CO) blobs (Lachica, Beck, & Casagrande, 1992; Yabuta & Callaway, 1998). In contrast, P pathway-recipient neurons in layer 4Cb connect most strongly to layer 3B, without any preference for blobs versus interblobs (Lachica et al., 1992; Yabuta & Callaway, 1998). These differences in feedforward input to layer 4B, layer 3 blobs, and layer 3 interblobs are reflected in differences in the functional properties of
627
neurons in these zones (Edwards, Purpura, & Kaplan, 1995; Livingstone & Hubel, 1988). Excitatory connections from deeper layers to layer 4 and superficial cortical layers have a less specific relationship to the functional architecture of V1. The lack of precision of these connections in the face of functional differences between compartments suggests that local connections from deep layers to layer 4 and above are modulatory connections (Fig. 2). For example, layer 6 pyramidal neurons make connections to layer 4C which lack specificity for ocular dominance columns (Wiser & Callaway, 1997). Despite this lack of specificity, layer 4C neurons are monocular; they can be driven by visual stimulation through one eye but not the other. Thus, under monocular viewing conditions which will activate layer 6 feedback to both ipsiand contralateral ocular dominance columns in layer 4C, layer 4C neurons in the column for the stimulated eye will be driven but in the other eye’s columns the active input from layer 6 is insufficient to drive visual responses. Thus the connections from layer 6 to layer 4C must be modulatory, not driving connections. Similarly, layer 5 and layer 6 pyramidal neurons that project to superficial layers of V1 have widespread axonal arbors that typically lack specificity for blobs, interblobs, or sublayers (e.g. 2/3A, 3B, 4B) (Callaway & Wiser, 1996; Wiser & Callaway, 1996). These observations suggest that these connections are also modulatory. Combining these inferences with information about the sources of input to deep layer neurons indicates that deep layers provide local modulatory feedback to superficial layers (Fig. 2). Layer 6 pyramidal neurons collect excitatory input from a broad range of local sources (Briggs & Callaway, 2001). Those that project axons to layer 4C receive their strongest excitatory input from layer 4C and deeper, but they often can also receive input from superficial layers. It is interesting to note that the input from the superficial layers is specific for different types of layer 6 pyramids that are associated with M versus P pathways. The pyramids that project axons only to M-recipient layer 4Ca can receive input from layer 4B but not from layer 2/3, while those that project axons to layer 4Cb can receive input from layer 2/3 but not layer 4B (Briggs & Callaway, 2001). Since layer 4B receives stronger input from layer 4Ca and layer 2/3 from 4Cb (see above), this organization suggests that stream-specific connections to and from these types of layer 6 neurons create trisynaptic rather than direct disynaptic feedback loops. Layer 4C cells connect to superficial cells which connect to layer 6 cells and then back to layer 4C. In particular, layer 4Ca connects most strongly to 4B which connects specifically to layer 6 cells that provide feedback to layer 4Ca. Layer 4Cb connects most strongly to layer 3B which connects specifically to layer 6 cells that provide feedback to layer 4Cb. Layer 5 and layer 6 pyramids that project axons to superficial layers appear to provide more direct, disynaptic feedback to the superficial layers (Fig. 2). Both layer 5 and layer 6 pyramidal neurons that project axons to layers 2-4B
628
E.M. Callaway / Neural Networks 17 (2004) 625–632
usually receive excitatory input from layer 2/3 (unpublished observations; (Briggs & Callaway, 2001)). But these same cells also receive input from other cortical layers, again highlighting the more general finding that neurons providing feedback typically integrate information from many sources. This observation is consistent with the suggestion that feedback neurons integrate information about both the input to and output from feedforward neurons and then modulate the activity of the feedforward neurons based on either a comparison or integration of this information (Callaway, 1998). 4. Alternate pathways 4.1. A feedforward cortico-cortical pathway from layer 5 via the pulvinar nucleus of the thalamus One of the more intriguing ideas to have emerged in recent years is the possibility that the classical cortico-cortical feedforward pathway has been overrated, or at the very least an important alternate pathway has been ignored (Guillery & Sherman, 2002). Guillery and Sherman propose that driving cortico-cortical communication can be mediated by layer 5 pyramidal neurons that project to the pulvinar nucleus, which in turn provides output to higher
Fig. 3. Schematic illustrating an alternative to the classical pathway for cortico-cortical communication. Primary sensory areas (1) receive input from first order thalamic nuclei (FO, e.g. LGN). The primary sensory area (1, e.g. V1) communicates via a proposed feedforward, driving pathway from layer 5 pyramidal neurons to a higher order thalamic nucleus (HO, e.g. pulvinar) which in turn provides input to the higher cortical area (2 or 4 in this diagram, e.g. V2). This mode of communication continues, always by way of higher order thalamic nuclei, through subsequent higher level processing stages. From Guillery and Sherman (2002) (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
cortical areas (Fig. 3). They make the case that many features of this pathway suggest that it constitutes a strong driving circuit. And the relatively precise and systematic organization of connections to and from the cortex and pulvinar (see (Shipp, 2003) for review) suggest that these connections are likely to mediate relatively direct interactions between cortical areas. A particular strength of this theory is that it provides a mechanism by which information used to generate motor responses is intimately linked to perceptual mechanisms. This is because the layer 5 neurons that project to the pulvinar also have axon collaterals that project to the superior colliculus (Lund, Lund, Hendrickson, Bunt, & Fuchs, 1975). Thus, instead of motor systems being connected to sensation only at the highest levels of visual processing, cortico-cortical communications utilize copies of the same signals sent to subcortical motor structures. 4.2. Shortcuts through V1 Although the majority of LGN input to primate V1 targets layer 4C, there are also substantial inputs to cytochrome oxidase blobs in layer 3, to layer 4A, and to layer 6. All of these regions contain neurons that make direct cortico-cortical connections. Thus a two-step feedforward pathway from LGN to extrastriate cortex may not be obligatory. Whether this is the case, however, depends on details of connectivity which are not known. In particular it is not known whether these sources of geniculate input directly connect to the same cells that make cortico-cortical connections, nor is it clear whether these are all driving connections. The best characterized of these alternative sources of geniculate input are those to the superficial layers of V1. LGN cells connecting to layer 4A and blobs have blueyellow color opponency, with the blob input being ‘blueOn’ and the 4A input ‘blue-Off’ (Chatterjee & Callaway, 2003). Since these pathways presumably play a crucial role in color vision, it is quite likely that they make driving connections and have a dominant role in determining the visual responses of color neurons in the visual cortex. The neurons which receive these connections may not, however, be neurons that project out of V1. Only about half of the layer 2/3 and 4A pyramidal neurons project out of V1, while the rest make only local connections (Callaway & Wiser, 1996). Furthermore, there is a precedent for selective connections to projecting versus local pyramids in layer 3 of V1—only local pyramids receive direct connections from layer 4Cb (Sawatari & Callaway, 2000). Thus it would not be surprising if processing of blue-yellow color opponent input to V1 included at least two-steps of feedforward processing, similar to the processing of input to layer 4C neurons. It is also likely that the blue-yellow input to layers 3 and 4A mixes directly with red-green opponent parvocellular input relayed from layer 4Cb (Sawatari & Callaway, 2000; Yabuta & Callaway, 1998). Thus, even if
E.M. Callaway / Neural Networks 17 (2004) 625–632
the superficial LGN input directly connected to corticocortical projecting pyramids, the local processing of the parvocellular input would still include multiple steps. Direct LGN input to layer 6 of V1 arises from collaterals of the same axons that target layer 4C (Blasdel & Lund, 1983; Freund, Martin, Soltesz, Somogyi, & Whitteridge, 1989). Most neurons in these layers do not make direct cortico-cortical connections, however the lone exception to this rule is particularly interesting. Very large pyramidal cells, or Meynert cells, in the deep layers of V1 include cells in layer 6 that project directly to the visual area MT (Fries, Keizer, & Kuypers, 1985). Although it is not known whether these cells directly receive LGN input, such a shortcut through V1 would be consistent with the role of MT in processing of visual motion information (Albright, 1984). The fast conduction velocity of the magnocellular afferent axons that would likely target these cells, combined with the fast conduction velocity suggested by the Meynert cells’ large axons, could provide an extremely fast conduit for information to reach area MT. Such a fast processing pathway could be important for animals to quickly assess and react to moving objects. Interestingly, it appears that the same Meynert cells that project to MT can also send collaterals to the superior colliculus (Fries et al., 1985). Thus, the implementation of a shortcut bypassing superficial cortex might simultaneously require an output to subcortical motor control structures. These cells might therefore also play a role in integrating perception with action, similar to that suggested for layer 5 pyramids that project to both the superior colliculus and the pulvinar nucleus (Guillery & Sherman, 2002). 5. Inhibitory connections Although there are a great diversity of inhibitory neuron types (e.g. see Kawaguchi & Kondo, 2002 for one recent review), the precise role of each cell type in the regulation of excitation in the cortex is not well understood. The purpose of this section of this review is not to discuss in detail each of these cell types or what is known about their roles in cortical circuits, but rather to discuss relationships between inhibitory connections and feedforward versus feedback excitatory connections. In particular I will discuss: (1) evidence that fast-spiking basket cells provide gain control for feedforward connections, (2) mechanisms by which other types of inhibitory neurons might provide more selective gating or gain control, and (3) a possible mechanism by which calretinin-expressing interneurons might dynamically route information between alternative feedforward pathways. 5.1. Fast-spiking cells provide gain control in the classical feedforward pathways An important role for inhibition is to provide gain control so that specificity of feedforward excitation can be
