This review summarizes the local circuit, interlaminar connections in adult mammalian neocortex. These were first demonstrated with anatomical techniques, which indicate some of the exquisite spatial precision present in the circuitry. Details, such as the class(es) of neurons targeted by some of these projections, have begun to be added in studies that combine paired/triple intracellular recordings with dye-filling of connected neurons. Clear patterns are emerging from these studies, with ‘forward’ projections from layer 4 to 3 and from 3 to 5 targeting both selected pyramidal cells and interneurons, while ‘back’ projections from layer 5 to 3 and from 3 to 4 target only interneurons. To place these data in a wider context, the major afferent inputs to and efferent outputs from each of the layers are discussed first.
Afferent Inputs to Neocortex
The afferent input to neocortex that has probably been the most thoroughly studied to date is that from the thalamus; this is partly because many studies have focused on primary sensory regions of neocortex. Although this review does not attempt to describe the many afferent systems, the complexity of the arrangement of thalamocortical inputs, particularly in species such as primates, where the sublaminae are clearly identifiable, indicates that the postsynaptic targets for each of these parallel ascending systems, carrying separate, but correlated information streams, may be different. One way in which the primate thalamus can be envisaged, which solves a number of previous apparent inconsistencies, is as a continuous matrix of calbindin-containing neurons within which regions or cores rich in parvalbumincontaining cells can be found, particularly in sensory and motor nuclei [for a review, see Jones (Jones, 2001)]. Nuclei rich in parvalbumin-containing ‘cores’ are the so-called ‘specific’ thalamic nuclei, relaying precise sensory information about one sensory modality to a specific region of neocortex and preserving, for example, retinotopic or somatotopic segregation. Nuclei with few or no parvalbumin-containing neurons are often termed ‘non-specific’ nuclei receiving more diffuse inputs and whose cortical projections do not respect areal boundaries. For example, while dorsal LGN (lateral geniculate nucleus) parvalbumin-containing ‘core’ cells project only to V1, the calbindin-containing ‘matrix’ cells of the LGN project to V1, V2 and beyond. Parvalbumin-containing cells project to the middle layers (layers 3 and 4) and to layer 6, while the calbindincontaining cells project primarily to layer 1 and to a lesser degree to layer 3. In the LGN, parvalbumin-containing cells are concentrated in the six principal parvocellular and magnocellular layers, while calbindin cells are concentrated in the s laminae and interlaminar plexuses and project to layer 1 (especially from layer K1) and to layer 3 in cytochrome oxidase rich blobs (especially from layer K3) (Ding and Casagrande, 1997). In the ventral posterior complex of the thalamus, only the VPM and VPL (ventro-posterio-medial and lateral) nuclei, the principal relays for somatosensory information, contain parvalbumin cells and like the principal layers of the LGN these regions stain intensely for cytochrome oxidase (Jones, 2001).
Further sublamina specific segregation of afferent inputs can be seen, especially in primary visual cortex (V1). Axons from the magnocellular layers of the LGN, with large receptive fields, project to a subdivision of layer 4C, layer 4Cα, and sometimes to only one half of the depth of that sublayer, as well as to layer 6. This input to layer 4Cα probably underlies the vigorous responsiveness of its cells to motion. Axons from the parvocellular layers with small receptive fields, on the other hand, terminate in layer 4Cβ, 4A, 3B, and in layer 6 [for a review, see Lund (Lund 1988)]. Layer 4B receives no direct thalamocortical input. There are, however, interspecies and/or interareal differences. In rat barrel cortex, for example, specific thalamocortical input spans the depth of layer 4 [for a review, see Castro-Alamancos and Connors (Castro-Alamancos and Connors, 1997)].
Inputs from other regions of cortex constitute a large proportion of the afferent input to any cortical region. Some of these result from long horizontal axon collaterals originating and running in layers 3–6. Others, such as the trans-callosal inputs, which involve layers 2–6, but are more concentrated in layers 2 and 3, and connections between more distant regions, travel via the white matter. ‘Forward’ projections from primary sensory regions, such as those from V1 to other visual areas, target layers 3 and 4 [e.g. in cat (Lowenstein and Somogyi, 1991)], forming the equivalent of a thalamocortical input in regions that lack a specific input from the core regions. So-called ‘back’ projections from what are often referred to as higher order areas, primarily target layer 1, with layers 3 and 6 also receiving input (Rockland and Drash, 1996). Again, this segregation of corticocortical ‘forward’ and ‘back’ projections indicates that they target very different neuronal elements in recipient regions. These major afferent projections to an idealized primary sensory region are summarized in Figure 1.
