1.1.

Problems Related to Variability in Tract-Tracing Methods and Data Analysis

The tract-tracing techniques used for connectivity studies have improved considerably over time, from early work involving stains for degenerating axons following targeted lesions and transport of radioactive amino acids to more modern research using injected plant lectins, synthetic chemicals, and viral vectors.^2,7 When collating results from prior studies to establish the electronic databases, the curators of these databases must select which techniques and studies they consider most reliable,³ as different neuronal tracers differ with respect to sensitivity, resolution, the degree to which they are preferentially transported either anterogradely or retrogradely, and the possibility of labeling fibers of passage. Different curators might make different judgments in this regard, a problem common to all meta-analyses.

Even among the studies that are selected for inclusion in a connectivity database, the results often differ considerably with respect to the presence or strength of a connection between the same two structures, which may reflect differences both in the tracer employed and in the way it was detected. In addition to differences in microscopes and detection systems (e.g., direct fluorescence versus immunostaining with or without secondary antibodies and other amplification steps), the apparent presence or strength of a connection can be influenced by subtleties in the location of the injection, the volume of the injection, the mechanism of delivery (pressure injection or iontophoresis), diffusion of the tracer molecule, the concentration of tracer, possible selectivity of uptake of the tracer by subclasses of neurons,⁹ and the duration of survival in relation to both rate of transport and turnover of the tracer.

Discrepancies in reports of labeling intensity also may be exacerbated by the semiquantitative scales that are common in anatomical studies, which often score labeling on a four-point scale (absent, +, ++, and +++). The standards for assigning such scores may not be the same between studies. For example, the vital difference between “no labeling” and “labeling exists” is a function of the sensitivity of the labeling method, which can be strongly influenced by minor particulars of the staining procedures, the sensitivity and resolution of the microscope and camera system, the thoroughness of the sampling strategy within each target region, and the false detection rate, which is the tendency to report labeling in control tissue (e.g., tissue from noninjected animals). It is also unclear whether different researchers implement these semiquantitative scales linearly or nonlinearly, and, if nonlinearly, what curve fit is applied. Despite these sources of variability, a single semiquantitative measure of staining intensity tends to be used for the data analysis associated with the electronic databases.⁸

1.2.

Problem of Brain Parcellation

Brain function reflects communication between individual neurons. A connectome detailing the neuron-to-neuron connectivity across the entire simple nervous system of the nematode C. elegans has been completed through serial reconstruction of axons across adjacent electron micrographs.¹⁰ Although neuron-to-neuron connectomes of portions of the Drosophila nervous system also have been completed,^11–13 a comparable neuron-to-neuron connectome is not currently envisioned for the rodent or human brain, given that there are orders of magnitude more synapses involved and considerable variation between individuals.^1,2 Instead, these connectomes are being constructed on the basis of connections between brain “areas,” a simplification that can be partially justified by observations that neighboring neurons tend to have similar properties. However, using brain regions as a basis for the connectome introduces several problems.

The problem with building a connectome using brain areas as units is that there is no a priori consensus about what these units should be.¹⁴ The brain can be parcellated using different schemes that can be based on differences in (1) the packing density of cell bodies, known as cytoarchitectonics, (2) the myelination density and patterns in gray matter, known as myeloarchitectonics, (3) the distribution of different proteins, mRNA, or neurotransmitters, (4) the activity of metabolic enzymes, such as cytochrome oxidase and succinate dehydrogenase, (5) brain activity under various conditions of peripheral stimulation and consciousness, recorded using various methods, and (6) patterns of connectivity to other brain regions (we will revisit this last potential criterion below). Although there can be some correspondence between the boundaries imposed on brain anatomy using these different criteria, mismatches also are common. In the mouse somatosensory cortex, for example, there is a correspondence between cytoarchitectonically defined whisker barrels that are detected through Nissl staining and patches of elevated cytochrome oxidase activity, whereas in rats, whisker barrels are harder to distinguish using cell body stains, but are evident in cytochrome oxidase staining patterns.¹⁵ Furthermore, even when considering the same cytoarchitectonic basis for parcellation of cerebral cortex, different researchers have chosen different boundaries.^2,16–18

The BAMS and neuroVIISAS databases provide tools to compare several of the parcellation schemes with one another, and curators of the sites attempt to resolve ambiguities introduced by the fact that the original studies may have reported results in reference to different parcellation schemes, so that all results can at least be presented in the same format.^3,4,8 Despite this type of standardization, none of the parcellation schemes have been independently validated with respect to the suitability of the brain regions as foundations for a connectome.²

1.3.