629
maintained in the face of increasingly strong activation of the feedforward pathway. In the primary visual cortex, this problem is exemplified by the contrast independence of orientation tuning—as visual stimulus contrast increases there are increasing activity levels and increasing nonlinearities in the responses of LGN neurons which must be managed by inhibitory cortical circuits in order to maintain selectivity (see Shapley, Hawken, & Ringach, 2003; Troyer, Krukowski, Priebe, & Miller, 1998; Troyer, Krukowski, & Miller, 2002 for review). Several lines of evidence strongly suggest that this gain control is mediated, at least in part, by fast-spiking basket cells. Feedforward connections both from thalamus to layer 4 (Agmon & Connors, 1992; Gibson, Beierlein, & Connors, 1999) and from layer 4 to layer 2/3 (Dantzker & Callaway, 2000) connect strongly and preferentially to fast-spiking basket cells. This contrasts with connections to other inhibitory cell types which often are not targeted by feedforward excitation (Agmon & Connors; Dantzker & Callaway, 2000; Gibson et al., 1999). (It should be noted that fast-spiking interneurons are not the only inhibitory cells that receive feedforward excitation (Dantzker & Callaway, 2000; Porter, Johnson, & Agmon, 2001)) Fast-spiking basket cells connect strongly to the somata of their neighboring pyramidal neurons and the placement of inhibition at the cell body allows these cells to provide strong control of the pyramidal neuron’s spiking (for example see Tamas, Buhl, Lorincz, & Somogyi, 2000). Thus, fast-spiking basket cells receive input from generally the same excitatory feedforward sources as the pyramidal neurons that they inhibit. Although the basic feedforward input sources are the same, the feedforward excitation to fast-spiking interneurons is probably more broadly tuned than the feedforward excitation to excitatory neurons (Hirsch et al., 2003; Swadlow, 2003). Thus, this inhibition is active regardless of the specificity of the stimulus and well-suited to shape response tuning by keeping the overall activity level of recipient neurons in a range where the tuning of feedforward excitation is most effective (e.g. Troyer et al., 1998, 2002). Strong and direct evidence for feedforward gain control by fast-spiking interneurons comes from Swadlow’s in vivo studies of barrel cortex (see Swadlow, 2002 for review). 5.2. Selective gating by dendrite targeting inhibitory neurons? In contrast to basket cells, which synapse mainly onto cell bodies and proximal dendrites, some inhibitory neuron types preferentially synapse onto more distal dendritic elements, including dendritic spines of pyramidal neurons (Kawaguchi & Kondo, 2002; Somogyi, Tamas, Lujan, & Buhl, 1998). Amongst the most notable of these cells are the somatostatin-expressing neurons found in both the hippocampus and cerebral cortex. In hippocampus these cells connect preferentially to the distal apical dendrites of CA1
630
E.M. Callaway / Neural Networks 17 (2004) 625–632
pyramidal neurons and it has therefore been suggested that they might preferentially inhibit excitatory inputs from the perforant path, which also connect to CA1 apical dendrites (Somogyi et al., 1998). Although there has not yet been a functional test of this hypothesis, the possibility that certain types of interneurons might be able to preferentially inhibit, and thus gate, excitatory input from specific sources suggests a powerful computational mechanism. Similarities between the features of somatostatin-expressing interneurons in cerebral cortex and hippocampus suggest that these cells might also play such a role in cerebral cortex. In the cerebral cortex these cells, often called ‘Martinotti cells’, have ascending axonal arbors that preferentially synapse onto distal dendritic elements (Kawaguchi & Kondo, 2002). Unlike in the hippocampus, however, the less distinct organization of the cerebral cortex leaves more uncertainty about the specific types of pyramidal cells that receive these inhibitory inputs. Although it is likely that connections are mostly to apical dendrites, these could be the dendrites of layer 2/3 pyramids or pyramids from deeper layers. A correlation between the expression of nitric oxide synthase (NOS) in some Martinotti cells and the NOS receptor, guanylyl cyclase, in the apical dendritic tufts of layer 5 pyramids (Vruwink, Schmidt, Weinberg, & Burette, 2001), suggests that layer 5 pyramids are a likely target of at least some of these connections. If Martinotti cells are able to selectively shunt excitatory input onto apical dendritic tufts, their role in cortical computations would depend on what excitatory inputs are located at those same dendritic locations and also what excitatory inputs activate the Martinotti cells. One possibility is that the Martinotti cells make synapses onto the same pyramidal neuron dendrites as excitatory feedback connections. This is most strongly suggested by the observation that cortico-cortical feedback often terminates in superficial cortical layers, particularly layer 1, where apical dendrites are most prominent. The input to Martinotti cells, on the other hand, is likely to include layer 5 pyramids. A subset of layer 2/3 regular spiking interneurons, likely including Martinotti cells, receive strong excitatory input predominantly from layer 5 (Dantzker & Callaway, 2000). Martinotti cells may therefore receive strong excitatory input from the same layer 5 cells whose apical dendritic tufts they in turn inhibit. This is suggestive of a pathwayspecific gain control. The amount of inhibition is dependent on the output of the same cell that is inhibited, but the inhibition is not generalized like it is for basket cells, instead it is specific to whatever excitatory input is targeting the apical tufts. 5.3. Pathway switching—dynamic routing by calretinin cells? The no strong loops hypothesis suggests that there should not be driving feedforward connections that create loops
(Crick & Koch, 1998). This hypothesis has been described in the framework of a static cortical circuit. But anatomical evidence that strong loops may exist suggests that it might be necessary to modify the hypothesis to integrate the possibility of dynamic changes in the flow of information. Driving circuits might be functionally silenced or become modulatory under some conditions. This hypothesis predicts that there should be specialized circuits which are able to dynamically switch the flow of information between feedforward pathways. Although there are many possible mechanisms by which such switching might be implemented, here I consider one scenario suggested by the anatomical relationships between calretinin expressing interneurons and the neurons that have been suggested to mediate feedforward connections in the classical pathway (e.g. Felleman & Van Essen, 1991) versus an alternate pathway through the pulvinar nucleus (Guillery & Sherman, 2002). An attractive mechanism by which the flow of information might be dynamically routed through different circuits is to have a single cell type (a ‘gating cell’) excite (or disinhibit) one pathway while inhibiting another. This cell type could then serve as a gate whose influence would in turn depend on the activation (or inhibition) of the gating cell. Based on anatomical observation it has been suggested that calretinin-expressing inhibitory neurons in monkey V1 disinhibit layer 2/3 pyramids while at the same time inhibiting deep layer pyramids (Meskenaite, 1997) (Fig. 4). Thus, these might be gating cells that can switch the flow of information between the classical cortico-cortical feedforward pathway from superficial layers and the alternate pathway from layer 5 tall pyramids via the pulvinar nucleus ((Guillery & Sherman, 2002), see above). Electron microscopic studies have demonstrated that inhibitory, calretinin-positive presynaptic terminals make a high proportion of their synapses (68 – 80%) onto other inhibitory neurons in layer 2/3 (Gonchar & Burkhalter, 1999; Meskenaite, 1997). Since these and other inhibitory neurons in layer 2/3 make strong inhibitory connections onto layer 2/3 pyramidal neurons, the effect of calretinin cell activity might be a double negative, or disinhibition of the layer 2/3 pyramids (Gonchar & Burkhalter; Meskenaite). Thus the anatomy of connections from calretinin-positive interneurons suggests that they might act to disinhibit the classical feedforward pathway (Fig. 4). In contrast to the anatomical observations in layer 2/3, calretinin-positive inhibitory axon terminals in layer 5 have been demonstrated to be mostly (60 – 80%) onto excitatory neurons (Gonchar & Burkhalter; Meskenaite). The calretinin cells might therefore inhibit the layer 5 pyramidal neurons that provide feedforward excitation to the pulvinar nucleus and mediate the alternative feedforward pathway proposed by Guillery and Sherman (2002). If both types of calretinin-positive synapses are coming from the same interneurons, then these cells could serve as a switch between the two pathways. When the calretinin cells were highly active they might
E.M. Callaway / Neural Networks 17 (2004) 625–632
631
Fig. 4. Schematic illustrating connections of calretinin-expressing inhibitory neurons (CR). Meskenaite (1997) infers that these ‘neurons appear to have a dual function of disinhibiting superficial layer neurons and inhibiting pyramidal output neurons in the deep layers.’ In superficial layers, the calretinin interneurons make connections primarily onto other inhibitory neurons (GABA). Since these other inhibitory neurons inhibit layer 2/3 pyramids, the net influence of the calretinin cells is to disinhibit layer 2/3 pyramids. In contrast, the calretinin neuron connections in layer 5 are primarily onto excitatory pyramidal neurons and they therefore directly inhibit excitatory output from layer 5. Viewed in light of Guillery and Sherman (2002) proposal that layer 5 pyramids might mediate an alternative feedforward pathway for cortico-cortical communication (see Fig. 3), the calretinin interneurons might provide a mechanism for switching the balance of cortico-cortical communication between the classical pathway from layer 2/3 pyramids versus the alternate pathway from layer 5. From Meskenaite (1997).