It has been suggested (White, 1986) that all cells that have dendrites in layer 4, including layer 4 spiny stellate cells, layer 4 pyramidal cells, the apical dendrites of layer 5 and 6 neurons and the basal dendrites of lower layer 3 neurons, receive thalamocortical input in layer 4. This with the possible exception of layer 6 pyramids in rodent barrel cortex, whose apical dendrites pass through the walls of the barrels where there are no thalamic axon terminals. That this innervation is, however, far from indiscriminate is indicated by the selective targeting of primate layer 4 subdivisions by different types of thalamic afferents (Fig. 1) and the sometimes sublayer specific arborization patterns of the apical dendrites of some deep layer pyramidal cells in layer 4. There may also be significant target selectivity amongst the inhibitory interneuronal population, with VIP (vasoactive intestinal polypeptide) containing (Hajos et al., 1997) and parvalbumin-containing interneurons (Staiger et al., 1996) being the primary layer 4 interneuronal targets of thalamic afferents in rodents.
A commonly reported difference between the thalamocortical and corticocortical inputs to layer 4 is the size of the boutons involved. Thalamic axons provide large en passant boutons that terminate predominantly on dendritic spines, while layer 6 axons provide complex side-spine arrays with small boutons [in cat and primate, for a review see Lund (Lund, 1988)]. This difference has been used to estimate the relative contributions made by different inputs to the innervation of layer 4 cells. Despite their large size and apparently dominant role in the activation of these cells in primary sensory regions, thalamic terminals have been estimated to contribute only 6% of the synapses onto spiny stellate neurons in layer 4, while layer 6 pyramidal cells contributed 45%, with 28% originating from spiny cells in cat layer 4 (Ahmed et al., 1994).
Pyramidal Cells and Efferent Projections
Although pyramidal cells share many distinctive anatomical features, they are far from a homogeneous group. Even within a layer there are several morphological subtypes, often projecting to different cortical and subcortical regions. Layer 5 (particularly 5A) contains large, burst-firing pyramidal cells with long apical dendrites that form an extensive tuft in layers 2 and 1. These large cells project, depending on cortical area, to subcortical regions such as the superior colliculus and the pons (Wang and McCormick, 1993). It is only these larger pyramidal cells in layer 5 that could access directly the afferent inputs to the superficial layers, including the back projections from higher cortical regions and inputs from calbindin-containing thalamic matrix cells. The apical dendrites of smaller, regular-spiking pyramids rarely extend beyond layer 3. These smaller pyramids in upper layer 5, project, for example, to the striatum [in primate (Catsman-Berrevoets and Kuypers, 1978)], those in lower layer 5 project to non-specific thalamic nuclei [in rodents, (White and Hersch, 1982)], forming a prominent source of efferents to the pulvinar, where they form quite different axonal arbours from those originating in large layer 5 pyramids, as well to as the superior colliculus [in primate (Rockland, 1996)]. Layer 5 cells do not, however, innervate the inhibitory thalamic nucleus, nucleus reticularis thalami (nRT) or the specific thalamic nuclei [in primate, for a review see Jones (Jones, 2001)].
Corticothalamic cells in upper layer 6 do project to the ‘specific thalamic’ nuclei and cores as well as to nRT, while lower layer 6 pyramidal cells project to both specific and nonspecific nuclei (Jones, 2001). For example, in rat somatosensory cortex (S1), corticothalamic cells in upper layer 6 project to VPM and have apical dendrites that terminate in layer 4 as well as a narrow axonal arbour that ascends to layer 4 and/or lower layer 3 where it ramifies. In upper layer 6 of primate V1, both large Meynert cells (which, unlike large layer 5 pyramids, lack an apical tuft in layers 1/2) and small pyramids project to area MT, to the superior colliculus and to area V2 [in primates (Rockland, 1996)]. Corticothalamic cells in lower layer 6 that project both to the specific nucleus VPM and to more posterior non-specific nuclei (Po) are also small, short pyramids, but their apical dendrites terminate in layer 5 and their axonal arbours are broader and ramify in layer 5 rather than in layers 4 and 3. Both of these groups might have access to thalamocortical input in layer 6, but only those projecting exclusively to VPM would have direct access to specific thalamic inputs in layer 4. Corticocortical cells in layer 6 display a range of morphologies, small, short pyramids, modified pyramids, inverted pyramids and spiny bipolar cells, the latter two having dendritic trees largely confined to layer 6, but which sometimes extend into the white matter. Their local axon collaterals arborize primarily in layers 5 and 6 in S1 and long horizontal branches extend into motor cortex, S2 and the callosum [in rodents (Zhang and Deschenes, 1997)]. The two subdivisions of layer 6 probably represent the different origins of layer 6 cells, the upper division (6A) with more vertically oriented pyramidal cells originating from the cortical plate, while the lower division (6B) with more horizontally oriented cells originates from the primordial plexiform layer.
