The human genome is composed of approximately 3 × 109
base pairs, containing around 20,000–30,000 genes [15
]. The compilation of the genome was aided by the fact that base pairs and genes are relatively straightforward choices as basic structural elements. The connectome will consist of two main descriptors defining its network architecture: neural elements and neural connections. Data fields for these elements specify superordinate or subordinate structures, a normalized position within a standard coordinate system, and additional parameters such as physiological/biophysical metadata that are crucial for specifying neural dynamics. The set of all N
elements constitutes the columns (targets) and rows (sources) of an N2
connection matrix A, whose aij
entries represent connections between individual elements i
. In keeping with conventions adopted by other authors [16
], confirmed absence of a connection is denoted by aij
= 0, while confirmed presence of a connection (irrespective of its strength or physiological characteristics) results in aij
= 1. Once a connection is confirmed to be present, its nonzero matrix element receives additional data entries cataloguing a range of structural and physiological parameters, such as fiber direction, connection density, strength, sign (excitatory/inhibitory), conduction delay, potential modulatory effects (voltage dependence), etc. The union of the binary connection matrix and connection-specific physiological data then results in a structural description that combines connection topology and biophysics.
Compiling the connectome faces two significant challenges not shared by other natural or technological networks. First, the human brain is a highly complex organ with a great number of structurally distinct, heterogeneous, yet interconnected components. Because a primary application of the connectome will likely be a structural substrate for understanding human cognitive function and interpreting neuroimaging studies, a first draft of the connectome may focus on the structural description of the corticothalamic system, including all regions of the cortex and their associated thalamic nuclei. Extensions of this first draft might include additional connected regions (striatum, cerebellum, etc.), with the ultimate aim of compiling the connectome of the whole brain.
A second challenge is that basic structural elements of the human brain, in terms of network nodes and connections, are difficult to define. Different kinds of structural descriptions could target at least three rather distinct levels of organization. At opposite ends of the scale are the level of single neurons and synapses (microscale) and the level of anatomically distinct brain regions and inter-regional pathways (macroscale). Between these two levels is the level of neuronal groups or populations (mesoscale). It is important to determine which level of description is the most appropriate for a first draft of the human connectome.
Microscale: Single Neurons and Synapses
Attempting to assemble the human connectome at the level of single neurons is unrealistic and will remain infeasible at least in the near future. With single neurons as the basic element, the size of the connectome would be several orders of magnitude larger than that of the genome, comprising an estimated 1011
neurons, with 1015
connections between them (approximately 1010
neurons and 1013
connections in the cortex alone) [18
]. Recording or tracing 1015
connections is not only technically impossible, it may also be unnecessary. While a genomic mutation in a single base pair can have dramatic consequences, alterations of single synapses or cells have not been shown to have similar macroscopic effects. Instead, there is overwhelming evidence that human cognitive functions depend on the activity and coactivity of large populations of neurons in distributed networks, including the corticothalamic system [20
]. Furthermore, individual neurons and connections are subject to rapid plastic changes. These changes include synaptic weights as well as structural remodeling of dendritic spines and presynaptic boutons [21
], possibly switching synaptic connections between large numbers of potential synaptic sites [22
]. We suggest that the vast number, high variability, and fast dynamics of individual neurons and synapses render them inappropriate as basic elements for an initial draft of the connectome.
Macroscale: Brain Regions and Pathways
An advantage of single neurons is that the elements themselves are relatively easily demarcated and well defined. In contrast, brain areas and neuronal populations are more difficult to delineate. No single universally accepted parcellation scheme currently exists for human brain regions (e.g., areas of the cerebral cortex), posing a significant obstacle to creating a unified resource such as the connectome. In the human cerebral cortex, neurons are arranged in an unknown number of anatomically distinct regions and areas, perhaps on the order of 100 [24
] or more. Different subdivisions of the human brain (e.g., brain stem, thalamus, cerebellum, or cortex) may require different criteria for parcellation.