effectively disinhibit layer 2/3 pyramids while at the same time inhibiting layer 5 pyramids. This would favor the classical feedforward pathway from layer 2/3. If the activity of the calretinin interneurons were reduced, this might lead to a net inhibition of layer 2/3 pyramids (reduction of disinhibition) while reducing the direct inhibition (net excitation) of layer 5 pyramids (Fig. 4). This would favor activity in the alternate, feedforward pathway through the pulvinar. Whether this is in fact a role of calretinin-expressing interneurons is of course uncertain. For example, a key question is whether the inhibitory connections to layer 5 pyramids arise from the same calretinin expressing interneurons that inhibit other inhibitory cells in layer 2/3. The anatomical association demonstrated so far does not link the two types of connections to the same cells, just to cells of the same type. The role of switching by calretinin cells would also depend on the circumstances which lead to their activation. For example, feedback cortico-cortical connections that target layer 1 might preferentially activate these cells (Gonchar & Burkhalter, 2003). Ideally, a functional test of this hypothesis might include the selective disruption of the activity of calretinin-expressing interneurons in vivo. The effects of such a manipulation on the activity of other cells in the network might then be directly measured and even correlated with behavioral effects. Ultimately, although this particular hypothesis might prove to be wrong, the more general hypothesis that certain interneuron types might differentially regulate the flow of information through cortical circuits (e.g. Somogyi et al.,
1998) is likely to have considerable merit. Testing these ideas will require the generation of specific hypotheses as well as development of tools better suited to probing and manipulating specific cell types.
References Agmon, A., & Connors, B. W. (1992). Correlation between intrinsic firing patterns and thalamocortical synaptic responses of neurons in mouse barrel cortex. Journal of Neuroscience, 12(1), 319–329. Albright, T. D. (1984). Direction and orientation selectivity of neurons in visual area MT of the macaque. Journal of Neurophysiology, 52(6), 1106–1130. Benevento, L. A., & Standage, G. P. (1982). Demonstration of lack of dorsal lateral geniculate nucleus input to extrastriate areas MT and visual 2 in the macaque monkey. Brain Research, 252(1), 161 –166. Blasdel, G. G., & Fitzpatrick, D. (1984). Physiological organization of layer 4 in macaque striate cortex. Journal of Neuroscience, 4(3), 880–895. Blasdel, G. G., & Lund, J. S. (1983). Termination of afferent axons in macaque striate cortex. Journal of Neuroscience, 3(7), 1389–1413. Briggs, F., & Callaway, E. M. (2001). Layer-specific input to distinct cell types in layer 6 of monkey primary visual cortex. Journal of Neuroscience, 21(10), 3600–3608. Callaway, E. M. (1998). Local circuits in primary visual cortex of the macaque monkey. Annual Review of Neuroscience, 21, 47– 74. Callaway, E. M., & Wiser, A. K. (1996). Contributions of individual layer 2-5 spiny neurons to local circuits in macaque primary visual cortex. Visual Neuroscience, 13(5), 907–922. Chatterjee, S., & Callaway, E. M. (2003). Parallel colour-opponent pathways to primary visual cortex. Nature, 426(6967), 668– 671. Crick, F., & Koch, C. (1998). Constraints on cortical and thalamic projections: the no-strong-loops hypothesis. Nature, 391(6664), 245– 250.
632
E.M. Callaway / Neural Networks 17 (2004) 625–632
Dantzker, J. L., & Callaway, E. M. (2000). Laminar sources of synaptic input to cortical inhibitory interneurons and pyramidal neurons. Nature Neuroscience, 3(7), 701 –707. Edwards, D. P., Purpura, K. P., & Kaplan, E. (1995). Contrast sensitivity and spatial frequency response of primate cortical neurons in and around the cytochrome oxidase blobs. Vision Research, 35(11), 1501–1523. Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex, 1(1), 1– 47. Freund, T. F., Martin, K. A., Soltesz, I., Somogyi, P., & Whitteridge, D. (1989). Arborisation pattern and postsynaptic targets of physiologically identified thalamocortical afferents in striate cortex of the macaque monkey. The Journal of Comparative Neurology, 289(2), 315– 336. Fries, W., Keizer, K., & Kuypers, H. G. (1985). Large layer VI cells in macaque striate cortex (Meynert cells) project to both superior colliculus and prestriate visual area V5. Experimental Brain Research, 58(3), 613– 616. Gibson, J. R., Beierlein, M., & Connors, B. W. (1999). Two networks of electrically coupled inhibitory neurons in neocortex. Nature, 402(6757), 75 –79. Gonchar, Y., & Burkhalter, A. (1999). Connectivity of GABAergic Calretinin-immunoreactive neurons in rat primary visual cortex. Cerebral Cortex, 9(7), 683–696. Gonchar, Y., & Burkhalter, A. (2003). Distinct GABAergic targets of feedforward and feedback connections between lower and higher areas of rat visual cortex. Journal of Neuroscience, 23(34), 10904–10912. Guillery, R. W., & Sherman, S. M. (2002). Thalamic relay functions and their role in corticocortical communication: generalizations from the visual system. Neuron, 33(2), 163–175. Hendrickson, A. E., Wilson, J. R., & Ogren, M. P. (1978). The neuroanatomical organization of pathways between the dorsal lateral geniculate nucleus and visual cortex in Old World and New World primates. The Journal of Comparative Neurology, 182(1), 123– 136. Hirsch, J. A., Martinez, L. M., Pillai, C., Alonso, J. M., Wang, Q., & Sommer, F. T. (2003). Functionally distinct inhibitory neurons at the first stage of visual cortical processing. Nature Neuroscience, 6(12), 1300–1308. Hubel, D., & Wiesel, T. (1972). Laminar and columnar distribution of geniculo-cortical fibers in the macaque monkey. The Journal of Comparative Neurology, 146(4), 421 –450. Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional architecture of monkey striate cortex. Journal of Physiology (Lond), 195(1), 215– 243. Kawaguchi, Y., & Kondo, S. (2002). Parvalbumin, somatostatin and cholecystokinin as chemical markers for specific GABAergic interneuron types in the rat frontal cortex. Journal of Neurocytology, 31(3– 5), 277– 287. Lachica, E. A., Beck, P. D., & Casagrande, V. A. (1992). Parallel pathways in macaque monkey striate cortex: anatomically defined columns in layer III. Proceedings of the National Academy of Sciences of the United States of America, 89(8), 3566–3570. Livingstone, M., & Hubel, D. (1988). Segregation of form, color, movement, and depth: anatomy, physiology, and perception. Science, 240(4853), 740 –749. Lund, J. S., Lund, R. D., Hendrickson, A. E., Bunt, A. H., & Fuchs, A. F. (1975). The origin of efferent pathways from the primary visual cortex, area 17, of the macaque monkey as shown by retrograde transport of horseradish peroxidase. The Journal of Comparative Neurology, 164(3), 287– 303. Maunsell, J. H., & van Essen, D. C. (1983). The connections of the middle temporal visual area (MT) and their relationship to a cortical hierarchy in the macaque monkey. Journal of Neuroscience, 3(12), 2563–2586.
