The nascent field of connectomics applies rapidly developing ultrastructural, trans-synaptic tracing, and whole brain imaging technologies to identify neural circuits at micro-, meso-, and macroscopic resolutions. The central tenet of the connectomic approach is that insights regarding both the functional properties of specific neural circuits and general brain organizational principles may be gained by definitively resolving network architecture (Carandini, 2012; Denk et al., 2012; Mitra, 2014).

Ultrastructural approaches such as serial block-face electron microscopy can comprehensively identify local synaptic connectivity, but are limited by the enormous time and costs associated with data acquisition and analysis, and are therefore best suited to the examination of small regions of brain in high detail (discussed by Lichtman and Denk, 2011; Wanner et al., 2015). Investigators interested in mapping more diffuse circuits have instead monitored the trans-synaptic spread of replication-competent neurotropic viruses such as rabies (Ugolini, 1995; Kelly and Strick, 2000; Dum et al., 2016) and alpha herpes variants (Strack et al., 1989; Rinaman and Schwartz, 2004; McGovern et al., 2012; reviewed by Nassi et al., 2015; Wojaczynski et al., 2015).

In recent years glycoprotein-deleted EnvA-pseudotyped rabies [SADΔG(EnvA)] has emerged as a flagship tool for tracing connectomes in experimental animals (Wickersham et al., 2007b; Callaway and Luo, 2015). The key insights made by Callaway and colleagues in developing this system are that the ability of SADΔG(EnvA) to enter populations of target neurons and retrogradely spread to monosynaptically connected partners can be controlled by selective expression of the EnvA receptor, TVA, and the rabies glycoprotein respectively. Here we apply this approach to resolve the afferent connectome of putative sympathetic premotor neurons in the rostral ventrolateral medulla (RVLM) of the Sprague Dawley rat.

This population, approximately half of which are adrenergic C1 neurons (Stornetta, 2009), is a major source of the glutamatergic drive that maintains sympathetic vasomotor tone and therefore determines arterial blood pressure (Guyenet, 2006). The factors that determine the ongoing activity of these neurons have for several decades remained an unresolved core issue in the field of autonomic neuroscience (Coote, 2007; Guyenet et al., 2013): electrophysiological recordings from anesthetized animals suggest that RVLM sympathetic premotor neurons are a point of convergence for inputs from visceral and somatic reflex pathways (Brown and Guyenet, 1984; McMullan et al., 2008), from the central respiratory pattern generator (McAllen, 1987; Miyawaki et al., 1995; Verberne et al., 1999; Moraes et al., 2013), hypothalamus (Yang and Coote, 1998; Allen, 2002; Horiuchi et al., 2004; Korim et al., 2014), and higher centers including the prefrontal cortex and amygdala (Gelsema et al., 1989; Verberne, 1996), and that ongoing synaptic drive supports their activity (Lipski et al., 1996). However, the relative contributions of inputs from these regions, and in particular the level of input derived from local medullary neurons, has remained elusive.

Our strategy was first to use a recombinant herpes vector with a retrograde transduction profile, HSV-hCMV-YTB, to drive the expression of TVA, rabies glycoprotein, and a fluorescent reporter (YFP) in neurons that project to the interomediolateral cell column of the spinal cord (IML), a major site of termination of sympathetic premotor neurons. We then focally microinjected SADΔG(EnvA)-mCherry into the RVLM, selectively restricting its access to TVA-expressing bulbospinal neurons. This enabled us to map brain-wide sources of synaptic input to bulbospinal RVLM neurons using an image alignment tool based on the Waxholm volumetric atlas of the rat brain (Papp et al., 2014; Kjonigsen et al., 2015).

Experiments were approved by Macquarie University Animal Ethics Committee and conformed to the Australian Code of Practice for the Care and Use of Animals for Scientific Purposes.

