Useful road maps: studying Drosophila larva’s central nervous system with the help of connectomics

Highlights • Drosophila larva enables combining comprehensive synapse resolution connectivity maps with cellular-resolution activity maps and behavior maps.• This approach provides a way to elucidate neural implementation of universal brain computations such as multisensory integration, learning, and action-selection.• Early multisensory convergence recruits specific sensorimotor loops for specific actions.• Higher-order brain areas integrate modalities in a more centralized way and produce high-order, valence-like signals for goal-oriented behavior.


Introduction
Ability to sense, act, remember, or anticipate emerges from the way the nervous system is organized into networks that allow signals to flow, interact, and change. Nerve cell types and numbers vary across different organisms, but many neuronal computations and behaviors appear to be done in a similar way across mammals and insects. For example, odor signals are processed via two parallel high-order pathways differing in representation and plasticity [1,2]. Feeding circuit is formed of multilayered loops linking external/enteric sensory inputs to motor/secretory outputs [3,4]. Dopaminergic neurons encoding reinforcement of different valences project onto spatially distinct associative regions [5].
Given the phylogenetic distance between insects and mammals (half a billion year), the fact that similar circuit solutions for complex problems have been conserved or reinvented through evolution points towards some fundamental principles linking structure and function in the central nervous system (CNS).
Here we review recently described circuits in the CNS of the larva of Drosophila melanogaster. The insect CNS comprises a brain, a subesophageal zone (SEZ), and a ventral nerve cord (VNC, Figure 1i). This tripartite organization is similar to the forebrain/cerebellum, brainstem, and spinal cord of vertebrates. The relatively small CNS of the Drosophila larva as a model offers a number of advantages for studying the circuit implementation of neural computations in a comprehensive way ( Figure 1).

Studying neural circuits in Drosophila larva
The early larval central nervous system contains fewer (ca. 15,000) and smaller neurons compared to the adult Drosophila making it amenable to relatively rapid electron microscopy imaging and circuit reconstruction with synaptic resolution [6] (Figure 1ii). So far, circuits for somatosensory processing [6-8] and motor programming [9][10][11] in the VNC, feeding [3] and neuromodulation [12] in the SEZ, as well as first-order sensory [14,15], and higher-order associative centers [16][17][18] in the brain have been reconstructed with synaptic resolution in the same EM volume of a first-instar (i.e. early larval stage) nervous system, and comprehensive reconstruction of the CNS is within reach. The reproducibility of this type of data has been tested by Gerhard et al. [19], who compared portions of nociceptive circuits in an early (first-instar) and a later (third-instar) stage larvae, and found that the fraction of total synaptic input associated with defined pre-synaptic partner is maintained despite a five-fold change in size. In many cases, the knowledge about connectivity could be augmented with immunohistology against neurotransmitters. This has allowed the identification of various types of circuit motifs [8,14-16,18].
The connectome alone is not sufficient for understanding circuit mechanisms [20], but it provides a necessary roadmap for mechanistic studies. In complement, Drosophila larva is amenable to a large variety of functional studies (Figure 1i,iii). With multiple genetic tools for selective targeting and manipulating individual cell types [21,22], functional connectivity between neurons can be tested by combining optogenetic activation of presynaptic neurons with e.g. electrophysiological recording [8,9,18] or calcium imaging of postsynaptic neurons [6,10,13,18]. Imaging the activity of motoneurons can also be used as a proxy for behavior and allows the visualization of fictive actions ii. Circuits for multisensory integration, learning, and action-selection As described in other organisms [57], specific sensory modalities act jointly early on in signal processing, at the first or second-order neuron, to build a meaningful representation of a stimulus. Later in processing, at higher-order brain regions, more sensory modalities can be combined to keep track of environmental variability. Circuit analyses in Drosophila larva are starting to elucidate the way in which all dimensions of multisensory experiences are integrated for appropriate action selection. touch sensors and nociceptors. Importantly, Wave neurons in posterior segments induce forward escape, whereas the 'Wave' neurons located in the anterior segments induce backward escape ( Figure 2i). EM reconstruction revealed Wave neurons in all segments receive synaptic input from nociceptive and touch-sensing neurons, but they have different output targets in different segments: the anterior Wave neurons synapse onto circuits in anterior segments that promote backward crawling, whereas the Wave neurons target posterior segments and promote forward crawling. Combining EM reconstruction with functional studies therefore revealed the way in which homologous neurons integrate and target different partners in different regions of the nervous system to mediate opposite behaviors.