Meskenaite, V. (1997). Calretinin-immunoreactive local circuit neurons in area 17 of the cynomolgus monkey, Macaca fascicularis. The Journal of Comparative Neurology, 379(1), 113 –132. Porter, J. T., Johnson, C. K., & Agmon, A. (2001). Diverse types of interneurons generate thalamus-evoked feedforward inhibition in the mouse barrel cortex. Journal of Neuroscience, 21(8), 2699–2710. Salin, P. A., & Bullier, J. (1995). Corticocortical connections in the visual system: structure and function. Physiological Reviews, 75(1), 107 –154. Sawatari, A., & Callaway, E. M. (2000). Diversity and cell type specificity of local excitatory connections to neurons in layer 3B of monkey primary visual cortex. Neuron, 25(2), 459–471. Shapley, R., Hawken, M., & Ringach, D. L. (2003). Dynamics of orientation selectivity in the primary visual cortex and the importance of cortical inhibition. Neuron, 38(5), 689 –699. Sherman, S. M., & Guillery, R. W. (1998). On the actions that one nerve cell can have on another: distinguishing ‘drivers’ from ‘modulators’. Proceedings of the National Academy of Sciences of the United States of America, 95(12), 7121–7126. Shipp, S. (2003). The functional logic of cortico-pulvinar connections. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 358(1438), 1605–1624. Somogyi, P., Tamas, G., Lujan, R., & Buhl, E. H. (1998). Salient features of synaptic organisation in the cerebral cortex. Brain Research Brain Research Review, 26(2–3), 113– 135. Swadlow, H. A. (2002). Thalamocortical control of feed-forward inhibition in awake somatosensory ‘barrel’ cortex. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 357(1428), 1717–1727. Swadlow, H. A. (2003). Fast-spike interneurons and feedforward inhibition in awake sensory neocortex. Cerebral Cortex, 13(1), 25– 32. Tamas, G., Buhl, E. H., Lorincz, A., & Somogyi, P. (2000). Proximally targeted GABAergic synapses and gap junctions synchronize cortical interneurons. Nature Neuroscience, 3(4), 366 –371. Troyer, T. W., Krukowski, A. E., & Miller, K. D. (2002). LGN input to simple cells and contrast-invariant orientation tuning: an analysis. Journal of Neurophysiology, 87(6), 2741–2752. Troyer, T. W., Krukowski, A. E., Priebe, N. J., & Miller, K. D. (1998). Contrast-invariant orientation tuning in cat visual cortex: thalamocortical input tuning and correlation-based intracortical connectivity. Journal of Neuroscience, 18(15), 5908–5927. Van Essen, D. C., & Deyoe, E. A. (1994). Concurrent processing in the primate visual cortex. In M. S. Gazzuniga (Ed.), (pp. 383– 400). The cognitive neuroscience, Cambridge, MA: MIT Press. Van Essen, D. C., & Maunsell, J. H. R. (1983). Hierarchical organization and functional streams in the visual cortex. Trends in Neuroscience, 6(9), 370 –375. Vruwink, M., Schmidt, H. H., Weinberg, R. J., & Burette, A. (2001). Substance P and nitric oxide signaling in cerebral cortex: anatomical evidence for reciprocal signaling between two classes of interneurons. The Journal of Comparative Neurology, 441(4), 288–301. Wiesel, T. N., Hubel, D. H., & Lam, D. M. (1974). Autoradiographic demonstration of ocular-dominance columns in the monkey striate cortex by means of transneuronal transport. Brain Research, 79(2), 273 –279. Wiser, A. K., & Callaway, E. M. (1996). Contributions of individual layer 6 pyramidal neurons to local circuitry in macaque primary visual cortex. Journal of Neuroscience, 16(8), 2724–2739. Wiser, A. K., & Callaway, E. M. (1997). Ocular dominance columns and local projections of layer 6 pyramidal neurons in macaque primary visual cortex. Visual Neuroscience, 14(2), 241 –251. Yabuta, N. H., & Callaway, E. M. (1998). Functional streams and local connections of layer 4C neurons in primary visual cortex of the macaque monkey. Journal of Neuroscience, 18(22), 9489–9499.
Feedforward, feedback and inhibitory connections in primate visual cortex Edward M. Callaway* Systems Neurobiology Laboratories, The Salk Institute for Biological Studies, 10010 North Torrey Pines Road, La Jolla, CA 92037, USA Received 20 April 2004; accepted 20 April 2004
Abstract Visual cortical circuits are organized at multiple levels of complexity including cortical areas, layers and columns, and specific cell types within these modules. Making sense of the functions of these circuits from anatomical observations requires linking these circuits to function at each of these levels of complexity. Observations of these relationships have become increasingly sophisticated over the last several decades, beginning with correlations between the connectivities and functions of various visual cortical areas and progressing toward cell type-specificity. These studies have informed current views about the functional interactions between cortical areas and modules and the mechanisms by which fine scale microcircuits influence interactions at more coarse levels of organization. q 2004 Elsevier Ltd. All rights reserved. Keywords: Local circuit; Driving; Modulatory; Gating; Inhibition; Excitation
1. Introduction and summary Anatomical studies of cortical circuits have revealed enormous complexity. At most levels of analysis the neurons in the cortex appear to form widespread distributed networks. For example, the primate primary visual cortex (V1) receives connections from at least 4 other cortical areas and from numerous subcortical structures and also provides output to most of the same structures (Felleman & Van Essen, 1991). But relating the various components of these circuits to their function reveals that connections differ in their influence on the behavior of the network (Salin & Bullier, 1995). Some connections appear to have a dominant, driving influence while others are modulatory (Callaway, 1998; Crick & Koch, 1998; Sherman & Guillery, 1998). Maunsell and van Essen (1983) and Van Essen and Maunsell (1983) used inferences about the functional relationships between different components of cortical circuits to derive anatomical rules that could be used to define hierarchical relationships between visual cortical areas. These rules led to the famous visual hierarchy described by Felleman and Van Essen. More recently, the connectivity of cortical neurons and their relationships to in vivo function have been studied with increasingly fine resolution. This has allowed feedforward (driving) and * Tel.: þ1-858-453-4100x1158. E-mail address: [email protected] (E.M. Callaway). 0893-6080/$ - see front matter q 2004 Elsevier Ltd. All rights reserved. doi:10.1016/j.neunet.2004.04.004
feedback (modulatory) connections to be identified within V1 local cortical circuits (Callaway, 1998) and also revealed how multiple diverging and converging streams (Van Essen & Deyoe, 1994) are embedded within this framework. There appear to be multiple interacting pathways for processing visual information, and the relative influences and interactions between pathways are likely to be dynamically regulated. Thus, a key problem which emerges for future studies is to understand whether and how neural circuits regulate these interactions and how such dynamic changes relate to visual perception. Inhibitory neurons are likely to play important and cell-type specific roles in these functions.
2. Classical feedforward and feedback connections between visual cortical areas Because the great majority of visual information reaching the cerebral cortex via the lateral geniculate nucleus of the thalamus (LGN) terminates in V1 (Benevento & Standage, 1982), and most (but not all) extrastriate cortical areas are dependent on V1 for their activation (Salin & Bullier, 1995), this area can be assigned to the lowest level in a hierarchy of visual cortical areas (Felleman & Van Essen, 1991). In contrast, other areas receiving input from V1 represent higher processing levels and therefore corticocortical connections out of V1 are defined as feedforward
626
E.M. Callaway / Neural Networks 17 (2004) 625–632
Fig. 1. Schematic diagram illustrating the anatomical features of cortico-cortical connections used by Felleman and Van Essen (1991) to assign hierarchical relationships between visual cortical areas. Forward connections terminate in layer 4 and originate from either superficial layers or from both superficial and deep layers. Feedback connections terminate outside layer 4 and originate either from deep layers or from both superficial and deep layers. From Felleman and Van Essen (1991).
and those to V1 are feedback. Connections from V1 terminate most densely in layer 4, as does the strongest input to V1 from the LGN. Thus termination of connections in layer 4 is a hallmark of feedforward cortico-cortical connections (Fig. 1; Felleman & Van Essen). With this rule as a starting point, it can then be seen that the feedforward connections that terminate in layer 4 originate from neurons in superficial cortical layers. And it can be shown that feedback connections terminate predominantly outside of layer 4 and originate from cells in deep layers (Fig. 1). (For more detailed reviews of these relationships see (Felleman & Van Essen, 1991; Salin & Bullier, 1995)).