The objective of the current study was to definitively and quantitatively map the afferent connectome of putative RVLM sympathetic premotor neurons, a regulatory axis through which convergent sensory and limbic inputs are thought to summate to produce baseline sympathetic nerve activity. We identified prominent inputs from well-established neurohumoral and viscero-sympathetic actuators such as the area postrema and NTS, and inputs that spanned medullary autonomic (A5, RVLM, ventromedial medulla, CVLM, LTF) and respiratory nuclei (Bötzinger, pre-Bötzinger Complex, PiCo, rostral ventral respiratory group), as well as supramedullary autonomic nuclei (paraventricular nucleus, ventrolateral PAG, lateral hypothalamus) (Dampney, 1994b; Guyenet, 2006; Card et al., 2011; Stornetta et al., 2016). However, the most striking feature of the connectome is its diffuse distribution, which contrasts with the nodal ball-and-stick schemes sometimes used to conceptualize it (Dampney, 1994a; Pilowsky and Goodchild, 2002; Guyenet, 2006), and its strong weighting toward local inputs, with around 14% of input neurons residing within the RVLM region and 30% of inputs lying within 1 mm of the RVLM epicenter. In contrast, the dataset includes a relatively low number of inputs from the forebrain (3%) and midbrain (3%), and a virtual absence of inputs from locus coeruleus, the amygdala, cortex and subfornical organ and median preoptic nucleus, despite compelling prior evidence for both functional and neuroanatomical connectivity (Dampney et al., 1987; Cassell and Gray, 1989; Gelsema et al., 1989; Verberne, 1996; Saha, 2005; Card et al., 2011; Bou Farah et al., 2016). This pattern of connectivity is consistent with qualitative data from in a similar study that targeted TH-synthesizing RVLM neurons in the mouse (Stornetta et al., 2016) and supports the argument that RVLM sympathetic premotor neurons play little, if any, role in the generation of sympathoexcitatory responses to acute psychological stress or conditioned fear (Dayas et al., 2001; Carrive and Gorissen, 2008; Furlong et al., 2014; Dampney, 2015), although it does not rule out the possibility that input from cortical and limbic structures could be indirect, perhaps involving disynaptic relays through the abundant monosynaptically connected RVLM interneurons identified in this study. Alternatively, as the current study focused exclusively on inputs received by sympathetic premotor neurons that project to the T2 thoracic spinal cord, it could be that higher centers provide a more (numerically) significant input to RVLM neurons that project to other spinal segments.

The minor input received from supramedullary centers is unlikely to represent a limitation of SADΔG(EnvA), as distance is not thought to impair its labeling efficiency (Callaway and Luo, 2015; Schwarz et al., 2015). Furthermore, inputs from the amygdala and cortex were observed in control experiments using rabies glycoprotein-transcomplemented SADΔG-mCherry and retrograde herpes vectors, and have been reported in studies that used classical chemical tracers (Bowman et al., 2013; Bou Farah et al., 2016), confirming that these structures innervate the RVLM region; apparently just not those neurons targeted in the current study. Taken at face value, these data suggest that inputs from numerous local brainstem structures predominate in their capacity to influence sympathetic nerve activity compared to more distant regions, with the caveat that identification of monosynaptically connected neurons does not necessarily denote activity of those inputs: rabies spread is a function of synaptic strength, not synaptic activity (Ugolini, 1995; Brennand et al., 2011).

RVLM sympathetic premotor neurons include both C1 and non-C1 neurons, which differ in their functional and neurochemical phenotypes (Schreihofer and Guyenet, 1997; Stornetta et al., 2001) and may play differing roles in the generation of baseline sympathetic nerve activity and the elaboration of sympathetic reflexes (Schreihofer and Guyenet, 2000; Schreihofer et al., 2000; Madden et al., 2006; Burke et al., 2011). We used a herpes vector with a retrograde tropism to transduce putative sympathetic premotor neurons based on their axonal trajectory, allowing us to target both C1 and non-C1 bulbospinal neurons. One potential limitation of this approach is that SADΔG(EnvA)-mCherry injected into the RVLM could have theoretically infected TVA-expressing spinally-projecting neurons from other brain regions, either as a result of trans-synaptic spread (to spinally projecting neurons that provide collateral input to RVLM sympathetic premotor neurons) or by direct infection of axons that traverse the RVLM. Evidence of such “ectopic” starter neurons was apparent in pilot experiments, but was largely eliminated by limiting the interval between herpes and rabies injections to 24 h and by diluting the herpes vector. We speculate that TVA is differentially expressed on the somata but not the axons of herpes-transduced neurons at 24 h, accounting for the selective accessibility of RVLM neurons at this time-point, and conclude that bulbospinal RVLM neurons do not receive collateral input from other spinally projecting neurons because we did not see any evidence of ectopically infected spinally projecting neurons in other brain regions.