Interlaminar Excitatory Projections
The projection patterns of pyramidal axons and multiple-field potential recordings in vitro (Bode-Greuel et al., 1987) have provided the basis for simple circuit diagrams of cortex. In their simplest form, specific thalamocortical input arriving in layer 4 (and 6) is perceived as being relayed from layer 4 to 3 (and thence to layer 2), from layer 3 to layer 5 and from 5 to 6, the processed information then leaving layer 3 for other cortical regions and from the deeper layers to other cortical and subcortical regions. That the real situation is rather more complex than this is indicated in the foregoing discussion, but this has been a useful starting point for many studies and discussions. The discussion in this section is based on anatomical studies of axonal arbours and the projections described as ‘weak’ are those that involve relatively small numbers of axon collaterals and relatively few synaptic boutons. In many cases the types of cells they contact, the laminar origin of these cells and the relative functional strengths of the inputs have yet to be determined. For example, if only specific classes of neurons are contacted in a given layer, what appears to be a numerically weak input to that layer could have a powerful effect on that specific class of neurons. Nor, until a large sample of target neurons is identified, can we be sure that all the neurons receiving a specific input in a given layer have their somata in that layer.
Excitatory Projections from Layer 4
In addition to a local axonal arbour of varying density in layer 4, the axons of layer 4 spiny excitatory cells send focused projections to layer 3, long horizontal and oblique branches within layer 4 and extending into layer 3, and less dense, more tightly focused projections to the deeper layers in all species studied (Lund, 1973; Valverde, 1976; Parnavalas et al., 1977; Feldman and Peters, 1978; Gilbert and Wiesel, 1979; Gilbert, 1983; Burkhalter, 1989; Anderson et al., 1994), with layer 4 pyramidal axons ramifying more extensively than spiny stellate axons in layer 6 (A.P. Bannister, unpublished). Where there are clear sublaminar divisions, the projections from layer 4 spiny cells are sublayer specific both in the origin and in the termination of the axons involved. For example, upper layer 4Cβ and lower 4Cα project to layers 3B and 4A, while upper layer 4Cα projects to layer 4B. Sublayers 4A and 4C (α and β) send weak projections to layer 5A in primate visual cortex, while layers 3A and 2 receive excitatory input from layer 4B, the layer 4 subdivision that does not receive direct thalamocortical input [for a review, see Lund (Lund, 1988)].
Excitatory Projections from Layer 3
‘Patchy’ horizontal connections within layer 3 are formed by long horizontal axon collaterals of layer 3 pyramidal cells (and interneurons). These largely contact the dendritic spines of presumed excitatory cells at distant sites in cat and primate (Rockland and Lund, 1982; Kisvárday et al., 1986; McGuire et al., 1991). Layer 3 pyramidal axons in all species studied, arborize primarily in layers 2/3 and 5, passing through layers 4 and 6, but ramifying little or not at all in these layers (Lorente de Nó, 1922; O’Leary, 1941; Spatz et al., 1970; Gilbert and Wiesel, 1983; Burkhalter, 1989; Lund et al., 1993; Yoshioka et al., 1994; Kritzer and Goldman-Rakic, 1995; Fujita and Fujita, 1996; Kisvárday et al., 1986).
Excitatory Projections from Layer 5
Layer 5 pyramidal axons arborize most densely within layer 5, but can also send projections to all other layers in rat and primate (Burkhalter, 1989; Keller, 1993; Yoshioka et al., 1994; Fujita and Fujita, 1996). Some of these cells have very long horizontal branches within layer 5 and in layer 6 in cat visual cortex (Gilbert and Wiesel, 1979), which can give rise to ascending branches distant from the soma. Layers 3A and 2 receive tightly focused but weak projections from layers 5A (and 6) and a broader axonal projection from layer 5B.
Excitatory Projections from Layer 6
Layer 6 receives no marked projections from more superficial layers, rather all layers seem to send weak, tightly focused inputs to this layer. In addition to the long horizontal and oblique collaterals of some layer 6 pyramidal axons (particularly the corticocortical cells) (Zhang and Deschenes, 1997) (O.T. Morris, unpublished), layer 6 pyramidal axons (particularly those of specific corticothalamic cells) often arborize densely in layer 4 [in cat (Gilbert and Wiesel, 1979)] [for sublayer selective arbours in primate see also Wiser and Callaway (Wiser and Callaway, 1996)] and in layer 5, with cat layer 6 cells classified physiologically as simple cells projecting to L4 and complex cells to layers 2 and 3 (Hirsch et al., 1998). The long horizontal branches of corticocortical layer 6 cells also give rise to ascending vertical or oblique collaterals that innervate layers 2–4.