3. Local feedforward and feedback connections in V1 The hierarchy of visual areas rests on a strong foundation. It is relatively straightforward to infer that V1 is the primary sensory area and most higher visual areas depend on this structure for their proper function. Understanding the flow of information within V1 is more
circuitous, but much can be learned from the rules that emerged from studies of cortico-cortical connections. Because feedforward cortico-cortical connections terminate predominantly in layer 4, it can be inferred that this layer represents an initial stage for processing by local circuits. And because feedforward cortico-cortical connections originate from superficial cortical layers this can be considered as the output stage. Finally because the layer 4 excitatory neurons that receive feedforward connections project axons most densely to more superficial layers (Yabuta & Callaway, 1998), it is natural to describe the local flow of information within a cortical area as being composed of two steps. The first step is the input to layer 4 and the second step from layer 4 to superficial layers (Fig. 2). The presence of numerous other local connections both to and from these layers illustrates that local circuits are not truly this simple (see Callaway, 1998 for review). Nevertheless, the presence of a basic, two-step framework for local cortical circuits is well-supported by correlations between connectivity and functional architecture in primate V1 (Callaway). These correlations suggest that LGN input
E.M. Callaway / Neural Networks 17 (2004) 625–632
Fig. 2. A two-level model of local cortical circuitry based on relationships between anatomical and functional properties in monkey V1 (Callaway, 1998). (See text for details.) Each processing level has a feedforward module (larger circles to the left) and a feedback module (smaller circles to the right). Feedforward modules receive driving excitatory input from the next lower level and make a similar driving connection to the next higher level (thick arrows). Feedback modules receive weaker input from the next lower level and from the feedforward module at the same level (dashed arrows). Finally, the feedback modules provide modulatory feedback connections to the feedforward module at the same level (thin arrows). Thus, feedforward modules relay information directly to the next level, while feedback modules combine information about the input to and output from the level and send it back to modulate the activity of the output neurons. From Callaway (1998).
to layer 4C, connections from layer 4 to more superficial layers, and connections from superficial layers to layer 4 of the next higher area are driving connections and thus define a dominant pathway for the feedforward processing of visual information (Fig. 2). The LGN input to layer 4C of V1 is organized into both sublayers and columns. LGN neurons driven by the ipsi- or contralateral eye terminate in distinct ocular dominance columns and the recipient layer 4C neurons are also driven only by visual stimulation of the corresponding eye (Hubel & Wiesel, 1968, 1972; Wiesel, Hubel, & Lam, 1974). Similarly, functionally distinct magno-(M) versus parvocellular (P) neurons in the LGN connect specifically to layers 4Ca and 4Cb, respectively (Blasdel & Lund, 1983; Hendrickson, Wilson, & Ogren, 1978; Hubel & Wiesel), and the recipient neurons have functional properties that strongly reflect these streamspecific inputs (Blasdel & Fitzpatrick, 1984). The relationships between functional properties and feedforward connections from layer 4C to more superficial layers of primate V1, suggest that these are also strong, driving connections. M pathway-recipient neurons in the upper part of layer 4Ca connect most strongly to layer 4B and within layer 3 have a bias toward cytochrome oxidase (CO) blobs (Lachica, Beck, & Casagrande, 1992; Yabuta & Callaway, 1998). In contrast, P pathway-recipient neurons in layer 4Cb connect most strongly to layer 3B, without any preference for blobs versus interblobs (Lachica et al., 1992; Yabuta & Callaway, 1998). These differences in feedforward input to layer 4B, layer 3 blobs, and layer 3 interblobs are reflected in differences in the functional properties of
627
neurons in these zones (Edwards, Purpura, & Kaplan, 1995; Livingstone & Hubel, 1988). Excitatory connections from deeper layers to layer 4 and superficial cortical layers have a less specific relationship to the functional architecture of V1. The lack of precision of these connections in the face of functional differences between compartments suggests that local connections from deep layers to layer 4 and above are modulatory connections (Fig. 2). For example, layer 6 pyramidal neurons make connections to layer 4C which lack specificity for ocular dominance columns (Wiser & Callaway, 1997). Despite this lack of specificity, layer 4C neurons are monocular; they can be driven by visual stimulation through one eye but not the other. Thus, under monocular viewing conditions which will activate layer 6 feedback to both ipsiand contralateral ocular dominance columns in layer 4C, layer 4C neurons in the column for the stimulated eye will be driven but in the other eye’s columns the active input from layer 6 is insufficient to drive visual responses. Thus the connections from layer 6 to layer 4C must be modulatory, not driving connections. Similarly, layer 5 and layer 6 pyramidal neurons that project to superficial layers of V1 have widespread axonal arbors that typically lack specificity for blobs, interblobs, or sublayers (e.g. 2/3A, 3B, 4B) (Callaway & Wiser, 1996; Wiser & Callaway, 1996). These observations suggest that these connections are also modulatory. Combining these inferences with information about the sources of input to deep layer neurons indicates that deep layers provide local modulatory feedback to superficial layers (Fig. 2). Layer 6 pyramidal neurons collect excitatory input from a broad range of local sources (Briggs & Callaway, 2001). Those that project axons to layer 4C receive their strongest excitatory input from layer 4C and deeper, but they often can also receive input from superficial layers. It is interesting to note that the input from the superficial layers is specific for different types of layer 6 pyramids that are associated with M versus P pathways. The pyramids that project axons only to M-recipient layer 4Ca can receive input from layer 4B but not from layer 2/3, while those that project axons to layer 4Cb can receive input from layer 2/3 but not layer 4B (Briggs & Callaway, 2001). Since layer 4B receives stronger input from layer 4Ca and layer 2/3 from 4Cb (see above), this organization suggests that stream-specific connections to and from these types of layer 6 neurons create trisynaptic rather than direct disynaptic feedback loops. Layer 4C cells connect to superficial cells which connect to layer 6 cells and then back to layer 4C. In particular, layer 4Ca connects most strongly to 4B which connects specifically to layer 6 cells that provide feedback to layer 4Ca. Layer 4Cb connects most strongly to layer 3B which connects specifically to layer 6 cells that provide feedback to layer 4Cb. Layer 5 and layer 6 pyramids that project axons to superficial layers appear to provide more direct, disynaptic feedback to the superficial layers (Fig. 2). Both layer 5 and layer 6 pyramidal neurons that project axons to layers 2-4B
628
E.M. Callaway / Neural Networks 17 (2004) 625–632
usually receive excitatory input from layer 2/3 (unpublished observations; (Briggs & Callaway, 2001)). But these same cells also receive input from other cortical layers, again highlighting the more general finding that neurons providing feedback typically integrate information from many sources. This observation is consistent with the suggestion that feedback neurons integrate information about both the input to and output from feedforward neurons and then modulate the activity of the feedforward neurons based on either a comparison or integration of this information (Callaway, 1998). 4. Alternate pathways 4.1. A feedforward cortico-cortical pathway from layer 5 via the pulvinar nucleus of the thalamus One of the more intriguing ideas to have emerged in recent years is the possibility that the classical cortico-cortical feedforward pathway has been overrated, or at the very least an important alternate pathway has been ignored (Guillery & Sherman, 2002). Guillery and Sherman propose that driving cortico-cortical communication can be mediated by layer 5 pyramidal neurons that project to the pulvinar nucleus, which in turn provides output to higher
Fig. 3. Schematic illustrating an alternative to the classical pathway for cortico-cortical communication. Primary sensory areas (1) receive input from first order thalamic nuclei (FO, e.g. LGN). The primary sensory area (1, e.g. V1) communicates via a proposed feedforward, driving pathway from layer 5 pyramidal neurons to a higher order thalamic nucleus (HO, e.g. pulvinar) which in turn provides input to the higher cortical area (2 or 4 in this diagram, e.g. V2). This mode of communication continues, always by way of higher order thalamic nuclei, through subsequent higher level processing stages. From Guillery and Sherman (2002) (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