A notable deviation from the standard autonomic circuit model is the discovery of a conspicuous input from the region of the prepositus hypoglossi, a brainstem nucleus conventionally associated with vestibulo-occulomotor integration (McCrea, 1988) but also previously shown to drive pressor responses to glutamate microinjection (Talman and Robertson, 1991). Rabies-labeled prepositus neurons reside along the axonal trajectory of RVLM bulbospinal neurons (Lipski et al., 1995; Stornetta et al., 2016) but are not labeled by traditional retrograde tracers injected at the RVLM pressor region (Dampney, 1994b) (or retrograde rabies/herpes vectors—data not shown): we speculate that trans-synaptic infection of this population may have been via axoaxonic contacts or via the distal dendrites of starter neurons. Future functional studies will be required to elucidate the functional significance of this input.

Our dataset provides unequivocal evidence of monosynaptic inputs from respiratory subnuclei, providing clues to the neuroanatomical substrate responsible for respiratory-sympathetic coupling, the entrainment of sympathetic nerve activity to the phasic bursting of the central respiratory rhythm generator. Although direct interaction between respiratory and cardiovascular neurons in the brainstem has long been suspected (reviewed by Pilowsky et al., 1996; Taylor et al., 1999; Zoccal et al., 2014), technical difficulties have until now precluded unambiguous examination of this hypothesis. We found no evidence that input neurons in the Bötzinger region were glycinergic, (a functional marker of Bötzinger respiratory neurons: Schreihofer et al., 1999; Ezure et al., 2003), undermining the hypothesis that Bötzinger neurons are a source of inhibitory input to sympathetic premotor neurons (Sun et al., 1997). This is consistent with the observation that blockade of RVLM glycinergic transmission does not alter respiratory-sympathetic coupling (Guyenet et al., 1990), and suggests that the close appositions identified by Sun et al. (1997) may not represent functional synapses, an acknowledged limitation of light microscopy for reliable identification of synaptic contacts (Murphy et al., 1995; Descarries and Mechawar, 2000).

On the other hand, we identified pre-Bötzinger Complex input neurons that were immunoreactive for NK1R, a putative marker of glutamatergic inspiratory neurons in this region (Gray et al., 2001; Guyenet and Wang, 2001; Stornetta et al., 2003), providing a plausible structural basis for the inspiratory-locked sympathoexcitation evident in some sympathetic outflows (reviewed by Häbler et al., 1994; Pilowsky et al., 1996) and RVLM sympathetic premotor neurons (McAllen, 1987; Moraes et al., 2013). We also identified inputs from ChAT-immunoreactive neurons in the region of the LTF that may correspond to the PiCo, a group of glutamatergic cholinergic neurons recently reported to underlie the generation of postinspiratory activity in the mouse (Anderson et al., 2016). Although functional studies of the PiCo are yet to be replicated in the rat, we tentatively hypothesize that excitatory input from its analog could underlie the prominent post-inspiratory activity seen in rat lumbar and splanchnic sympathetic nerve activities (Haselton and Guyenet, 1989; Burke et al., 2010; Korim et al., 2012).

Few input neurons were found in the medullary retrotrapezoid/parafacial region that lies close to the ventral brainstem surface immediately adjacent to the RVLM. As this population is not thought to directly contribute to respiratory-sympathetic coupling (Moraes et al., 2012) their neurochemical phenotype was not studied in detail: such inputs that were identified did not conform to the small fusiform morphology or reticular distribution typical of RTN CO₂-sensing neurons (Stornetta et al., 2006).

Like other mesoscale connectomes, the RVLM connectome is diffuse and highly non-uniform in its distribution. This level of complexity makes it difficult to conceptualize as a whole, and makes identification of its major subdivisions susceptible to operator bias. Principle component analysis (cluster analysis) provides a simple and unbiased way to do this but suffers some potential drawbacks. First, selection of the appropriate number of clusters (the value of k) is inexact; it can be estimated by examining the degree of variance accounted for by iterative increases in k and identifying the “elbow” after which the relationship between k and variance becomes linear (Thorndike, 1953: in the current study k = 12). However, as this value is a function of the variance of the dataset, one would expect k to increase if more experimental data were added—in other words, if one analyzed 10,000 input neurons instead of 1,000, the sensitivity of the analysis would be higher and it would resolve clusters not currently detectable. Second, cluster analysis was performed on the pooled connectomic dataset, rather than on each individual experiment. Although the overall normalized distribution of inputs was reproducible between experiments, this means that experiments in which a large number of input neurons were detected influence cluster detection more than those in which fewer inputs were detected. Taken together, cluster analysis of connectomic datasets should be seen as a useful platform for breaking complicated datasets down into a readily digestible overview, rather than a definitive structural analysis tool.

