Introduction
Connectivity maps allow insights into the structure of the brain, for instance through graph-theoretical analyses (Jouve et al.
1998; Rubinov and Sporns
2010), and help to create hypotheses on neural processing strategies (Maunsell and Newsome
1987; Felleman and Van Essen
1991; Nassi and Callaway
2009). For instance, experimental knowledge about laminar patterns of connectivity (Felleman and Van Essen
1991; Markov et al.
2014b) in combination with experimental studies on cortical activity (van Kerkoerle et al.
2014; Bastos et al.
2015a) have inspired theories about hierarchical processing and communication between cortical areas (Bastos et al.
2012,
2015b). Furthermore, connectivity maps provide a structural basis for dynamical models of the brain. They have been derived at different levels of detail and for different species such as the mouse (Oh et al.
2014) and macaque monkey (Stephan et al.
2001; Bakker et al.
2012). Such maps inherently possess uncertainties, for example, due to gaps in the experimental data or deformations associated with the mapping to standard brains. Consequently, there is an ongoing need for improvement, gradual refinement, and theoretical integration.
The connectivity of the brain is closely linked to its cellular architecture. Systematic relations have been identified in cortex using the notion of architectural types (Barbas
1986; Barbas and Rempel-Clower
1997), which classify the distinctiveness of the laminar cortical architecture as well as the thickness of the granular layer (Dombrowski et al.
2001). A set of connectivity features, including the existence or absence of connections and laminar patterns of cortico-cortical connections, are linked to structural differences between areas (Barbas and Rempel-Clower
1997; Hilgetag and Grant
2010; Hilgetag et al.
2016; Beul et al.
2017). The concept of architectural types represents a discretization of a continuum of structural features across areas (von Economo and Van Bogaert
1927). Types relate also to neuron density, as types with low ordinal number have low overall neuron density. Statistical relationships between cortical architecture and connectivity may have a developmental origin, with areas of low type developing earlier and having a larger time window for interconnecting with other areas (Barbas and García-Cabezas
2016; Beul et al.
2017). Regardless of the underlying cause, such regularities help to fill gaps in existing connectivity maps.
Network science describes the connectivity of neuronal networks in different ways, for instance in terms of total numbers of synapses, pairwise connection probabilities, or in- and outdegrees of nodes, but also by more abstract measures of connection strength (Hagmann et al.
2007; Wedeen et al.
2008). Some of these different measures of connectivity are related through neural population sizes, for instance, average indegrees are obtained by dividing the total number of synapses by the size of the target population. Knowledge about the cellular architecture of the brain thus allows researchers to translate between different measures of connectivity. Furthermore, combining network connectivity with a quantification of the cellular architecture leads to a cellular-level network description, necessary for dynamical model simulations at this resolution.
In the present study, we investigate the network of vision-related areas of macaque cortex, a system that has garnered intense interest in experimental studies (e.g., De Valois et al.
1982; Luck et al.
1997; van Kerkoerle et al.
2014; Bastos et al.
2015a). The available experimental data on the cellular architecture and connectivity of the system are extensive, yet still incomplete. However, structural relations and distances between areas expose statistical regularities that we employ to bridge some of the missing data.
The microcircuit model of Potjans and Diesmann (
2014), which constitutes a synthesis of local connectivity data from electrophysiological and anatomical studies, forms the basis for the intra-area connectivity in our network. Although the data originate mainly from studies on rat somatosensory and cat primary visual cortex, the comprehensive collation of local connectivity by this model is unparalleled for macaque cortex, let alone for the individual areas we consider. Our choice is justified by predominant similarities between the local cortical connectivity in different species and areas, as formalized by the concept of a ‘canonical microcircuit’ (Douglas et al.
1989; Douglas and Martin
2004). We nevertheless take into account variability across areas as resulting from known differences in laminar compositions and their degree of connectivity.
The connectivity between areas in our model combines information from a recent release of the CoCoMac connectivity database (Stephan et al.
2001; Bakker et al.
2012) with quantitative data on cortico-cortical connection densities (Markov et al.
2014a) and laminar patterns (Markov et al.