cortical areas (Fig. 3). They make the case that many features of this pathway suggest that it constitutes a strong driving circuit. And the relatively precise and systematic organization of connections to and from the cortex and pulvinar (see (Shipp, 2003) for review) suggest that these connections are likely to mediate relatively direct interactions between cortical areas. A particular strength of this theory is that it provides a mechanism by which information used to generate motor responses is intimately linked to perceptual mechanisms. This is because the layer 5 neurons that project to the pulvinar also have axon collaterals that project to the superior colliculus (Lund, Lund, Hendrickson, Bunt, & Fuchs, 1975). Thus, instead of motor systems being connected to sensation only at the highest levels of visual processing, cortico-cortical communications utilize copies of the same signals sent to subcortical motor structures. 4.2. Shortcuts through V1 Although the majority of LGN input to primate V1 targets layer 4C, there are also substantial inputs to cytochrome oxidase blobs in layer 3, to layer 4A, and to layer 6. All of these regions contain neurons that make direct cortico-cortical connections. Thus a two-step feedforward pathway from LGN to extrastriate cortex may not be obligatory. Whether this is the case, however, depends on details of connectivity which are not known. In particular it is not known whether these sources of geniculate input directly connect to the same cells that make cortico-cortical connections, nor is it clear whether these are all driving connections. The best characterized of these alternative sources of geniculate input are those to the superficial layers of V1. LGN cells connecting to layer 4A and blobs have blueyellow color opponency, with the blob input being ‘blueOn’ and the 4A input ‘blue-Off’ (Chatterjee & Callaway, 2003). Since these pathways presumably play a crucial role in color vision, it is quite likely that they make driving connections and have a dominant role in determining the visual responses of color neurons in the visual cortex. The neurons which receive these connections may not, however, be neurons that project out of V1. Only about half of the layer 2/3 and 4A pyramidal neurons project out of V1, while the rest make only local connections (Callaway & Wiser, 1996). Furthermore, there is a precedent for selective connections to projecting versus local pyramids in layer 3 of V1—only local pyramids receive direct connections from layer 4Cb (Sawatari & Callaway, 2000). Thus it would not be surprising if processing of blue-yellow color opponent input to V1 included at least two-steps of feedforward processing, similar to the processing of input to layer 4C neurons. It is also likely that the blue-yellow input to layers 3 and 4A mixes directly with red-green opponent parvocellular input relayed from layer 4Cb (Sawatari & Callaway, 2000; Yabuta & Callaway, 1998). Thus, even if
E.M. Callaway / Neural Networks 17 (2004) 625–632
the superficial LGN input directly connected to corticocortical projecting pyramids, the local processing of the parvocellular input would still include multiple steps. Direct LGN input to layer 6 of V1 arises from collaterals of the same axons that target layer 4C (Blasdel & Lund, 1983; Freund, Martin, Soltesz, Somogyi, & Whitteridge, 1989). Most neurons in these layers do not make direct cortico-cortical connections, however the lone exception to this rule is particularly interesting. Very large pyramidal cells, or Meynert cells, in the deep layers of V1 include cells in layer 6 that project directly to the visual area MT (Fries, Keizer, & Kuypers, 1985). Although it is not known whether these cells directly receive LGN input, such a shortcut through V1 would be consistent with the role of MT in processing of visual motion information (Albright, 1984). The fast conduction velocity of the magnocellular afferent axons that would likely target these cells, combined with the fast conduction velocity suggested by the Meynert cells’ large axons, could provide an extremely fast conduit for information to reach area MT. Such a fast processing pathway could be important for animals to quickly assess and react to moving objects. Interestingly, it appears that the same Meynert cells that project to MT can also send collaterals to the superior colliculus (Fries et al., 1985). Thus, the implementation of a shortcut bypassing superficial cortex might simultaneously require an output to subcortical motor control structures. These cells might therefore also play a role in integrating perception with action, similar to that suggested for layer 5 pyramids that project to both the superior colliculus and the pulvinar nucleus (Guillery & Sherman, 2002). 5. Inhibitory connections Although there are a great diversity of inhibitory neuron types (e.g. see Kawaguchi & Kondo, 2002 for one recent review), the precise role of each cell type in the regulation of excitation in the cortex is not well understood. The purpose of this section of this review is not to discuss in detail each of these cell types or what is known about their roles in cortical circuits, but rather to discuss relationships between inhibitory connections and feedforward versus feedback excitatory connections. In particular I will discuss: (1) evidence that fast-spiking basket cells provide gain control for feedforward connections, (2) mechanisms by which other types of inhibitory neurons might provide more selective gating or gain control, and (3) a possible mechanism by which calretinin-expressing interneurons might dynamically route information between alternative feedforward pathways. 5.1. Fast-spiking cells provide gain control in the classical feedforward pathways An important role for inhibition is to provide gain control so that specificity of feedforward excitation can be
629
maintained in the face of increasingly strong activation of the feedforward pathway. In the primary visual cortex, this problem is exemplified by the contrast independence of orientation tuning—as visual stimulus contrast increases there are increasing activity levels and increasing nonlinearities in the responses of LGN neurons which must be managed by inhibitory cortical circuits in order to maintain selectivity (see Shapley, Hawken, & Ringach, 2003; Troyer, Krukowski, Priebe, & Miller, 1998; Troyer, Krukowski, & Miller, 2002 for review). Several lines of evidence strongly suggest that this gain control is mediated, at least in part, by fast-spiking basket cells. Feedforward connections both from thalamus to layer 4 (Agmon & Connors, 1992; Gibson, Beierlein, & Connors, 1999) and from layer 4 to layer 2/3 (Dantzker & Callaway, 2000) connect strongly and preferentially to fast-spiking basket cells. This contrasts with connections to other inhibitory cell types which often are not targeted by feedforward excitation (Agmon & Connors; Dantzker & Callaway, 2000; Gibson et al., 1999). (It should be noted that fast-spiking interneurons are not the only inhibitory cells that receive feedforward excitation (Dantzker & Callaway, 2000; Porter, Johnson, & Agmon, 2001)) Fast-spiking basket cells connect strongly to the somata of their neighboring pyramidal neurons and the placement of inhibition at the cell body allows these cells to provide strong control of the pyramidal neuron’s spiking (for example see Tamas, Buhl, Lorincz, & Somogyi, 2000). Thus, fast-spiking basket cells receive input from generally the same excitatory feedforward sources as the pyramidal neurons that they inhibit. Although the basic feedforward input sources are the same, the feedforward excitation to fast-spiking interneurons is probably more broadly tuned than the feedforward excitation to excitatory neurons (Hirsch et al., 2003; Swadlow, 2003). Thus, this inhibition is active regardless of the specificity of the stimulus and well-suited to shape response tuning by keeping the overall activity level of recipient neurons in a range where the tuning of feedforward excitation is most effective (e.g. Troyer et al., 1998, 2002). Strong and direct evidence for feedforward gain control by fast-spiking interneurons comes from Swadlow’s in vivo studies of barrel cortex (see Swadlow, 2002 for review). 5.2. Selective gating by dendrite targeting inhibitory neurons? In contrast to basket cells, which synapse mainly onto cell bodies and proximal dendrites, some inhibitory neuron types preferentially synapse onto more distal dendritic elements, including dendritic spines of pyramidal neurons (Kawaguchi & Kondo, 2002; Somogyi, Tamas, Lujan, & Buhl, 1998). Amongst the most notable of these cells are the somatostatin-expressing neurons found in both the hippocampus and cerebral cortex. In hippocampus these cells connect preferentially to the distal apical dendrites of CA1
630
E.M. Callaway / Neural Networks 17 (2004) 625–632