Examples of Important Forms of Cortico-Cortical Connectivity that Can Be Obscured in Post Hoc, Parcellation-Based Connectomes

Parcellation schemes typically divide the brain into regions that differ in size, and within the cortex, some of the areas considered as units in the schemes used in the electronic databases are both quite large and can be subdivided into smaller regions using other parcellation schemes. Projections from one large cortical region to another may involve neurons that are located in a small subdivision of the first region and that send axons to a small subdivision of the other region. For example, neurons in different whisker barrels of primary somatosensory cortex project topographically to different subdomains of primary motor cortex¹⁹ and secondary somatosensory cortex.²⁰ Knowledge of this organization is lost in some electronic databases, because the entire barrel field of somatosensory cortex is considered as a single unit, as is the entire primary motor cortex and the entire secondary somatosensory cortex. It has been argued that topographic maps provide an efficient neural solution to many problems in sensory processing,²¹ therefore, it seems important that these relationships be evident in the connectome.

Projections that involve only a part of a predefined brain region also raise an important problem for quantification. It is unclear how (or whether) intense staining within a small portion of a brain area would be distinguished quantitatively from moderate staining distributed throughout the entire area when establishing quantitative descriptions of network relationships. Particular brain regions also can contain intermingled neurons with very distinct projection patterns. For example, in somatosensory cortex, the whisker barrel field contains distinct barrel column neurons in layer 4 whose projections are almost entirely local, as well as septal neurons that send diffuse collaterals that extend horizontally for great distances in all directions.²² Combining different projection patterns in an area-to-area model of connectivity could cause confusion.

In general, digitizing patterns of connectivity into discrete brain regions for display in correlation matrices and network graphs strips the neural connectivity data of its inherent spatial structure, obscuring strong rules such as the greater connectivity between neighboring regions that have been noted in descriptions of the human connectome.²³ As will be described in greater detail below, the distribution of labeling following injections of a tracer can result in a gradient of decreasing intensity that crosses continuously across a boundary between supposedly distinct cytoarchitectonic regions,^24,25 another potentially important spatial feature that is lost after binning the data into discrete regions. In the section below, we will discuss in detail the diffuse, radiating and tapering horizontal projection that is the subject of our own research and that is effectively obscured in the post hoc connectome databases. (Here and throughout the remainder of this paper, we use the term “diffuse” to mean simply “spread out over a large space, not concentrated in one area” and do not imply any particular function of the connectivity.) We then will consider to what extent knowledge of this horizontal connectivity is preserved in the more recent connectome projects.

4.1.

Mouse Connectome Project

For the Mouse Connectome Project, two iontophoretic injections were made into different cortical locations of each brain, each injection containing a different pair of one anterograde (Phaseolus vulgaris leucoagglutinin or fluoro-ruby) and one retrograde (fluorescently tagged cholera toxin b subunit or Fluorogold) tract tracer.⁶ After one week of survival for transport of the tracers, the brains were sectioned coronally at a thickness of $50\mu \mathrm{m}$, and every fourth slice was counterstained with a fluorescent Nissl stain and immunostained using a fluorescent secondary antibody to label the leucoagglutinin tracer. The coronal plane of section allows evaluation of projections to and from subcortical brain regions in the same slices. All five fluorescent labels were imaged in separate channels of a high-throughput microscope.

Because only every fourth $50\text{-}\mu \mathrm{m}$ slice is analyzed, $150\mu \mathrm{m}$ of data along the anterior-posterior axis is missing between each analyzed slice, which prevents true tractography and the reconstruction of most horizontal axons. Also, in coronal slices, most horizontal axons are cut obliquely, leaving small fragments that may be overlooked in the manual annotation of the staining employed by this group. Nevertheless, some of the images of sections available online in the Mouse Connectome Project iConnectome⁶ show evidence of lightly stained small fragments in surrounding cortical regions after injections of anterograde tracers in the barrel field [Fig. 4(a)]. These fragments seem to be distributed in a pattern consistent with the long-range, diffuse horizontal axonal radiation that we have detected in rat cortex.