2014b). For long-distance connections, tracing data are more reliable than diffusion MRI (Thomas et al.
2014), which enters into most current multi-area modeling work (Deco and Jirsa
2012; Sanz Leon et al.
2013; Kunze et al.
2016). The observed exponential fall-off of connection density with spatial distance (Ercsey-Ravasz et al.
2013) helps to estimate connection densities for area pairs where quantitative data are lacking. The categorization of areas into architectural types predicts cell densities and laminar thicknesses in case of missing data. Such structural differences between areas are in turn linked to and help fill in laminar patterns of cortico-cortical projections. A unique feature of our connectivity map is that it enables layer-specific polysynaptic pathways to be characterized, as synapse locations are statistically mapped (based on morphological reconstructions; Binzegger et al.
2004) to the locations of the target cell bodies forwarding the synaptic input. In this study, we aim to derive a consistent picture of the connectivity within and between vision-related areas within one hemisphere of macaque cortex. A treatment of callosal and subcortical connections therefore lies beyond the scope of the current study, but represents an important extension for a future revision of the model.
Besides uncovering layer-specific pathways for routing cortico-cortical communication, the resulting network description reveals a modular structure that resembles a functional categorization of areas. The derivations of the connectivity and the numbers of neurons necessarily entail choices which, given the available data, yield a compromise between detail and conciseness. Due to these simplifying assumptions and presently unexplained biological variability, the entries of the resulting connectivity matrix are only estimated up to a certain precision, and therefore the individual entries should be interpreted with care. The advantage of our approach is that it makes the assumptions explicit, which enables their consequences to be studied in a systematic manner. Furthermore, the matrix as a whole already provides a multi-scale connectivity substrate for the investigation of cortical dynamics via analytical theory and numerical simulation in a way that an incomplete matrix cannot, and various validations demonstrate the plausibility of its community structure and layer-specific pathways.
The remainder of this paper is organized as follows. In the “
Materials and methods” section, we provide an overview of the processing of the experimental data contributing to the model. In the “
Results” section, we detail the derivation of the network description including population sizes and the multi-scale cortical connectivity. Subsequently, we analyze the resulting connectivity map with regard to community structure and emerging paths in the network. In terms of source and target layers, we find that feedforward paths follow a stereotypical pattern, also shared by lateral paths, while feedback paths feature a high degree of heterogeneity. However, in pathways passing through several areas, the intermediate laminar patterns of lateral paths more closely resemble those of feedback paths. Finally, we discuss the implications of our results and suggest future directions in the "
Discussion" section. Preliminary results have been published as preprint in Schmidt et al. (
2016).
Discussion
The present study integrates data on cortical architecture, geometry, and connectivity into a comprehensive unihemispheric network description of the vision-related areas of macaque cortex. A number of simplifying assumptions and heuristics that are based on established and novel statistical regularities complement the measurements in view of the sparseness of quantitative species- and area-specific data. Our study thus represents a compromise between detail and conciseness, where avenues for future improvements are explicitly identified. The multi-scale network description consists of a population-, layer- and area-specific connectivity map together with neural population sizes, which resolve ambiguities in connectivity measures. In the derived connectivity, we find multiple clusters reflecting the anatomical and functional partition of visual cortex into early visual areas, ventral and dorsal streams, and frontal areas, showing that the network construction yields a meaningful structure. The laminar resolution of the model, along with a statistical mapping of synapse to target cell body locations, enables a novel characterization of direct and indirect paths across neural populations in the cortex. Our findings stand up to validation with varied network models defined based on moderately pruned connectivity data and models where the employed heuristics are relaxed.
The cortico-cortical connectivity is based on axonal tracing data collected in a new release of CoCoMac (Bakker et al.
2012) combined with recent quantitative and layer-specific retrograde tracing experiments (Markov et al.
2014b,
a). The projections revealed by these axonal tracing data are complex and not strictly sequential, including bypass connections such as those from V1 to V4 bypassing V2 (Nakamura et al.