Interneurons in the Neocortex
The first attempts to categorize non-pyramidal cells in neocortex were based on the dendritic and axonal morphology visible with Golgi staining. More recently, dye-filling of single cells has demonstrated that the axonal arbours of some of these cells are more complex and/or more extensive than was originally apparent, while immuno-cytochemical and single-cell PCR (polymerase chain reaction) studies have provided an additional way of subdividing inhibitory interneurons. On the basis of their neurochemistry, there appear to be three, largely nonoverlapping groups of interneurons in adult mammalian cortex, those that express parvalbumin, which are typically proximally targeting cells, those that express somatostatin, calbindin and/or NPY (neuropeptide Y), which typically target more distal compartments of pyramidal dendrites and those that express calretinin, VIP (vasoactive intestinal polypeptide) and/or CCK (cholecystokinin) in a variety of combinations (Kawaguchi and Kubota, 1997) and include both proximally and distally targeting subtypes. Some of the classes described below are extremely distinctive and readily identifiable with even a cursory light microscopic examination of the axonal and/or dendritic morphology. Others, particularly those whose dendrites and axons are more or less evenly distributed around the soma, can be difficult to separate into distinct subcategories without information about their neurochemistry, and/or target preferences.
Interneurons That Target Proximal Portions of Pyramidal Cells
Chandelier or axo-axonic cells (Somogyi, 1977) are one of the most distinctive classes, with a stereotypical axonal arbour consisting of horizontal or oblique branches bearing short, vertically oriented collaterals that give rise to strings of large boutons that innervate the axon initial segments of pyramidal cells. In all species studied, these cells are found in layers 2–6, with a bias towards more superficial layers (Valverde, 1983). Many, though not all, chandelier axon terminals are parvalbumin immunopositive (Lewis and Lund, 1990; De Felipe and Fariñas, 1992). Their axonal arbours are largely local to the soma/ dendrites, but not necessarily concentric with the dendritic tree and can terminate in more than one non-contiguous layer, e.g. in layers 4Cα and 5A in primate visual cortex (Lund, 1987).
Basket cells, as a group, cover a vast range of morphologies, some with spherical axonal arbours confined to the layer of origin, some with arbours that are split between two noncontiguous layers and some with elongate arbours spanning several sequential layers. Some are large cells with long horizontal (and sometimes vertical) myelinated axonal branches (Fig. 2B, cell k), others are very small with discrete axonal and dendritic arbours which may be partially and sporadically myelinated; the single common distinguishing feature being their preferential innervation of the somata and proximal dendrites of pyramidal cells. It can therefore be difficult to distinguish basket cells from dendrite-targeting cells with similar gross morphology unless their targets can be identified [see Wang et al. (Wang et al., 2002), for a detailed analysis of three classes of basket cells in immature rat cortex]. There are two major neurochemical groups of basket cells, those that contain parvalbumin and are predominantly fast-spiking cells and the largely regular-spiking basket cells that contain CCK and sometimes VIP (Kawaguchi and Kubota, 1997, 1998; Kubota and Kawaguchi, 1997). At one time, both parvalbuminand CCKcontaining cells were thought to be exclusively proximally targeting in both neocortex and hippocampus. However, recent studies in hippocampus have demonstrated that although the majority of cells in both groups selectively innervated the somata and proximal dendrites of pyramidal cells, a significant proportion were dendrite-targeting cells (e.g. bistratified cells). The CCK-immunopositive cell in Figure 2B (cell m) for example, innervated fourthand fifth-order dendrites of a postsynaptic pyramidal cell. Nor were parvalbumin-containing cells exclusively fast-spiking, nor CCK-containing cells exclusively regularspiking cells (Pawelzik et al., 2002). Thus, although a large population of fast-spiking, parvalbumin-containing, basket cells exists, there are significant numbers of parvalbumin-containing cells that are neither basket (nor chandelier) cells nor fastspiking cells and there are fast-spiking cells that are neither parvalbumin-containing nor proximally targeting. In the present discussion about circuitry, layer 4 parvalbuminand VIPcontaining interneurons are of particular interest as the major interneuronal recipients of specific thalamocortical input. It is not yet clear, however, whether this input selectively targets subdivisions of these populations. For example, are the VIPcontaining cells basket cells that also contain CCK, or dendritetargeting double bouquet cells?
Interneurons That Target the Dendrites of Pyramidal Cells
Double bouquet cells, first named for their bipolar or bitufted dendritic arbour, are most readily identifiable by their narrow bundles of unmyelinated, vertically oriented, descending and sometimes ascending axon collaterals. Other interneurons have one or more descending axon collaterals, but these bundles are instantly recognizable under the light microscope and are distinctive in cat and primate cortex [e.g. Somogyi and Cowey (Somogyi and Cowey, 1984), their Fig. 2]. The equivalent cell type in rodents may have less tightly concentrated axon bundles (Kawaguchi and Kubota, 1997). Dye-filled double bouquet cells can also be seen to have an extremely dense, local axonal arbour (e.g. Tamas et al. (Tamas et al., 1998), Fig. 2A, cell a) broader than the vertical bundles and sometimes spanning more than one layer. They target dendritic spines, small calibre dendrites of spiny cells and the somata and dendrites of smooth cells. They are found in layers 2–4 and, like several other types of dendritetargeting interneurons, are often immunoreactive for calbindin, calretinin and/or VIP (DeFelipe and Jones, 1992; Conde et al., 1994; Del Rio and De Felipe, 1997; Kawaguchi and Kubota, 1997; Peters and Sethares, 1997).