Multisensory integration at early stages
The most vigorous and energetically costly rolling escape is elicited in response to predator attack [38]. Presenting mechanosensory cues with nociceptive ones facilitates rolling [6,61], likely because it better approximates predator attack which also stimulates multiple senses. Rolling is mediated by a command-like 'Goro' neuron that receives indirect functional inputs from both mechanosensory and nociceptive sensory neurons [6,61]. EM reconstructions of circuits downstream of mechanosensory and nociceptive neurons and upstream of the Goro neurons elucidated precisely where and how the information from these distinct sensory modalities is integrated. This revealed that mechanosensory and nociceptive information converges early on in the sensory processing hierarchy onto first-order multisensory interneurons that integrate the information superadditively [6] ( Figure 2ii). Multiple interneurons, gathering slightly different somatosensory modalities, relay multisensory nociceptive inputs to Goro and are sufficient to evoke rolling [6,61]. Additionally, later stages of multisensory integration at higher-order nodes enhance action selection [6]. Furthermore, when activated in combination with nociceptive neurons, touch-sensing neurons integrate the multiple mechanosensory inputs through sNPF feedback release and facilitate rolling [60]. Thus, knowing both how the different nodes of circuits are connected and their functional properties provided a mechanistic insight into the way in which nocifensive behaviors are selected ( Figure 2).

Higher-order integration: learning and value coding
Larval behavior is plastic and adapts to experience (review in [47]). Mushroom Bodies (MB) in the larval brain are necessary to form associative memory [43]. This memory is expressed by changing navigation towards a cue (e.g. an odor) that has been associated with a reward (e.g. sucrose) or a punishment (e.g. quinine).  With its few neurons and low redundancy, the brain of Drosophila larva is a model of choice to combine more experimental and theoretical approaches and deepen our understanding of these mechanisms.

Action selection
Drosophila larvae can generate many exclusive actions. Studies are beginning to elucidate the way in which multisensory inputs, higher-order valence signals, and context are used for action selection. In recent years progress has been made in understanding the circuits that mediate the selection of distinct types of escape responses in response to threatening somatosensory stimuli: roll, fast crawl, turn, back-up, and hunch. The selection of the most vigorous escape, roll, is enhanced by integrating nociceptive with mechanosensory inputs [6,61] (Figure 2). So far, many circuit motifs involved in the selection of these actions and their organization into sequences have been identified in the VNC. For example, reciprocally connected inhibitory interneurons mediate behavioral choice between hunch and turn in response to an air-puff [8], lateral disinhibition promotes sequence transitions between these actions, and specialized local feedback disinhibition provides positive feedback that stabilizes a behavior and prevents reversals to the preceding one. The combination of these interconnected circuit motifs can implement both behavior selection and the organization of behaviors into a sequence. Interestingly the connectome reveals that descending neurons from the brain and SEZ synapse onto many of the VNC interneurons involved in somatosensory responses [6,8]. The way in which brain circuits bias somatosensory choices is an exciting open question for the future.
A second major action selection paradigm in the larva is the choice whether to crawl or turn and which way to turn when navigating gradients of aversive or repulsive cues. The alternation between runs (forward crawl events), stops, and turns can be done without brain inputs [68] but their transitions based on sensory inputs rely on the brain (besides likely modalities detected by VNC sensory neurons, [8,68]). These choices require the computation of the value of the cue, which is done based on both innate and learnt valences. Ongoing reconstruction of neurons downstream MB and LH will inform about the way in which innate and learnt values are integrated. In addition, different modalities converge onto the same pattern of navigation responses [13, 37,40,41,53]. Different sensory inputs have been shown to be translated into turn action following the same signal transformation function [54] (Figure 3ii). It is therefore possible that all modalities converge onto a common center involved in computing an overall integrated value of a cue and guiding navigation. This interpretation also fits the convergence of projection neurons of various modalities onto the two brain structures MB

Conclusion
Providing a precise roadmap with the connectome, Drosophila larva helps formulate and test new hypotheses about the way in which neural circuits implement fundamental computations such as multisensory integration, learning, value computation, and action-selection.
Future research will include richer behavioral situations (e.g. [42,72,73]), in vivo recording of whole-brain activity [23-25], modelling approaches (e.g. [8,11,16,18,52]). In parallel, whole-brain RNAseq reveals genes expressed in individual neurons [74,75] that might be essential for these computations. The comparatively small size enables rapid reconstruction of connectomes from multiple individuals (e.g. [76]) and opens doors to an exciting new area of experimental connectomics to address questions about the structural correlates of specific memory traces, individual differences in circuits that underlie distinct personality traces, and discovering the effects of various mutants on the circuit architecture.  18. Eschbach C et al.: Recurrent architecture for adaptive regulation of learning in the insect brain. Nat Neurosci 2020, 23:544-555. EM reconstruction of the larval MB second-order circuit, combined with functional study shows that dopaminergic neurons with similar functions receive similar inputs. It also found that 50% of the inputs received by MB modulatory neurons are feedbacks from MB compartments of same and/ or of opposite valence. Artificial neural networks constrained with the reconstructed connectivity address the role of these feedbacks.
19. Gerhard S, Andrade I, Fetter RD, Cardona A, Schneider-Mizell CM: Conserved neural circuit structure across drosophila larval development revealed by comparative connectomics. Elife 2017, 6:e29089. EM reconstruction of segments of the VNC allows comparison of nociceptive circuits across larval individuals and stages showing remarkable conservation of circuits despite five-fold difference in size. In particular, the authors point that the fraction of total synaptic input associated with defined pre-synaptic partner is maintained.