Nonetheless, anatomically distinct brain regions and inter-regional pathways represent perhaps the most feasible organizational level for compiling a first draft of the human connectome. Several neuroinformatics resources recording large-scale connection patterns in the cerebral cortex of various mammalian species already exist, for example, for most cortical regions of the macaque monkey [16
], cat [26
], and rat [27
]. Computational analyses of these datasets have revealed a broad range of network characteristics [28
], including the existence of clusters of brain regions [29
], hierarchical organization [30
], small-world attributes [32
], distinct functional streams [34
], motifs [35
], and areal contributions to global network measures [36
A broad range of experimental approaches exist at the macroscale. Cerebral white matter has traditionally been taken as a marker of the amount of connectivity within a cortical system. The relative contribution of cerebral white matter has increased throughout phylogeny to such an extent that its volume and metabolic requirements may present a limitation to further increases in connectional complexity [37
]. The structural organization of white matter has been investigated by dissection, histological staining, degeneration methods, and axonal tracing [38
]. Axonal tracing methods are the main basis for existing mammalian large-scale connection matrices [16
], and their systematic compilation is currently being refined [17
] to extract more sophisticated data, perhaps with the help of automated text analysis. In the human brain, postmortem tracing with carbocyanine dyes has provided details of connectivity within and between adjacent areas [40
]. However, because of the slow transport and the length of fibers in the human brain, this method fails to reveal more remote connections. Another tracing method employs the in vivo detection by magnetic resonance imaging of high-contrast rare earth ions (e.g., manganese [42
]) that have been injected into fiber tracts or inhaled, and taken up by neurons. The invasive nature and potential toxicity of the procedure makes it an unlikely candidate for human connectivity analyses.
Recent noninvasive imaging methods (diffusion tensor imaging [DTI] in its several variants, commonly followed by computational tractography) have been shown to produce results that are consistent with known pathways formed by major fiber tracts in the human brain [43
], although there continues to be some limitations in data acquisition and processing algorithms [46
]. To disambiguate signals produced by crossing or intersecting fibers, advances in diffusion imaging technology may allow the resolution of multiple axonal pathways within a single image voxel [47
]. Despite the promise of diffusion imaging, a systematic atlas of DTI-based neuroanatomy, including probabilistic data gathered from individual brains, has not yet been produced, and the relationship among tensor fields, fiber tracts, and neuronal connections remains controversial. This controversy is likely to be resolved only through a combination of anatomical tracing techniques with noninvasive diffusion and functional imaging.
Perhaps the most promising avenue for compiling the human connectome originates from the notion that individual brain regions maintain individual connection profiles. What defines a segregated brain region is that all its structural elements share highly similar long-range connectivity patterns, and that these patterns are dissimilar between brain regions. These connectivity patterns determine the region's functional properties [49
], and also allow their anatomical delineation and mapping. Diffusion imaging can be used to identify borders between cortical areas [50
], most clearly on the basis of differences in long-range connections to the thalamus [51
]. The idea that patterns of connectivity can be used to identify areal boundaries has also been tested in a combination of functional imaging and DTI/tractography in the human medial frontal cortex [53
]. First, connection probabilities from voxels within the medial frontal cortex to all other voxels in the rest of the brain were obtained. Binarized connection patterns were then used to calculate a cross-correlation matrix, which was examined for the existence of distinct clusters of voxels with shared connection patterns. Such clusters were taken to represent anatomically segregated areas corresponding to human supplementary motor area and pre- supplementary motor area, respectively. Functional mapping revealed close correspondence between DTI and functional activation patterns. While this combined structural–functional approach is computationally intensive, nothing prevents its application to the entire corticothalamic system. We suggest that the correlated use of noninvasive structural and functional imaging methods offers the most promising experimental route toward the human connectome.
Structural connection data obtained using this combined methodology can in principle be validated by histological or tracing methods. Likely, no single method will turn out to be sufficiently powerful or comprehensive. A systematic application of sophisticated diffusion-weighted imaging combined with spatially registered high-resolution anatomical or spectroscopic imaging, regional activation, and coactivation data (e.g., electroencephalography, magnetoencephalography) obtained within the same individual subject, may offer the most feasible strategy for mapping the human connectome at the macroscale. This first draft of the human connectome would take the form of a probabilistic map of voxel-by-voxel connectivity embedded within standard coordinates containing approximately 104–105 elements and approximately 105–107 structural connections. It would not only provide information on the large-scale connection patterns within the corticothalamic system, but also on parcellation of human cortex into distinct areas based on a combination of structural and functional data in the same individual. Since this connectivity matrix is voxel-based, it can be cross-referenced with existing reference templates and with population-based brain atlases.