pyramidal neurons and it has therefore been suggested that they might preferentially inhibit excitatory inputs from the perforant path, which also connect to CA1 apical dendrites (Somogyi et al., 1998). Although there has not yet been a functional test of this hypothesis, the possibility that certain types of interneurons might be able to preferentially inhibit, and thus gate, excitatory input from specific sources suggests a powerful computational mechanism. Similarities between the features of somatostatin-expressing interneurons in cerebral cortex and hippocampus suggest that these cells might also play such a role in cerebral cortex. In the cerebral cortex these cells, often called ‘Martinotti cells’, have ascending axonal arbors that preferentially synapse onto distal dendritic elements (Kawaguchi & Kondo, 2002). Unlike in the hippocampus, however, the less distinct organization of the cerebral cortex leaves more uncertainty about the specific types of pyramidal cells that receive these inhibitory inputs. Although it is likely that connections are mostly to apical dendrites, these could be the dendrites of layer 2/3 pyramids or pyramids from deeper layers. A correlation between the expression of nitric oxide synthase (NOS) in some Martinotti cells and the NOS receptor, guanylyl cyclase, in the apical dendritic tufts of layer 5 pyramids (Vruwink, Schmidt, Weinberg, & Burette, 2001), suggests that layer 5 pyramids are a likely target of at least some of these connections. If Martinotti cells are able to selectively shunt excitatory input onto apical dendritic tufts, their role in cortical computations would depend on what excitatory inputs are located at those same dendritic locations and also what excitatory inputs activate the Martinotti cells. One possibility is that the Martinotti cells make synapses onto the same pyramidal neuron dendrites as excitatory feedback connections. This is most strongly suggested by the observation that cortico-cortical feedback often terminates in superficial cortical layers, particularly layer 1, where apical dendrites are most prominent. The input to Martinotti cells, on the other hand, is likely to include layer 5 pyramids. A subset of layer 2/3 regular spiking interneurons, likely including Martinotti cells, receive strong excitatory input predominantly from layer 5 (Dantzker & Callaway, 2000). Martinotti cells may therefore receive strong excitatory input from the same layer 5 cells whose apical dendritic tufts they in turn inhibit. This is suggestive of a pathwayspecific gain control. The amount of inhibition is dependent on the output of the same cell that is inhibited, but the inhibition is not generalized like it is for basket cells, instead it is specific to whatever excitatory input is targeting the apical tufts. 5.3. Pathway switching—dynamic routing by calretinin cells? The no strong loops hypothesis suggests that there should not be driving feedforward connections that create loops
(Crick & Koch, 1998). This hypothesis has been described in the framework of a static cortical circuit. But anatomical evidence that strong loops may exist suggests that it might be necessary to modify the hypothesis to integrate the possibility of dynamic changes in the flow of information. Driving circuits might be functionally silenced or become modulatory under some conditions. This hypothesis predicts that there should be specialized circuits which are able to dynamically switch the flow of information between feedforward pathways. Although there are many possible mechanisms by which such switching might be implemented, here I consider one scenario suggested by the anatomical relationships between calretinin expressing interneurons and the neurons that have been suggested to mediate feedforward connections in the classical pathway (e.g. Felleman & Van Essen, 1991) versus an alternate pathway through the pulvinar nucleus (Guillery & Sherman, 2002). An attractive mechanism by which the flow of information might be dynamically routed through different circuits is to have a single cell type (a ‘gating cell’) excite (or disinhibit) one pathway while inhibiting another. This cell type could then serve as a gate whose influence would in turn depend on the activation (or inhibition) of the gating cell. Based on anatomical observation it has been suggested that calretinin-expressing inhibitory neurons in monkey V1 disinhibit layer 2/3 pyramids while at the same time inhibiting deep layer pyramids (Meskenaite, 1997) (Fig. 4). Thus, these might be gating cells that can switch the flow of information between the classical cortico-cortical feedforward pathway from superficial layers and the alternate pathway from layer 5 tall pyramids via the pulvinar nucleus ((Guillery & Sherman, 2002), see above). Electron microscopic studies have demonstrated that inhibitory, calretinin-positive presynaptic terminals make a high proportion of their synapses (68 – 80%) onto other inhibitory neurons in layer 2/3 (Gonchar & Burkhalter, 1999; Meskenaite, 1997). Since these and other inhibitory neurons in layer 2/3 make strong inhibitory connections onto layer 2/3 pyramidal neurons, the effect of calretinin cell activity might be a double negative, or disinhibition of the layer 2/3 pyramids (Gonchar & Burkhalter; Meskenaite). Thus the anatomy of connections from calretinin-positive interneurons suggests that they might act to disinhibit the classical feedforward pathway (Fig. 4). In contrast to the anatomical observations in layer 2/3, calretinin-positive inhibitory axon terminals in layer 5 have been demonstrated to be mostly (60 – 80%) onto excitatory neurons (Gonchar & Burkhalter; Meskenaite). The calretinin cells might therefore inhibit the layer 5 pyramidal neurons that provide feedforward excitation to the pulvinar nucleus and mediate the alternative feedforward pathway proposed by Guillery and Sherman (2002). If both types of calretinin-positive synapses are coming from the same interneurons, then these cells could serve as a switch between the two pathways. When the calretinin cells were highly active they might
E.M. Callaway / Neural Networks 17 (2004) 625–632
631
Fig. 4. Schematic illustrating connections of calretinin-expressing inhibitory neurons (CR). Meskenaite (1997) infers that these ‘neurons appear to have a dual function of disinhibiting superficial layer neurons and inhibiting pyramidal output neurons in the deep layers.’ In superficial layers, the calretinin interneurons make connections primarily onto other inhibitory neurons (GABA). Since these other inhibitory neurons inhibit layer 2/3 pyramids, the net influence of the calretinin cells is to disinhibit layer 2/3 pyramids. In contrast, the calretinin neuron connections in layer 5 are primarily onto excitatory pyramidal neurons and they therefore directly inhibit excitatory output from layer 5. Viewed in light of Guillery and Sherman (2002) proposal that layer 5 pyramids might mediate an alternative feedforward pathway for cortico-cortical communication (see Fig. 3), the calretinin interneurons might provide a mechanism for switching the balance of cortico-cortical communication between the classical pathway from layer 2/3 pyramids versus the alternate pathway from layer 5. From Meskenaite (1997).
effectively disinhibit layer 2/3 pyramids while at the same time inhibiting layer 5 pyramids. This would favor the classical feedforward pathway from layer 2/3. If the activity of the calretinin interneurons were reduced, this might lead to a net inhibition of layer 2/3 pyramids (reduction of disinhibition) while reducing the direct inhibition (net excitation) of layer 5 pyramids (Fig. 4). This would favor activity in the alternate, feedforward pathway through the pulvinar. Whether this is in fact a role of calretinin-expressing interneurons is of course uncertain. For example, a key question is whether the inhibitory connections to layer 5 pyramids arise from the same calretinin expressing interneurons that inhibit other inhibitory cells in layer 2/3. The anatomical association demonstrated so far does not link the two types of connections to the same cells, just to cells of the same type. The role of switching by calretinin cells would also depend on the circumstances which lead to their activation. For example, feedback cortico-cortical connections that target layer 1 might preferentially activate these cells (Gonchar & Burkhalter, 2003). Ideally, a functional test of this hypothesis might include the selective disruption of the activity of calretinin-expressing interneurons in vivo. The effects of such a manipulation on the activity of other cells in the network might then be directly measured and even correlated with behavioral effects. Ultimately, although this particular hypothesis might prove to be wrong, the more general hypothesis that certain interneuron types might differentially regulate the flow of information through cortical circuits (e.g. Somogyi et al.,
1998) is likely to have considerable merit. Testing these ideas will require the generation of specific hypotheses as well as development of tools better suited to probing and manipulating specific cell types.
References Agmon, A., & Connors, B. W. (1992). Correlation between intrinsic firing patterns and thalamocortical synaptic responses of neurons in mouse barrel cortex. Journal of Neuroscience, 12(1), 319–329. Albright, T. D. (1984). Direction and orientation selectivity of neurons in visual area MT of the macaque. Journal of Neurophysiology, 52(6), 1106–1130. Benevento, L. A., & Standage, G. P. (1982). Demonstration of lack of dorsal lateral geniculate nucleus input to extrastriate areas MT and visual 2 in the macaque monkey. Brain Research, 252(1), 161 –166. Blasdel, G. G., & Fitzpatrick, D. (1984). Physiological organization of layer 4 in macaque striate cortex. Journal of Neuroscience, 4(3), 880–895. Blasdel, G. G., & Lund, J. S. (1983). Termination of afferent axons in macaque striate cortex. Journal of Neuroscience, 3(7), 1389–1413. Briggs, F., & Callaway, E. M. (2001). Layer-specific input to distinct cell types in layer 6 of monkey primary visual cortex. Journal of Neuroscience, 21(10), 3600–3608. Callaway, E. M. (1998). Local circuits in primary visual cortex of the macaque monkey. Annual Review of Neuroscience, 21, 47– 74. Callaway, E. M., & Wiser, A. K. (1996). Contributions of individual layer 2-5 spiny neurons to local circuits in macaque primary visual cortex. Visual Neuroscience, 13(5), 907–922. Chatterjee, S., & Callaway, E. M. (2003). Parallel colour-opponent pathways to primary visual cortex. Nature, 426(6967), 668– 671. Crick, F., & Koch, C. (1998). Constraints on cortical and thalamic projections: the no-strong-loops hypothesis. Nature, 391(6664), 245– 250.
