Connectivity in the protocerebral bridge
A functional connectome is, by construction, sparser than can be predicted by light-level anatomy. Our study shows this most clearly in one neuropil, the PB (Figure \ref{704309}B). E-PG neurons are the only columnar type that are presynaptic in the PB, but their activation did not trigger a significant response in any of the 5 other columnar neurons we tested. This came as a surprise as we assumed the E-PG population would constitute the relay between EB and FB columnar systems. To verify that this lack of observed connectivity was not due to the recruitment of global inhibitory circuits, we also ran these experiments in the presence of picrotoxin, and did not observe any difference in responses (see Supplementary Figure \ref{841071}B). By contrast, the Δ7 interneurons are strongly activated by E-PG neurons, and their activation leads to significant responses in several columnar neuron types (E-PG, P-EN1, P-EN2, P-F1N3 and P-F3N2v). The Δ7 neurons, therefore, appear to constitute an important bottleneck in the system (Figure \ref{704309}B), and may serve as the only strong link between columnar neurons in the PB. The response profiles following Δ7 activation are also unusually complex (see Supplementary Figure \ref{302685}): P-ENs display mild activation, E-PG and P-F3N2v inhibition, and P-F1N3 strong rebound excitation (see Figure \ref{489066}Biii).
Connectivity in the EB columnar system, the ring attractor network
Figure \ref{704309}Dii shows the subpart of the network that has been proposed to sustain the ring attractor representation of heading \cite{Green2017,Turner-Evans2017}. One hypothesized feature of such a circuit is a large degree of recurrence between the different EB columnar types. In particular, P-EN to E-PG reciprocal connections are important for models of the rotation of the bump. While we found strong support for the P-EN1 to E-PG connection, the E-PG to P-EN1 connection that we reported functionally under a stronger stimulation protocol \cite{Turner-Evans2017} may be mediated through the Δ7 interneurons. A few other connections were found in the EB (for example, P-EN1 to P-EN2), but it is important to stress that not all combinations could be tested due to limitations in the genetic reagents available. For example, the role of the P-EG neurons in this circuit, remains unclear. A key additional type that our results suggest may contribute in important ways to the persistence of activity in this circuit is the AMPG-E neuron, which appears to provide localized excitatory feedback to the E-PG neurons.
Discussion
The dataset presented in this study constitutes a resource for the growing community of researchers interested in the central complex. While similar coarse functional connectivity techniques have been used to map short pathways in previous studies, this is, to our knowledge, the first extensive dataset of its kind. We hope that it will become an evolving source of information, which we expect to be most useful when combined with other complementary data sources, such as EM-level anatomical connectivity and high-resolution gene expression profiles. Such combined data would constitute a solid base to build constrained network models of the central complex, and to generate detailed hypotheses of its function. As with any large dataset, we see this effort mainly as a starting point for more detailed research.
Limitations of the method
The connectivity technique we applied has several limitations that are important to keep in mind. First, the combination of CsChrimson and GCaMP does not guarantee that the connections observed are direct and monosynaptic. However, the large set of controls with cell-type pairs whose processes do not overlap provides a statistical framework to interpret the results — not surprisingly, uncertainty is highest for weak connections. A more pressing issue concerns sensitivity: what can be detected is limited by the stimulation protocol and the sensitivity of GCaMP6m. Thus, an absence of a post-synaptic response cannot be interpreted as an absence of a connection. The fact that inhibitory responses are visible, and that strong responses saturate with the range of stimulations used (see Supplementary Figure \ref{723169}) is reassuring. It is likely, however, that EM-level anatomy will reveal that some weak synaptic connections have been missed by this technique. Their functional importance will need to be investigated using more sensitive methods, for example, intracellular electrophysiology. Further, we relied on full-field stimulation of populations of specific neuronal types, which comes with its own drawbacks: this approach provides no access to connectivity between neurons of the same class, and does not account for potential non-physiological network effects. One such effect would be the recruitment of global inhibitory networks that could mask an otherwise excitatory connection. However, whenever we suspected this could be a possibility, we controlled for it by blocking inhibition with picrotoxin, and never saw evidence of a significant effect (Supplementary Figure \ref{841071}B). Even though picrotoxin was effective in blocking the inhibitory responses we observed (Supplementary Figure \ref{841071}A), we cannot exclude the possibility that picrotoxin-insensitive inhibition (e.g. GABA-B, \citealt{Olsen_2008}) might be present in the network. Furthermore, the fact that we stimulate whole presynaptic populations means that the strength of connections we report is influenced both by neuron-to-neuron transmission strength and the degree of convergence in the network. Finally, all our experiments were performed on ex-vivo brain preparations. Given the variety of neuromodulators that operate in the central complex \cite{Kahsai2012}, it is likely that functional connectivity within this region is modulated by brain state. Consistent with this possibility, we saw that the fluorescence baseline tended to fluctuate spontaneously during the course of our experiments in most types recorded (Supplementary Figure \ref{998826}A). Although increases in baseline activity allowed us to detect inhibitory responses, we noticed that excitatory responses also occasionally depended on this baseline fluctuation (Supplementary Figure \ref{998826}C).
