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

Abstract

The larva of Drosophila melanogaster is emerging as a powerful model system for comprehensive brain-wide understanding of the circuit implementation of neural computations. With an unprecedented amount of tools in hand, including synaptic-resolution connectomics, whole-brain imaging, and genetic tools for selective targeting of single neuron types, it is possible to dissect which circuits and computations are at work behind behaviors that have an interesting level of complexity. Here we present some of the recent advances regarding multisensory integration, learning, and action selection in Drosophila larva.

Figure 1

Approaches for studying neural circuit in Drosophila larva. i. Drosophila larva beneficiates from a comprehensive genetic toolkit for selective targeting of uniquely identified neuron types, often a single pair of left-right homologous neurons [21,22]. Left, schematic of the tripartite organization of the larval CNS: brain, SEZ and VNC. Right, confocal image of a CNS immunostained against n-cadherin (blue) and GFP expressed in a single pair of MB output neurons (green). ii. Synaptic resolution connectome. A full CNS imaged with EM has been reconstructed; the connectivity of many neural circuits at synaptic resolution is now known. Quantitative studies have shown that strong connections are symmetrical between left and right side [7] and conserved across individuals and larval stages [19]. iii. Accessibility to physiological activity: in vivo imaging [26, 27, 28, 29, 30, 31], ex vivo whole-brain imaging [23], intracellular recordings of uniquely identified neurons that can be selectively labelled by GFP [8,9,18,77]. iv. Substantial behavioral repertoire: discrete actions such as run, stop, head cast, turn, hunch, backup, roll can be automatically tracked and categorized. v. Modelling approaches. Thanks to the few number of larval neurons (ca. 15,000 including 2500 brain neurons), reconstructed connectivity can be implemented in an artificial neural network [8,16,18]. Behavioral hallmarks can also be reproduced by an agent-based model [52,56,67].

Figure 2

Circuits for nocifensive behavior in larva. Sensorimotor circuits for escape response selection to a nociceptive stimulus have been particularly well studied in Drosophila larva. The particular nocifensive behavior depends on the type of threat: predator attack, noxious heat or harsh touch can evoke rolling response [6,38,60,61], instead weaker and less threatening stimuli can evoke fast forward crawling or back-up [34,39,62]. The nature of the response relies on the type of somatosensory neurons co-activated with the polymodal larval nociceptors (multidendritic class IV ‘md-4’, magenta). Components of the circuits reconstructed downstream of the nociceptive neurons [6,61] and of the main somatosensory neurons (chordotonal ‘cho’, ‘md-2’ and ‘md-3’, greens) [7,8,34] are depicted. Circles indicate pairs or class of neurons, plain lines are direct connections, dotted lines indirect ones, all reconstructed in the same EM volume. i. ‘Wave’ neurons repeatedly tiling the VNC integrate spatially defined multimodal inputs to promote either run or back-up in response to, respectively, posterior or anterior harsh touch stimulation [34]. These neurons also indirectly contact the roll-promoting neuron ‘Goro’ [6,34]. ii. Multiple levels of multisensory integration occur during the selection of rolling: both early in sensory processing, and at more downstream nodes in the network [6, 7, 8,19,34,60,61]. Modelling shows that multiple levels of multisensory integration enhances action selection [6]. iii. ‘Basin 4’ super-additively integrates ‘cho’ and md-4’ inputs. Increased Basin-4 activation by multisensory cues, in turn increases the likelihood of rolling behavior, through Goro activation [6].

Figure 3

Circuits for high-order integration and value computation. In the larval brain the MB and the LH receive convergent sensory inputs from olfactory, thermosensory and visual projection neurons (green and light blue) [16,14,15]. The LH neurons (light brown) are assumed to parse the information according to innate valence. The MB Kenyon cells (dark brown) are thought to decorrelate sensory signals via highly divergent connections, expose them to reinforcement-gated plasticity (red open circles), and generate learnt valence signals (purple). MB is characterized by a recurrent architecture: a GABAergic neuron (black) gathers signals at the KCs axons and feeds back onto the KCs dendrites [16,17]; in addition, KCs signal back to the teaching neurons (mostly DANs, red) [16,49] and so do the MB output neurons (purple and orange feedbacks) [16,18]. These multilayered paths provide neural substrates for integrating prediction to the teaching signals, in addition to other inputs from the SEZ, for adaptive memory update (See Box i [18]). Further down the circuit, MB and LH are likely integrated for valence signals that can be used as instructive signal for navigation (orange, See Box ii [54] and iii [35]). The SEZ (grey) transforms more [16,35] or less [3] integrated signals into behavior. i.Top panel, the experimental exploration of the responses of some teaching DANs, combining calcium transient and optogenetic stimulations, confirms that DANs integrate external (multisensory nociceptive neurons Basin) and memory-related (MB outputs neurons) inputs combining the recording of cell activity and optogenetic stimulations. Bottom panel, In an artificial neural network incorporating the connectivity of the MB and the response tuning of the DANs, discriminative signal for conditioned stimuli that predict (CS+) or not (CS-) the unconditioned stimulus (US) emerges in the DANs after associative training [18]. ii. Reverse-correlation experiments [13,53, 54, 55] link olfactory and visual inputs to navigation [54]. Under fictive odor stimulation (the optogenetic activation of receptors for attractive odors continuously varying in intensity), redirecting turns are initiated by a decrease in the signal, while under real visual stimulation (aversive blue light varying in intensity) turns are initiated when the signal increases. The sensorimotor transformation estimated for the combination of inputs suggests a linear integration of olfactory and visual inputs, probably readable at the level of LH output. iii. PDM-DN, a neuron reconstructed in EM, downstream of LH neurons and descending to the SEZ, is necessary for navigating towards an odor source [35]. The optogenetic activation of this neuron triggers stop and redirection in crawling animals by stopping the wave of body contraction that goes through the body and shutting off the activity of the corresponding motoneurons.