632
E.M. Callaway / Neural Networks 17 (2004) 625–632
Dantzker, J. L., & Callaway, E. M. (2000). Laminar sources of synaptic input to cortical inhibitory interneurons and pyramidal neurons. Nature Neuroscience, 3(7), 701 –707. Edwards, D. P., Purpura, K. P., & Kaplan, E. (1995). Contrast sensitivity and spatial frequency response of primate cortical neurons in and around the cytochrome oxidase blobs. Vision Research, 35(11), 1501–1523. Felleman, D. J., & Van Essen, D. C. (1991). Distributed hierarchical processing in the primate cerebral cortex. Cerebral Cortex, 1(1), 1– 47. Freund, T. F., Martin, K. A., Soltesz, I., Somogyi, P., & Whitteridge, D. (1989). Arborisation pattern and postsynaptic targets of physiologically identified thalamocortical afferents in striate cortex of the macaque monkey. The Journal of Comparative Neurology, 289(2), 315– 336. Fries, W., Keizer, K., & Kuypers, H. G. (1985). Large layer VI cells in macaque striate cortex (Meynert cells) project to both superior colliculus and prestriate visual area V5. Experimental Brain Research, 58(3), 613– 616. Gibson, J. R., Beierlein, M., & Connors, B. W. (1999). Two networks of electrically coupled inhibitory neurons in neocortex. Nature, 402(6757), 75 –79. Gonchar, Y., & Burkhalter, A. (1999). Connectivity of GABAergic Calretinin-immunoreactive neurons in rat primary visual cortex. Cerebral Cortex, 9(7), 683–696. Gonchar, Y., & Burkhalter, A. (2003). Distinct GABAergic targets of feedforward and feedback connections between lower and higher areas of rat visual cortex. Journal of Neuroscience, 23(34), 10904–10912. Guillery, R. W., & Sherman, S. M. (2002). Thalamic relay functions and their role in corticocortical communication: generalizations from the visual system. Neuron, 33(2), 163–175. Hendrickson, A. E., Wilson, J. R., & Ogren, M. P. (1978). The neuroanatomical organization of pathways between the dorsal lateral geniculate nucleus and visual cortex in Old World and New World primates. The Journal of Comparative Neurology, 182(1), 123– 136. Hirsch, J. A., Martinez, L. M., Pillai, C., Alonso, J. M., Wang, Q., & Sommer, F. T. (2003). Functionally distinct inhibitory neurons at the first stage of visual cortical processing. Nature Neuroscience, 6(12), 1300–1308. Hubel, D., & Wiesel, T. (1972). Laminar and columnar distribution of geniculo-cortical fibers in the macaque monkey. The Journal of Comparative Neurology, 146(4), 421 –450. Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional architecture of monkey striate cortex. Journal of Physiology (Lond), 195(1), 215– 243. Kawaguchi, Y., & Kondo, S. (2002). Parvalbumin, somatostatin and cholecystokinin as chemical markers for specific GABAergic interneuron types in the rat frontal cortex. Journal of Neurocytology, 31(3– 5), 277– 287. Lachica, E. A., Beck, P. D., & Casagrande, V. A. (1992). Parallel pathways in macaque monkey striate cortex: anatomically defined columns in layer III. Proceedings of the National Academy of Sciences of the United States of America, 89(8), 3566–3570. Livingstone, M., & Hubel, D. (1988). Segregation of form, color, movement, and depth: anatomy, physiology, and perception. Science, 240(4853), 740 –749. Lund, J. S., Lund, R. D., Hendrickson, A. E., Bunt, A. H., & Fuchs, A. F. (1975). The origin of efferent pathways from the primary visual cortex, area 17, of the macaque monkey as shown by retrograde transport of horseradish peroxidase. The Journal of Comparative Neurology, 164(3), 287– 303. Maunsell, J. H., & van Essen, D. C. (1983). The connections of the middle temporal visual area (MT) and their relationship to a cortical hierarchy in the macaque monkey. Journal of Neuroscience, 3(12), 2563–2586.
Meskenaite, V. (1997). Calretinin-immunoreactive local circuit neurons in area 17 of the cynomolgus monkey, Macaca fascicularis. The Journal of Comparative Neurology, 379(1), 113 –132. Porter, J. T., Johnson, C. K., & Agmon, A. (2001). Diverse types of interneurons generate thalamus-evoked feedforward inhibition in the mouse barrel cortex. Journal of Neuroscience, 21(8), 2699–2710. Salin, P. A., & Bullier, J. (1995). Corticocortical connections in the visual system: structure and function. Physiological Reviews, 75(1), 107 –154. Sawatari, A., & Callaway, E. M. (2000). Diversity and cell type specificity of local excitatory connections to neurons in layer 3B of monkey primary visual cortex. Neuron, 25(2), 459–471. Shapley, R., Hawken, M., & Ringach, D. L. (2003). Dynamics of orientation selectivity in the primary visual cortex and the importance of cortical inhibition. Neuron, 38(5), 689 –699. Sherman, S. M., & Guillery, R. W. (1998). On the actions that one nerve cell can have on another: distinguishing ‘drivers’ from ‘modulators’. Proceedings of the National Academy of Sciences of the United States of America, 95(12), 7121–7126. Shipp, S. (2003). The functional logic of cortico-pulvinar connections. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 358(1438), 1605–1624. Somogyi, P., Tamas, G., Lujan, R., & Buhl, E. H. (1998). Salient features of synaptic organisation in the cerebral cortex. Brain Research Brain Research Review, 26(2–3), 113– 135. Swadlow, H. A. (2002). Thalamocortical control of feed-forward inhibition in awake somatosensory ‘barrel’ cortex. Philosophical Transactions of the Royal Society of London. Series B: Biological Sciences, 357(1428), 1717–1727. Swadlow, H. A. (2003). Fast-spike interneurons and feedforward inhibition in awake sensory neocortex. Cerebral Cortex, 13(1), 25– 32. Tamas, G., Buhl, E. H., Lorincz, A., & Somogyi, P. (2000). Proximally targeted GABAergic synapses and gap junctions synchronize cortical interneurons. Nature Neuroscience, 3(4), 366 –371. Troyer, T. W., Krukowski, A. E., & Miller, K. D. (2002). LGN input to simple cells and contrast-invariant orientation tuning: an analysis. Journal of Neurophysiology, 87(6), 2741–2752. Troyer, T. W., Krukowski, A. E., Priebe, N. J., & Miller, K. D. (1998). Contrast-invariant orientation tuning in cat visual cortex: thalamocortical input tuning and correlation-based intracortical connectivity. Journal of Neuroscience, 18(15), 5908–5927. Van Essen, D. C., & Deyoe, E. A. (1994). Concurrent processing in the primate visual cortex. In M. S. Gazzuniga (Ed.), (pp. 383– 400). The cognitive neuroscience, Cambridge, MA: MIT Press. Van Essen, D. C., & Maunsell, J. H. R. (1983). Hierarchical organization and functional streams in the visual cortex. Trends in Neuroscience, 6(9), 370 –375. Vruwink, M., Schmidt, H. H., Weinberg, R. J., & Burette, A. (2001). Substance P and nitric oxide signaling in cerebral cortex: anatomical evidence for reciprocal signaling between two classes of interneurons. The Journal of Comparative Neurology, 441(4), 288–301. Wiesel, T. N., Hubel, D. H., & Lam, D. M. (1974). Autoradiographic demonstration of ocular-dominance columns in the monkey striate cortex by means of transneuronal transport. Brain Research, 79(2), 273 –279. Wiser, A. K., & Callaway, E. M. (1996). Contributions of individual layer 6 pyramidal neurons to local circuitry in macaque primary visual cortex. Journal of Neuroscience, 16(8), 2724–2739. Wiser, A. K., & Callaway, E. M. (1997). Ocular dominance columns and local projections of layer 6 pyramidal neurons in macaque primary visual cortex. Visual Neuroscience, 14(2), 241 –251. Yabuta, N. H., & Callaway, E. M. (1998). Functional streams and local connections of layer 4C neurons in primary visual cortex of the macaque monkey. Journal of Neuroscience, 18(22), 9489–9499.