1993). To translate FLN data into connection densities, we assume that the number of synapses established in the target area does not differ across projecting areas. Implicitly, other studies that interpret FLN in terms of connection strengths (e.g., Markov et al.
2013; Goulas et al.
2014) make the same assumption. There is, however, evidence that the number of cortico-cortical synapses per neuron in a projection depends on its direction (Rockland
2003).
We fill in missing data using relationships between laminar source and target patterns (Felleman and Van Essen
1991; Markov et al.
2014b) and architectural differentiation (Hilgetag et al.
2016; Beul et al.
2017), an approach for which Barbas (
1986) and Barbas and Rempel-Clower (
1997) laid the groundwork. To estimate missing data on connection densities, we use the exponential decay of FLN with inter-areal distance, which relies on the exponential distribution of axon lengths combined with the parcellation of cortical space into areas (Ercsey-Ravasz et al.
2013). For simplicity, we here assume an isotropic distribution of connection densities, in line with Ercsey-Ravasz et al. (
2013), but data from hamster cortex suggest that axons may extend further along the mediolateral axis than along the anterior–posterior axis (Cahalane et al.
2011).
The use of axonal tracing results avoids the pitfalls of tractography based on diffusion MRI data, which strongly depends on parameter choices (Thomas et al.
2014), has limited spatial resolution, cannot sense the direction of connections, and has been found to both underestimate (Calabrese et al.
2015b) and overestimate (Maier-Hein et al.
2016) cortical connectivity. A recent study comparing dMRI-based tractography on macaque cortex with retrograde tracing data shows that tractography after removal of false positives and false negatives is modestly informative about connection strengths (Donahue et al.
2016). Since axonal tracing data need to be combined across individuals whereas dMRI maps are obtained in individual brains, the two approaches are complementary.
The local connectivity of our network customizes that of the microcircuit model of Potjans and Diesmann (
2014) according to the specific architecture of each area, taking into account neuronal densities and laminar thicknesses. Although the model of Potjans and Diesmann (
2014) is based on data from rat and cat cortex, it serves as a prototype for the local circuits in our study due to the lack of similarly comprehensive quantitative data on pairwise connection probabilities in macaque cortex. Future revisions of the model can refine the analysis by incorporating additional knowledge on the local structure of macaque cortex as it becomes available, for instance information on cell morphologies in different areas (e.g., Gilman et al.
2017). Neuronal densities decrease from high to low-type visual areas, resulting in an apparent caudal-to-rostral gradient (Charvet et al.
2015). Combined with the assumption of a constant volume density of synapses (O’Kusky and Colonnier
1982; Cragg
1967) this yields higher indegrees in low-type areas. This trend matches an increase in dendritic spines per pyramidal neuron (Elston and Rosa
2000; Elston
2000; Elston et al.
2011). We thus clarify how volume densities of neurons and synapses together determine such an increase in per-neuron connectivity along the architectonic gradient of visual areas.
Total cortical thickness decreases with overall neuron density (cf., von Economo and Van Bogaert
1927; la Fougère et al.
2011; Cahalane et al.
2012). Similarly, total thicknesses from MR measurements decrease with increasing architectural type (Wagstyl et al.
2015), which has a strong positive correlation with cell density (Hilgetag et al.
2016). Laminar and total cortical thicknesses are determined from micrographs, which has the drawback that this approach covers only a small fraction of the surface of each cortical area. For absolute, but not relative, thicknesses, another caveat is potential shrinkage and obliqueness of sections. It has also been found that laminar and total thicknesses depend on the sulcal or gyral location of areas, which is not offset by a change in neuron densities (Hilgetag and Barbas
2006). However, regressing our relative thickness data against cortical depth of the areas registered to F99 revealed no significant trends of this type (Supplementary Fig. S2). Laminar thickness data are surprisingly incomplete, considering that this is a basic anatomical feature of cortex. Total thicknesses have already recently been measured across cortex (Calabrese et al.
2015a; Wagstyl et al.
2015), and could complement the data set used here covering 14 of the 32 areas. However, when computing numbers of neurons, using histological data may be preferable, because shrinkage effects on neuronal densities and laminar thicknesses partially cancel out.