Fig. 4

Possible evidence for long-range horizontal axons in the Mouse Connectome Project. (a) Images of BDA staining in a coronal slice taken through the core of an injection site in mouse barrel cortex (left panels) and in a slice 1.4 mm posterior, where long-range horizontal axons in cortex might be expected based on our findings. The right panel shows a higher magnification of light staining that may represent labeled horizontal axons sectioned obliquely in the coronal slice. Images are courtesy of Ref. 43 and come from brain SW120819-02A (slices 17/43 and 24/43). (b) Images of the distribution of labeling following injections of anterograde tracers into the mouse barrel field in the Mouse Connectome Project Map Viewer. Areas deemed to show significant labeling have been shaded, with the cores of the injection sites labeled using darker circles (top panels for each brain). For each brain, a coronal slice 0.7 to 0.8 mm posterior to the injection site is also shown (bottom panels – positions relative to Bregma are indicated). Although there is some sense of a horizontal spread of axon density toward these posterior slices, the axons do not extend as far as would be expected from our data, especially within the coronal slices.

For the Mouse Connectome Project, the density of projections were annotated manually in reference to brain regions determined from an a priori parcellation scheme using a four-step semiquantitative scale (0, 1, 2, or 3), with more than one evaluator scoring each case.⁶ The criterion for receiving a score of 1 (sparse) included verification of each anterograde projection by retrograde labeling (at least a few labeled neurons). This confirmation is potentially important, given that most tract tracers label neurons both anterogradely and retrogradely to some extent, including rAAV.^44–47 However, with this stringent criterion, it is difficult to determine whether the infrequently encountered short fragments of obliquely cut horizontal axons (which likely correspond to our horizontal radiation) would receive a score of 0 or 1 in this annotation. The use of a priori parcellation schemes also could obscure the horizontal projection, as was discussed above in the section on connectomes constructed from prior data.

In addition to scoring brain areas using a four-point scale, the Mouse Connectome Project also includes manually shaded areas receiving projections on a standard reference atlas. From the four examples of this shading currently available on the website for barrel field injections [Fig. 4(b)], there is some sense of a horizontal spread of fibers out from the injection site along the anterior-posterior axis, but the axons do not extend as far as the horizontal projection in our own studies, and the threshold criteria for shading are not clear.

4.2.

Mouse Brain Architecture Project

In a parallel effort, the Mouse Brain Architecture Project,⁴⁸ provides a database of images of coronal brain slices following tracer injections placed throughout the mouse brain. Tracers include BDA (anterograde) and cholera toxin B (retrograde), as well as viral vectors. The viral vectors include rAAV vectors with permissive CMV early enhancer/chicken β-actin (CAG) promoters to drive high levels of expression of green or red fluorescent proteins for anterograde tracing, as well as two different modified rabies virus vectors, one for retrograde and one for anterograde tracing. For cortical rAAV injections, both green and red fluorescent protein vectors are used at each site to infect neurons at different depths; otherwise, one tracer is used at each injection site. The brains are sliced at a thickness of $20\mu \mathrm{m}$, and alternate sections are used for tracer localization and Nissl staining. The missing tissue disrupts tractography, but the slices are closer together than those in the Mouse Connectome Project. Very similar to the case for the Mouse Connectome Project, images from brains injected with anterograde tracers into the somatosensory cortex show evidence of labeled axon fragments throughout the surrounding cortical regions, including other primary sensory cortices and associational regions (see for example, visual cortical regions in section 146 from brain 1612F and section 153 from brain 1088F), which is fully consistent with the horizontal fiber radiation we have described for rat brains.

For both the Mouse Connectome Project and the Mouse Brain Architecture project, slices were processed by standard procedures on microscope slides. As a result, there are occasional section irregularities (e.g., fluorescence associated with small tears in the tissue) that might interfere with the accurate detection of sparse fragments from obliquely cut axons. The use of two-photon tomography by the Allen Institute that is described below may eliminate the noise contributed by such section irregularities and, thus, allow more sensitive detection of the sparse fragments.