Central complex motifs
The connectivity matrix we obtained is sparser than that predicted by light level anatomy. Our results suggest that the Δ7 interneurons are a bottleneck for information processing in the PB. This is all the more interesting given the range of responses evoked by Δ7 stimulation (Supplementary Figure \ref{302685}). Properties of the synapses that Δ7 neurons make with their post-synaptic partners may play a primary role in the way that a heading signal is generated and maintained in the EB columnar system, and also in how it may be transferred to the FB columnar system. Every Δ7 neuron innervates all columns of the PB, and has presynaptic-looking processes in two columns. The fact that a neuron with such extensive arbors participates in a circuit where representations are spatially restricted (the bump of activity is limited to a few neighboring columns) suggests that understanding local processing at the single neuron level might be critical to a complete understanding of how the circuit operates. Interestingly, the same puzzle occurs at the input side of the system with the ring neurons, which similarly innervate the entire circumference of the EB.
The fact that several sources of input are inhibitory raises the question of how activity is maintained in the region. Candidate mechanisms are the uncovered excitatory inputs into the PB and EB, recurrent connections in the EB and intrinsic properties of neurons \cite{Egorov2002,Yoshida2009,Russell1982} — some cell types, for example, showed robust post-stimulation rebounds (see Figure \ref{489066}Bii). It is also possible that our selection of cell types and our methods missed some sources of excitation.
The range of inputs revealed here opens many avenues for investigation. Whereas some ring neuron subtypes have received considerable attention \cite{Sun_2017,Shiozaki_2017,Seelig2013}, most PB inputs and LAL-noduli interneurons have not yet been characterized. A recent study in the sweat bee \cite{Stone2017}, for example, reported that one of the LAL-noduli interneurons — a likely input to the FB system — carries regressive optic flow signals.
The specific functions subserved by the network motifs that we have uncovered may only become clear with functional studies in behaving animals. A key puzzle set up by our findings is the small number of output channels of the central complex. Our results are consistent with the LAL being the primary output structure for the central complex \cite{Chiang2011,Hanesch1989}, although the structure also acts as an input region (via ring neurons and potentially via IMPF-L neurons). While it is possible that our selection of Gal4 lines was unintentionally biased against output neurons, or that our technique otherwise missed a number of output pathways, the picture of the central complex that emerges is of a densely recurrent sensorimotor hub with relatively low dimensional output (much as proposed by some models e.g. \citealt{Stone2017,Fiore2015,Strauss_2010}). The implications of this bottleneck for motor control remains a challenge for future studies to resolve.
Materials and methods
Fly stocks and crosses
For any given pair of neurons, drivers were chosen, and the overlap between pre- and post-synaptic looking regions assessed based on publicly available expression patterns (\citealt{Tirian2017,Jenett2012}, see Supplementary Figure \ref{147797}) digitally aligned on a common reference brain (as described in \citealt{Aso2014}). For every LexA driver used, we prepared two stocks containing GCaMP6-m \cite{chen_ultrasensitive_2013} and CsChrimson \cite{klapoetke_independent_2014} under LexAop (resp. UAS) or UAS (resp. LexAop) control : XXX-LexA;13XLexAop2-IVS-p10-GCaMP6m 50.629 in VK00005, 20xUAS-CsChrimson-mCherry-trafficked in su(Hw)attP1 and XXX-LexA;20xUAS-IVS-GCaMP6m 15.629 in attP2, 13XLexAop2-CsChrimson-tdTomato in VK00005. Those stocks were then crossed to a Gal4 driver or a split-Gal4 \cite{luan_refined_2006} driver for the experiment. In the split-Gal4 case, the two split halves are inserted in attP40 and attP2 respectively. To avoid transvection between the split and the LexA driver \cite{mellert_transvection_2012}, we inserted the LexA drivers in alternative sites, either su(Hw)attP5 \cite{pfeiffer_refinement_2010} or VK22 \cite{venken_pacman:_2006}, and used the splits exclusively in combination with those lines after checking their expression patterns. The list of drivers used and the corresponding cell types are given in Table \ref{786393}. Throughout this paper we follow the naming convention set out in \citealt{Wolff2015} for full names, and abbreviated following the scheme described in \citealt{Kakaria2017} and used in \citealt{Green2017} and \citealt{Turner-Evans2017}. For each cell type, we labeled every region innervated as pre- or post-synaptic (or both): this was done at the resolution of the glomerulus for the PB, the layer for the FB and the individual nodulus. We divided the LAL in three zones based on the overlap between the lines used. Existing subdivisions for the EB and Gall were preserved. This labeling was used to evaluate whether the arbors of a given cell-type-pair overlapped.