We statistically assign cortico-cortical synapses to target neurons based on anatomical reconstructions (Binzegger et al.
2004). This assumes that the anatomical strength of a connection between two different types of neurons depends on the product of the average number of synapses formed by the source neuron in a particular layer and the dendritic density of the target neurons in that layer, an extended version of Peters’s rule (Braitenberg and Schüz
1991). Axo-dendritic overlap predicts connectivity to some extent, but the actual multiplicity and synaptic strength of connections between individual neurons show large variations (Shepherd et al.
2005; Kasthuri et al.
2015). However, Rees et al. (
2016) review existing literature and conclude that using Peters’s rule at the level of cell types instead of individual cells can deliver a reasonable approximation to cortical circuitry. On the target side, the assignment of synapses based on dendritic extent yields laminar cell body distributions for feedforward and feedback projections that mostly follow the classical scheme for laminar synapse distributions of Felleman and Van Essen (
1991). However, in our network, layer 4 neurons receive substantial feedback input, stressing the importance of distinguishing between synapse and cell body positions, as previously pointed out by De Pasquale and Sherman (
2011). This prediction can be tested for example with glutamate uncaging in the source area combined with patch-clamp recording in the target area (Covic and Sherman
2011), or via axonal tracing combined with morphological reconstruction of the target neurons (Porter
1997). Covic and Sherman (
2011) found feedback onto layer 4 neurons in mouse auditory cortex; however, such a pattern remains to be shown in primates. This finding would shed a new perspective on the role of L4 neurons in cortical processing. In predictive coding for instance, L4 neurons are hypothesized to process forward prediction errors using their feedforward inputs, while layer 5 pyramidal cells process feedback predictions via their apical dendrites in the supragranular layers (Bastos et al.
2012). With L4 neurons receiving additional feedback via dendrites reaching into layer 2/3, their role could be more intricate and involve processing of both feedforward and feedback signals.
Our analysis includes target patterns from the CoCoMac database, which enables us to link target patterns to quantitatively defined laminar projection patterns of bilaminar origin, refining the classification of Felleman and Van Essen (
1991). Markov et al. (
2014b) combined their source patterns from retrograde tracing with target patterns from previous anterograde tracing studies in different species and distinguished feedback and feedforward connections further into hierarchically short-range and long-range projections, respectively. They found subtle differences in target patterns, e.g., that feedforward connections from high-type visual areas terminate in layers 3B and 4 of intermediate areas, but exclusively in layer 4 in low-type areas. However, the anterograde data used by Markov et al. (
2014b) cover target patterns for connections in only a small subset of visual areas. Our data from CoCoMac include target patterns for all visual areas with 29% coverage of all connections in our network, but do not allow us to draw conclusions on such a fine classification into hierarchically short-range and long-range connections. Future work could test if a revised version of the full CoCoMac dataset using a finer layer distinction supports the findings of Markov et al. (
2014b). Laminar specificity of cortico-cortical connections is important because it can support complementary channels for feedforward and feedback communication in cortex (Bastos et al.
2015b). In particular, anatomical segregation of communication channels likely plays a role in enabling directional differences in oscillation frequencies associated with inter-area communication (van Kerkoerle et al.
2014; Bastos et al.
2015a; Michalareas et al.
2016). This segregation can occur even in single cells that combine feedback and feedforward processing on their apical and basal dendrites (Körding and König
2001; Urbanczik and Senn
2014), again stressing the importance of taking cell morphologies into account.
The connectivity of neuronal networks can be described in terms of different measures, each highlighting a specific aspect of the network and relating differently to its dynamics. For instance, in mean-field descriptions of network dynamics, indegrees tend to be most directly related to stationary firing rates, while fluctuations around this stationary state depend on the population size, and therefore, correlations are determined by a combination of indegrees and connection probabilities (Brunel
2000; Helias et al.
2013). On the other hand, outdegrees relate more directly to the overall influence of each node. Our network description consisting of population sizes and numbers of synapses for each connection allows us to translate between these measures, showing how they differ in their relative strength across connections. Using the appropriate connectivity measures can facilitate the interpretation of observed dynamics.