Another distinctive class of bitufted, calbindin(in layer 5) and/or somatostatin-immunoreactive (Conde et al., 1994; Gabbott et al., 1997; Kawaguchi and Kubota, 1997) dendritetargeting interneurons are the Martinotti cells, most prevalent in the deep layers but reported also in layer 3 (Fairén et al., 1984) (see the Martinotti-like cell, Fig. 2A, cell c) [(Gupta et al., 2000), their Fig. 4]. The distinguishing feature of these cells is their prominent axonal projection to layer 1. Many fine, unmyelinated, ascending axon collaterals that fan out as they ascend, bear en passant boutons in intermediate layers, with some reaching, and branching along layer 1. Again, many other interneurons have ascending axon collaterals – it is the prominent axonal arbour in layer 1 that distinguishes Martinotti cells.
The third, clearly identifiable class of dendrite-targeting (Kisvárday et al., 1990) interneurons are the late-spiking (Kawaguchi and Kubota, 1997), neurogliaform, or spider web cells (Fairén et al., 1984) (Fig. 2A, cell d; 2B, cell l). These cells have very small but extremely dense, convoluted dendritic and axonal arbours that are often concentric.
In addition to these, at least in rodent layers 2/3 and 5, are sparsely spiny, burst-firing (or low-threshold-spiking) interneurons with large, but relatively sparse dendritic and axonal arbours (Thomson et al., 1995). These cells target distal pyramidal dendrites and are again calbindinor somatostatinimmunoreactive (Kawaguchi and Kubota, 1993, 1997) and appear to be the neocortical equivalent of the very readily identifiable OLM (oriens-lacunosum moleculare) cells found in the hippocampus, which target the most distal dendrites of pyramidal cells (Blasco-Ibanez and Freund, 1995; Ali and Thomson, 1998) and receive, like the neocortical equivalent, powerfully facilitating EPSPs from neighbouring pyramidal cells.
VIP and calretinin are often co-expressed in interneurons, but this population is largely distinct from the VIP/CCK expressing cells, especially in the superficial layers (Kawaguchi and Kubota, 1997). In layers 1–3, calretinin immunopositive interneurons are, like many VIP-positive cells, typically bipolar, with vertically oriented dendrites that can reach and run horizontally in layer 1. Their axons are fine and often send collaterals to layers 5 and 6. Calretinin-positive axon terminals target fine GABA-negative dendrites, but preferentially innervate other interneurons in the superficial layers, particularly other calretinin-containing cells. In this they resemble one class of interneuron-specific interneurons in the hippocampus [for a review, see Freund and Buzsáki (Freund and Buzsáki, 1996)]. In the deeper layers, however, the preferred targets of calretinin-containing terminals include a large proportion of GABA-negative somata [in rat (Gonchar and Burkhalter, 1999)].
It is important to note that while many of the interneurons belonging to the classes described above are readily identifiable if well filled with reaction product, there are many that do not comply with these classifications, neurons that meet one or two, but not all the criteria for a particular classification or whose near-spherical axonal and dendritic arbours defy classification on gross morphological criteria. A range of interneurons recorded and dye-filled in layers 3 and 4 of cat visual cortex are shown in Figure 2A, B to illustrate this point. It is perhaps worth noting here that in young mammals many interneurons that will bear smooth dendrites once mature, are sparsely, or even densely, spiny. In addition, many of the neuronal markers that are widely used for interneuronal classification are either not yet expressed, or are expressed in different neurons from those that will contain them on reaching maturity. Although the functions of these temporary spines and transient calcium-binding proteins and peptides and the reasons for their appearance and disappearance are fascinating questions, for dendritic spines and cellular neurochemistry to be unambiguous categorization criteria, it may be important to study mature cortex.
Interlaminar Inhibitory Projections
In addition to the dense local axonal arbours of many interneurons and the long horizontal collaterals of some of the larger cells, many of the interlaminar projections of excitatory neurons are accompanied by inhibitory axons. However, individual interneurons, particularly those that target relatively proximal portions of pyramidal cells, often have highly layeror even sublayer-specific axonal projection patterns. The two most readily recognizable dendrite-targeting interneuronal classes appear less selective, Martinotti cells with ascending collaterals that innervate all layers between their origin and layer 1, and double bouquet cells with descending collaterals that innervate all layers between their origin and layer 5 or 6. In Figure 3 therefore the cartoons depicting dendrite-targeting inhibitory projections are based on the known projections of these two classes, while those depicting more proximally targeting cells are based on cells that have either been unambiguously identified as basket, chandelier or proximal dendrite-targeting cells, or on cells with similar morphology.