4.3.

Allen Mouse Brain Connectivity Atlas

The Allen Institute for Brain Science has employed small iontophoretic injections of rAAV(2/1) expressing GFP under a synaptophysin promoter, which combines with the tropism of the serotype 1 capsids to further restrict expression of GFP to neurons.⁵ After three weeks of expression and transport, the mouse brains are sectioned coronally at 100-micron intervals, and after each slice is removed, the surface of the tissue block is interrogated with scanning two-photon fluorescence tomography to collect a single focal plane of fluorescence data indicating the distribution of endogenous GFP fluorescence at a high level of sensitivity.⁵ The technique of tomography ensures that data in different sections can be aligned precisely without distortion. The method can be automated and generates $\sim 750\mathrm{GB}$ of image data from a single brain in $\sim 18.5\mathrm{h}$. Coronal sections are ideal for the overall goal of creating the Allen Mouse Brain Connectivity Atlas, which involves tracer injections and connectivity mapping of subcortical brain regions as well as cortical ones, because the coronal plane is most commonly used in other atlases, and its familiarity allows a more intuitive appreciation of the brain regions involved.

The large image files in the Allen Mouse Brain Connectivity Atlas reflect the high spatial resolution of the coronal images, but despite the resolution in this plane, $100\mu \mathrm{m}$ of depth information along the anterior-posterior dimension is missing between the collected coronal planes. The missing data prevent true tractography and the reconstruction of most of the horizontally oriented axons that are the subject of our research. Nevertheless, the individual histological sections available online show clear evidence of obliquely cut, small axonal fragments in the cortical regions surrounding injection sites in the barrel field of mouse cortex, and the distribution of these axonal fragments seems to be consistent with the horizontal axonal radiation that is more directly evident in our sections of flattened rat cortices [Fig. 5(a)]. In addition, the Allen group interpolates the data between coronal sections to create a three-dimensional (3-D) model of the overall projection patterns. The horizontal axonal radiation may be preserved in this rendering [Fig. 5(b)], although it is difficult to see this diffuse projection pattern when looking down on the modeled cortical surface due to underlying and overlying projections.

Fig. 5

Possible evidence for long-range horizontal axons in the Allen Mouse Brain Connectivity Atlas. (a) Images of green fluorescent protein fluorescence in three coronal slices following injection of a recombinant AAV vector into mouse barrel cortex. Details from the two-photon tomography images show fluorescent structures consistent with long-range horizontal cortical axons sliced obliquely at substantial distances posterior from the injection site (arrowheads). Images of this brain (378-1488, Experiment 127866392-SSp-bfd) were downloaded from Ref. 49. (b) Summaries of the projection densities from Experiment 127866392-SSp-bfd downloaded from Ref. 50 (top row) and from Brain Explorer 2 software (downloaded from Ref. 51) (bottom row). The radiation of long-range horizontal axons through all cortical layers probably contributes to the axonal density surrounding the injection site in these images; however, it is more difficult to parse the horizontal projection from specific projections in these images than it is in our slices of flattened cortices.

As part of the analysis of their connectivity data, the Allen group counts pixels of segmented GFP labeling within $100\mu \mathrm{m}\times 100\mu \mathrm{m}$ areas of their coronal sections to generate a quantitative measure of the projection density, which we think offers important advantages to semiquantitative scales. These areas are extended along the A-P axis to create voxels that are then fit to a standard 3-D reference atlas. However, most of the network analysis using correlation matrices and graph theory does not involve these equally sized volumes, but rather involves parcellated brain regions similar to what was described above for connectomes constructed from preexisting, published data.⁵ The use of a priori parcellation schemes in these analyses, therefore, may obscure certain details of connectivity that are available in the original data, including topographical projections and diffuse, long-range horizontal projections that decline gradually in density. The Allen group provides software that can summarize projection patterns and strengths within 3-D models of the mouse brain, which reestablishes some of the spatial relationships that are not evident from network graphs and correlation matrices.