The population-level connectivity enables us to identify the most prominent laminar projection patterns in shortest paths between areas. While pathways from high-type to low-type areas and horizontal pathways (between structurally similar areas) both follow a stereotypical pattern originating in the supragranular layers and targeting layer 4, projections from low-type to high-type areas feature a richer repertoire of layer-specific paths. At relay stages in indirect paths, horizontal pathways more closely resemble low-to-high-type pathways. These findings suggest that areas of equal architectural type communicate via similar pathways as connections from structurally more differentiated to less differentiated areas in terms of their start-end pattern, but that these pathways are often relayed via pathways similar to those from structurally less differentiated to more differentiated areas. The hypothesis that dynamical interactions follow these anatomical paths could be tested in experiments as well as numerical simulations. The anatomical paths in our model are fairly independent of whether they are categorized based on SLN or the architectural types. An exception is that a significant number of low-to-high-type paths originate in supragranular layers, while the origin of feedback paths is strongly dominated by the infragranular layers. Still, these similarities suggest that functional signatures of connections categorized based on the structural gradient are similar to those observed for hierarchical projections (van Kerkoerle et al.
2014; Bastos et al.
2015a; Michalareas et al.
2016).
We here concentrate on aspects of cortical structure for which substantial datasets are available, leaving aside insights on specific details in individual areas for which the available information is highly incomplete. Our algorithmic approach makes the network amenable to the integration of additional details, such as more diverse neuronal populations (Defelipe et al.
1999; Binzegger et al.
2004; Markram et al.
2015), additional area specificity of local circuits (Beul and Hilgetag
2015), connectivity patterns beyond pairwise connection probabilities (Song et al.
2005; Kasthuri et al.
2015; Markram et al.
2015), or spatial properties of connectivity (Colby et al.
1988; Salin et al.
1989; Gattass et al.
1997; Markov et al.
2014b). The cortico-cortical connectivity may be further refined by incorporating a dual counterstream organization of feedforward and feedback connections (Markov et al.
2014b), by including different numbers of cortico-cortical synapses per neuron in feedforward and feedback directions (Rockland
2003), and by incorporating cortico-cortical projection patterns on the single-cell level as found in mouse V1 (Han et al.
2017). It is also worth investigating whether the preferential targeting of excitatory neurons by feedback projections is part of a more gradual reduction in inhibition–excitation ratio from feedforward to feedback projections, as is the case for optogenetically determined EPSCs (D’Souza et al.
2016).
In this study, we concentrate on the network of vision-related areas within one hemisphere of cortex, thereby leaving aside callosal and subcortical connections as well as connections with other cortical areas. Since most tracing studies concentrate on one hemisphere, knowledge about callosal connections is sparse; however, tracing data from mouse cortex (Goulas et al.
2017) and rhesus monkey prefrontal cortex (Barbas et al.
2005) suggest similar construction principles to those of ipsilateral connections, which can be used to inform a future revision of the model. The integration of thalamo-cortical loops is an important extension of the model, but in view of the added complexity beyond the scope of the current study. Since the corresponding connectivity has been measured for parts of cortex only, it would be necessary to fill gaps in the data by empirical regularities similar to those used in the present study, possibly employing more advanced graph-theoretical techniques similar to Jouve et al. (
1998). This would help ensure the realism of graph-theoretical properties of the connectivity matrix not tested for in the present study, and would enhance the reliability of individual entries of the matrix that are currently only first-order estimates.
Our study can thus be the starting point for iterative refinement and more detailed descriptions of cortical connectivity, contributing to a better understanding of cortical structure. It also provides the basis for numerical simulations that investigate the relation between structure and dynamics (Schmidt et al.
2016; Schuecker et al.
2017). In contrast to previous simulation studies, which are based on binary or coarsely weighted tracing data or on diffusion MRI (Honey et al.
2007; Knock et al.
2009; Deco et al.
2009), the weighted and directed graph resulting from our integration of axonal tracing data enables studying the activity supported by the highly heterogeneous connectivity of cortex.