Inhibitory Inputs to Layer 4
As a population, layer 5A interneurons innervate all the thalamic recipient layers (4A, 4Cα and β, and 6 and 3), as well as layers 2 and 1. The projection to layer 4 is predominantly to the thalamic recipient sublayers 4C and 4A in primates (Lund, 1988). Some of the layer 5 cells that innervate layer 4 (rat), have few or even no axon collaterals in other layers and either resemble clutch cells with dense, narrow axonal arbours in layer 4, or have a dense local arbour in layer 4 plus prominent horizontal, myelinated collaterals over a millimetre in length that give rise to short vertical or oblique branches [in rat (Thomson et al., 1996)]. Other layer 5 interneurons with ascending axons innervate several, or all, layers between 2 and 5. Only a minority of layer 3 interneurons appear to send significant descending inhibition into layer 4, however. The most prominent of these being layer 3 double bouquet cells with a dense local axonal arbour in layers 3 and 4 and a narrower vertical projection (the ‘horse’s tail’) into the deeper layers. Aspinous layer 6 cells projecting to layer 4Cα have also been described (Lund, 1988).
Inhibitory Inputs to Layer 3
Layer 4 interneurons with a wide range of morphologies and including both soma/proximal dendriteand more distal dendrite-targeting cells, innervate layer 3B. However, unlike layer 5 interneurons, the axons of layer 4 interneurons rarely, if ever, penetrate above layer 3 and typically not beyond layer 3B. Layer 3B receives inhibitory input from interneurons in layers 4C in primate (Lund, 1987, 1988) and upper layer 4 in rat and cat (Thomson et al., 2002). In rat (Thomson et al., 1996) and in primate (Lund, 1988) there are layer 5 interneurons (including proximal dendrite-targeting cells) whose axons arborize densely in layer 5 and in layers 2 and 3, but whose single ascending collateral passes through layer 4 with little or less arborization there, while other layer 5 interneurons (including basket cells and Martinotti cells) innervate several, or all, layers between layer 2 (or layer 1 for Martinotti cells) and layer 5.
Inhibitory Inputs to Layers 5 and 6
In primate visual cortex, inhibitory cells in both subdivisions of layer 4C also innervate layers 5A and 6, although the axons of neither subdivision ramify in 5B. The descending inhibitory projections from layers 3 and 4 to the deeper layers are often weak and/or focused, although layer 4A is described as sending a strong inhibitory input to 5A (Lund, 1988). Many interneurons in layers 3 and 4 have either local axonal arbours, restricted to the layer of origin, or ascending arbours. However, in addition to the double bouquet cells, which send tightly focused axonal bundles to the deeper layers, some other interneurons in these layers, including those described as large basket cells, send one or two, often myelinated, axon collaterals to layer 5 where they generate discrete arbours (Fig. 2A, cell f; 2B, cell m).
Although the main descending trunk of a pyramidal axon becomes myelinated soon after leaving the initial segment, the local, intra-cortical axon collaterals of spiny excitatory cells are less frequently myelinated, even those that project for millimetres horizontally can be relatively fine and largely unmyelinated. Each millimetre of fine, unmyelinated axon could introduce a conduction delay of at least 2 ms. In contrast, the horizontal and some of the vertical projections of many proximally targeting interneuronal axons are much thicker and strongly myelinated. Conduction delays in these collaterals could therefore be more than an order of magnitude briefer, indicating that the horizontal spread of inhibition will often be faster than the horizontal spread of excitation.
Identification of the Targets of Interlaminar Projections
Some of the details missing from maps based on axonal projection patterns include the classes and subclasses of neurons preferentially innervated by these axons in each of the layers and sublayers, the properties of the connections and the relative probabilities of one type of connection, or another, being made. Recently, attempts have been made to provide some of this information by recording one neuron (that is dye-filled and later identified morphologically) and activating synaptic inputs to it by focal release of caged glutamate in a cortical slice. This is more selective than electrical stimulation as a method for activating local corticocortical inputs, since it will not activate axons en passage and has provided some valuable clues [e.g. Dantzker and Callaway (Dantzker and Callaway, 2000)]. It is not, however, always possible to distinguish between inputs that result from neurons activated at their somata from those that were activated at dendritic sites, making the location of the presynaptic soma(ta) ambiguous. It can also be difficult to distinguish monosynaptic from disynaptic inputs, since the timing of the presynaptic spikes cannot be determined with accuracy. For the unambiguous identification of both the preand the post-synaptic neuron and to study the properties of identified connections in any detail there appears as yet no less labour-intensive alternative to dual intracellular recordings with biocytin labelling. We therefore focus here on data obtained in this way and particularly on interlaminar connections in adult neocortex. Intra-laminar connections have been more intensely studied with these methods and are summarized in Figure 4; further details can be found in the cited literature.