Mesoscale: Minicolumns and Their Connection Patterns
A first draft of the corticothalamic connectome at the macroscale might provide a comprehensive dataset comprising several hundred brain regions and thousands of pathways, but it does not incorporate information on functional subdivisions or segregated subcircuits within each brain region. While such a macroscale description will provide many fundamental insights into the large-scale organization of human cortex, it is an insufficient basis for a complete understanding of the human brain's functional dynamics and information processing capacities. A further step in this direction involves the characterization of connection patterns among elementary processing units, corresponding to local populations of neurons such as cortical minicolumns. Mountcastle [54
] originally proposed the cortical minicolumn as a basic functional unit of mammalian cerebral cortex. While details of minicolumn architecture are likely to vary across different brain regions [56
], minicolumns generally contain approximately 80–100 neurons, spanning all cortical layers, with a diameter of approximately 30–50 μm [55
]. Minicolumns may possess relatively stereotypic internal processing, and maintain generic patterns of inputs and outputs with minicolumns in other regions [56
Recent studies have provided evidence for functionally specialized and precisely wired subnetworks of neurons coexisting within single cortical columns [60
]. These studies have shown that cortical columns may contain segregated subnetworks, corresponding to minicolumns, which promote greater intracolumnar functional independence and informational heterogeneity. The members of such subnetworks are selectively interconnected with each other, indicating that connections within and between minicolumns follow more complex rules than simple random patterns or Gaussian (distance-dependent) connection profiles.
Minicolumns may be a sensible choice for neural elements at the mesoscale of the connectome because they may represent basic functional elements that are crucial for cortical information processing. While tracing or recording all minicolumns in an individual brain is still impossible, it may be feasible to collect data on minicolumn anatomy for selected brain regions, which can then be “fit” into the appropriate positions within the macroscale connection matrix. Functional imaging at columnar resolution has been carried out in animal experiments using high field strength [62
], and may be facilitated in the future through selective imaging of fast and spatially precise capillary cerebral blood flow response components [63
]. Especially important for determining functional responses of brain regions are connection patterns between each region's constituent elements. Axonal tracing methods have revealed specific patterns of horizontal connections between individual cells and cell populations within a brain region, which are often found to preferentially link cells with similar response characteristics [64
], resulting in clustered intra-areal connectivity. Such patterns might also be accessible in the human brain [40
], and regional variations in such patterns may provide important clues regarding the way in which information is processed within each region.
The axonal tracing approach delivers minicolumn maps for each distinct brain region, including information on their functional segregation and local (intra-areal) interconnectivity. While parcellated brain regions at the macroscale can be identified across individuals, we have no means to resolve the locations of corresponding minicolumns across different brains. The mapping of smaller-than-macroscopic units, therefore, requires coordinate-independent mapping approaches [66
], which preserve the anatomical relationships and basic physiological properties of these units. Thus, macroscale and minicolumn descriptions deliver complementary datasets that need cross-level integration to achieve a single mesoscale version of the connectome. Effectively, minicolumn maps need to be mapped onto brain region–specific voxel sets rather than individual voxels, with voxel sets providing spatial embedding and probabilistic long-range connections, and minicolumn maps providing local connectivity, processing, and coding information. The minicolumn description provides a functionally heterogeneous architecture that is unique to each parcellated brain region, with specific (probabilistic) patterns of intra-areal and interareal minicolumn connections. A crucial task will be to convert long-range, nondirected voxel-by-voxel connectivity into directed, functionally heterogeneous long-range minicolumn interconnectivity. This requires intermediate descriptive levels such as stripes, bands, and blobs in the early visual system, for which specific connection patterns have been demonstrated across areal boundaries. When accomplished, this cross-level integration will result in a mesoscale connection matrix of the human brain that might comprise as many as 107
structural elements (comparable in size to the 2005 World Wide Web), with minicolumn elements directly cross-referenced to voxels in the macroscale map.