Overall, the spatial distribution of the RVLM connectome adheres to two key organizational principles evident when examining interregional connectivity profiles [i.e., at the macroscopic resolution: reviewed by Bullmore and Sporns (2012)]. The first is reciprocity, that is, regions innervated by the collateral branches of putative sympathetic premotor neurons such as the NTS, RVLM, CVLM/A1, raphe and PAG (Card et al., 2006; McMullan and Pilowsky, 2012; Stornetta et al., 2016) all in turn provide afferent input to RVLM sympathetic premotor neurons. The second is spatial embedding: RVLM sympathetic premotor neurons are more likely to receive input from nearby neurons than distant ones. The spatial embedding evident in the rat RVLM connectome is in striking accordance with the nodal edge-length distribution of the mouse mesoscale connectome, the most detailed whole brain connectivity atlas currently available (Oh et al., 2014), and in particular the spatial distribution of reciprocally connected nodes (Henriksen et al., 2016) (see Supplementary Image 3 for direct comparison). This supports the proposition, reviewed by Bullmore and Sporns (2012), that highly conserved general principles likely govern circuit topography irrespective of species, brain region, cell type or scale, and provides for the first time cell-type specific connectomic data that can be used to test general brain connectivity models.

Cell-specific connectome tracing provides unique insights into the organization of circuits that control populations of neurons selected based on their genetic, functional, or projection profiles (Yonehara et al., 2013; Pollak Dorocic et al., 2014; Pollock et al., 2014; Wertz et al., 2015). However, one of the challenges to this approach lies in the analysis and presentation of the data. A common strategy is to align histological images with corresponding plates from a stereotaxic atlas, count the number of neurons that reside within each segmented region, and then tabulate the output (Pollak Dorocic et al., 2014; Schwarz et al., 2015). Although superficially straightforward, we found that the reliability of this approach depends greatly on the expertise of the operator and can be compromised by even small misalignments in cutting plane. We calculate that each degree in deviation in the lateral plane results in an offset of ~280 μm rostrocaudal at the widest point of the rat brain, (which is further exaggerated by any deviation in the dorsoventral plane). Perfect histological cutting planes are virtually impossible to achieve, and so data presented in this way are intrinsically error-prone. The longevity of data presented as #cells/region are further undermined by the evolving taxonomy of the brain, variation in segmentation and nomenclature used by different reference atlases, and the ongoing discovery of new cell groups. An alternative approach is to present imaging data in its entirety, either as an online repository of still (Pollak Dorocic et al., 2014) or video images (Stanek et al., 2014). This allows other investigators to access the raw data, but independent analysis is hampered by the enormous volume of imaging data and the time commitment required to reanalyze it.

Our approach sidesteps both of these limitations: alignment of a 3d reference atlas to the histological image (rather than the other way around) eliminates potential errors caused by imperfect cutting angle or tissue distortion. Second, the compact size (current study <100 kb) and standardization of the positional metadata make it amenable to sharing and independent reanalysis, and eliminates the need for other investigators to directly interact with cumbersome raw imaging data (current study >300 Gb). Routine publication of such metadata would enable researchers to present complex neuroanatomical datasets in a standardized and transparent format that improves utility and reproducibility.

The promise of the connectomic approach is reflected by its widespread adoption and the continued refinement of connectome-tracing viral vectors (Osakada et al., 2011; McGovern et al., 2015; Kim et al., 2016; Reardon et al., 2016; Zingg et al., 2017), but tools that compile and quantify connectomic datasets remain primitive. The approach described here allows the quantification, standardization, and sharing of complete neuroanatomical datasets, and therefore provides a platform by which researchers may independently visualize, analyze, and compare each other's data. Bioinformatics at a single-cell resolution may allow researchers to move beyond the ball-and-stick diagrams often used to conceptualize the organization of spatially diffuse real-world neuronal networks (Rockland, 2015) and realize the full potential of novel connectomic technologies.

Conceptualization: SM, AG, and AA; Methodology: BD, RN, and SM; Software: JB; Formal analysis: SL, BD, AT, and SM; Investigation: BD, AT, SL, CM, and SM; Resources: RR, PB, RN, and JB; Data curation: SL, BD, and SM; Writing—Original Draft: BD, AG, and SM; Writing—Review and Editing: All authors; Visualization: BD and SM; Supervision: SM; Project Administration: SM; Funding Acquisition: SM, AA, and AG.