Interlaminar Connections Between Layers 3 and 5
That interlamina connections can be highly selective was indicated when the postsynaptic pyramidal targets of layer 3 pyramidal axons in layer 5 were studied with dual intracellular recordings in adult rat neocortex (Thomson and Bannister, 1998). All the postsynaptic layer 5 pyramidal cells identified were large, burst-firing cells with long apical dendrites forming a tuft in layers 1/2. Layer 5 interneurons were also innervated by layer 3 pyramidal axons, including one interneuron that had strong axonal arbours in both layers 5 and 3 (Thomson et al., 1996; Thomson and Deuchars, 1997), but the smaller layer 5 pyramidal cells that lacked a tuft in the superficial layers were not innervated by layer 3 pyramidal axons. Large layer 5 pyramidal cells can access inputs to layer 1 both directly via their own apical dendritic tufts and indirectly via their dense inputs from layer 3 pyramidal cells in the same column. These would include both the ‘back’ projections from ‘higher’ cortical areas and the inputs from ‘non-specific’ thalamic matrix cells. The smaller layer 5 pyramidal cells would not have access to these inputs via either route, since their apical dendrites do not reach layer 1 and they are not activated by layer 3 pyramidal cells. The extent to which they receive this information from large layer 5 cells is unclear. In our earlier study in rat, which included both groups, this question was not directly addressed (Thomson et al., 1993; Deuchars et al., 1994). It is only possible from existing data to state that while connections between small, regularspiking and large, burst-firing cells were frequently recorded, burst-firing cells were more commonly found to be the postsynaptic cell of the partnership (with a ratio of ∼1:10, A.M. Thomson and A.P. Bannister, unpublished data). Parallel, but separate, information processing streams appear to lie side by side within a layer and within a column. That these streams may also be differentially regulated by local interneurons is indicated by the differential distributions of inhibitory terminals on somata and axon initial segments of layer 5 pyramidal cells with different subcortical targets in rodent [e.g. Czeiger and White (Czeiger and White, 1997)].
Equally strikingly, rat layer 3 pyramidal cells almost never received inputs from ascending layer 5 pyramidal axons and the inputs that did occur were extremely weak (Thomson and Bannister, 1998). These ascending axons therefore select other targets in layer 3, the distal dendrites of other layer 5 pyramidal cells (Deuchars et al., 1994) and certain subclass(es) of interneurons (Dantzker and Callaway, 2000).
Interlaminar Connections Between Layers 3 and 4
In a more recent triple recording study of the connections within and between layers 3 and 4, a similarly selective pattern emerged (Thomson et al., 2002). The excitatory input to layer 3 pyramidal cells from layer 4 spiny excitatory cells was as frequent and as strong as their intra-laminar inputs from other layer 3 pyramidal cells in both adult rat and cat neocortex. However, in none of the pairs studied in either rat or cat was the layer 3 pyramidal cell presynaptic to the layer 4 excitatory cell. It is not perhaps surprising that descending layer 3 axons, which typically do not arborize in layer 4, do not innervate the smaller layer 4 cells such as spiny stellates and small interneurons, whose dendrites are largely confined to that layer. It is more surprising that their collaterals in layer 3 do not contact the apical dendrites of layer 4 pyramidal cells which pass through layer 3. The axons of layer 3 pyramidal cells can therefore distinguish between the dendrites of layers 2/3 pyramids which they do innervate and the apical dendrites of layer 4 pyramidal cells which they do not innervate. Alternatively, of course, the apical dendrites of layer 4 pyramids can reject innervation by the axons of layer 3 pyramids, but not those of other excitatory layer 4 cells. If these cells are to maintain tightly organized receptive field properties, it might be inappropriate for them to receive more integrated information from the superficial layers. What excitatory inputs these layer 4 pyramidal cells receive on their spiny dendrites in layers 2 and 3 is therefore unclear, possibly ascending layer 4 axons or longer distance inputs from other cortical regions.
Layer 3 pyramidal axons can also distinguish between layer 4 excitatory cells and layer 4 inhibitory cells. Layer 4 contains many different types of interneurons and many of these, with a wide range of morphologies, have dendrites extending into layer 3. Almost half the layer 4 interneurons tested with simultaneously recorded layer 3 pyramidal cells (and one in five of the tested pairs that involved a pyramidal cell in layer 3 and an interneuron in layer 4) were found to receive an excitatory input from layer 3. Many of these, including both basket cells and dendrite-targeting cells, also innervated lower layer 3 and inhibited pyramidal cells there. Where multiple inputs to a single layer 4 interneuron could be studied with triple intracellular recordings, these inhibitory interneurons were found to receive excitatory inputs from both layers 3 and 4 with equivalent frequencies (e.g. Fig. 2B, cell k). These interneurons included parvalbumin-immunopositive interneurons in upper layer 4 that received excitatory inputs from both layers 3 and 4 and innervated both layers. Since parvalbumin-containing interneurons are a major interneuronal target of direct thalamocortical input (Staiger et al., 1996), these cells could integrate this input with activity in the two layers and via their inhibitory outputs coordinate the temporal properties of activity in the two layers (Thomson et al., 2002).
Interestingly, despite their prominent ascending axonal projections and frequent innervation of layer 3 pyramidal cells, layer 4 spiny cell axons may be highly selective in the layer 3 interneurons they target. Only one of nine layer 3 interneurons in cat and one of five in rat tested received an EPSP from a simultaneously recorded layer 4 excitatory cell (31 and 10 tested pairs, respectively). Nor did layer 3 interneurons inhibit layer 4 spiny cells in any of the tested pairs. The single postsynaptic layer 3 interneuron identified in cat (Fig. 2A, cell f) was the largest layer 3 smooth cell recovered and was excited by all three simultaneously recorded layer 4 spiny cells (as well as by a layer 3 pyramidal cell). It was a large parvalbumin immunonegative cell with long, myelinated, horizontal axon collaterals, most of its axonal arbour within layers 2 and 3, but with a single myelinated collateral that descended through layer 4 without branching and arbourized in layer 5. From the available evidence, therefore, many layer 3 interneurons appear to be primarily local circuit neurons, connected to other superficial layer cells, with a minority sending a focused projection to layer 5, but few being connected to layer 4 spiny cells. It will be interesting to determine which of these cells receives a significant excitatory input from layer 5.
Interlaminar Connections Between Layers 4 And 6
The connections between layer 6 and layer 4 are also of great interest as layer 6 can be seen both as an input and as an output layer where thalamic connections are concerned and therefore to involve perhaps both ‘forward’ and ‘back’ connections. Connections between excitatory cells in these two layers have been recorded in one study of adult cat striate cortex (Tarczy-Hornoch et al., 1999). The EPSPs elicited by layer 6 cells were reported to be smaller on average than those elicited by layer 4 spiny cells and unlike the layer 4 to layer 4 connections, exhibited facilitation, rather than depression. That study also reports a single connection from a spiny layer 4 cell to a layer 6 pyramidal cell whose morphology resembled that of cells projecting to the claustrum. It will be of interest to determine which of the several types of layer 6 pyramidal cells activates layer 4 cells (the prediction from the previous discussion being the thalamocortical cells in upper layer 6 which most densely innervate layer 4) and whether interneurons, and interneurons of any specific type(s), are also activated.
A brief synopsis of the connectivity between layers 3, 4 and 5 might therefore be that excitatory ‘forward projections’ (from layer 4 to 3 and from 3 to 5) contact selected pyramidal cells as well as a selected subset of interneurons, while the ‘back projections’ (from 3 to 4 and from 5 to 3) contact only, or primarily, inhibitory interneurons. There are a number of reasons why strong ‘feedback’ excitation might not be valuable. Firstly, thalamocortical input from core or specific thalamic regions is often highly specific in modality or submodality, spatially focused and organized as it enters the recipient layers. This information is modified in various ways and correlated with other inputs as it passes ‘onwards’ through the circuit. Strong ‘back’ projections would confound the precision of this primary signal. In addition, apart from the inherent dangers associated with positive feedback, reverberating excitation would destroy any temporal precision in the code.
Not only are many cortical neurons more readily excitable after a period of hyperpolarization, but cortical synapses have very frequency-dependent properties. Although these temporal properties are beyond the scope of this review, it is useful in this context to remember that the majority of connections between spiny excitatory cells and from these cells to a large proportion of inhibitory interneurons (including many parvalbumincontaining cells) are so called ‘depressing synapses’. Simplistically, the response to the first action potential of a train is relatively strong, but subsequent EPSPs in a high-frequency train are depressed. Inhibitory ‘forward and back’ projections may play a role in suppressing excitability, but a much more interesting possibility is that they coordinate the recipient layers in the temporal domain so that appropriate, behaviourally relevant inputs arrive when the recipient cells are most receptive and when their output synapses are most effective, i.e. after they have been silenced for a few tens of milliseconds. Fast-spiking interneurons and the dense connections between them are thought to form the substrate for the generation of the fast gamma oscillations that occur during attention and arousal. Although pyramidal cells rarely fire repeatedly on every cycle of this rhythm, their firing, particularly the coincident firing of cells responding to a single object, is phase-locked to this rhythm. The thick, strongly myelinated interlaminar axon collaterals of some interneurons could play an important role in coordinating this activity across layers. The possibility that the axons providing intracortical connections represent ‘delay lines’ with highly tuned latencies, depending on their diameters and myelination, and contribute to, for example, motion detection is also an intriguing possibility.
Work from this laboratory cited herein was supported by the Medical Research Council, Novartis Pharma (Basel) and the Wellcome Trust. We would like to thank Miss Audrey Mercer and Mrs Hilary McPhail for assistance in preparing figures.