Connectomic reconstruction of a female Drosophila ventral nerve cord

Reconstruction of a female fly VNC

We previously imaged an electron microscopy volume of a Drosophila FANC and manually reconstructed a small number of VNC neurons and synapses¹. Here, to accelerate efforts to map the VNC connectome, we applied deep learning techniques to automatically reconstruct neurons and synaptic connections throughout the FANC. We first refined the alignment of serial-section image data using self-supervised convolutional neural networks (CNNs) (Fig. 1). We then used other CNNs to segment the dataset into neurons (Fig. 1c,g), and to predict synapse locations and pre- and postsynaptic partners²⁰ (Fig. 1c). To identify all cells intrinsic to the VNC, we detected nuclei using another CNN (Fig. 1d–f and Methods), which identified 17,076 putative cells. By manual inspection, we classified 14,621 (85.6%) of these nuclei as neuronal. The remaining objects were glia nuclei (11.9%) or non-nucleus objects (2.5%) (Extended Data Fig. 1). The CNN, trained on neuronal nuclei, likely did not detect most glia nuclei.

The automated cell segmentation was ingested into a Chunked-Graph²¹, which enabled human experts to proofread errors in the reconstructions, such as attaching cell bodies and splitting neuronal processes from glia. We imported synapse and cell body predictions into the connectome annotation versioning engine (CAVE), so that the associated cell segmentation for each annotation would be dynamically updated during proofreading. This system enabled us to query up-to-date connectivity and associated metadata, such as cell-type annotations. The suite of tools available for analysis of the FANC dataset are described in Extended Data Fig. 2a–e, and includes a web platform (https://braincircuits.io) for interacting with the FANC dataset. Following the example of FlyWire²¹, we organized a community currently comprising more than 200 users from more than 30 laboratories who used these tools to proofread the FANC cell segmentation and analyse their circuits of interest. Representative examples of proofread motor and premotor neurons (preMNs) are shown in Fig. 1g.

To assess the quality of the FANC automated segmentation and synapse prediction, we focused on leg MNs (Fig. 2a). MNs are among the largest cells in the Drosophila nervous system (Extended Data Fig. 1a–c). The largest leg MN in the fly, the fast tibia flexor²², has nearly 1 cm of total dendritic length (a typical Drosophila body is around 3 mm in length). Manual tracing and synapse annotation of the fast tibia flexor MN took an expert tracer approximately 200 h. For comparison, proofreading the automated reconstruction of this same MN to correct errors in the automated segmentation took 2 h. Overlaying the manual and automated reconstructions for individual MNs (Fig. 2b,c) illustrates that the automated methods, followed by proofreading, effectively captured the major dendritic branches and more than 80% of the total cable (Fig. 2d).

Most differences between manual tracing and automatic segmentation occurred at fine branch endpoints (Fig. 2b). The number of manually and automatically annotated synapses differed by less than 9% for the largest tibia flexor neuron and the left accessory tibia flexor (Fig. 2e). Automatic synapse detection found over 50% more synapses for the other two MNs, owing to incomplete manual tracing in synapse-dense regions of the neuropil (Extended Data Fig. 3e). The mean distance from a manually annotated synapse to a predicted synapse was less than 1.5 μm (Extended Data Fig. 3b), demonstrating the accuracy of automatic synapse prediction.

To assess how the automatic segmentation and synapse prediction recovered synaptic connectivity, we computed precision and recall curves for each of the four MNs (Fig. 2f). We created lists of presynaptic partners from either the automatically predicted synapses or the manually annotated synapses, sorted the lists according to the number of synapses in the connection, and applied a minimum threshold of three synapses to be included as a connected partner. Precision is the percentage of the top k predicted partners that are in the list of ‘true’ partners based on manual synapse annotation (Fig. 2f, left). Recall is the fraction of the top k ground-truth partners that are in the list of predicted partners (Fig. 2f, right). For each MN, we found that over 80% of predicted partners with more than three synapses are true partners (precision), and over 80% of true top k partners with more than three synapses are predicted partners (recall).

In summary, we present the alignment, segmentation and automatic synapse prediction for the FANC electron microscopy volume. Our comparison with manual tracing and synapse annotation provides confidence that the automated segmentation, followed by appropriate proofreading (see Methods), provides a level of accuracy similar to that from manual tracing. On average, proofreading MNs and their presynaptic partners required 20 edits per neuron and took 100× less time, markedly accelerating reconstruction of the connectome.

Identification of leg motor neurons

Knowing which MNs innervate which muscles is crucial to interpreting the function of motor circuits in the connectome. We therefore set out to identify the peripheral muscle targets of all MNs innervating the front (T1) legs of the fly^5,6,23,24. To reconstruct the MN projection map, we integrated information across three imaging datasets that collectively span the VNC and leg (Fig. 3a). First, we identified and proofread all the MNs innervating the T1 leg in the FANC electron microscopy dataset (69 in left T1, 70 in right T1). Second, we traced the motor nerves and axons into their target muscles using an X-ray holographic nano-tomographic (XNH) dataset of the fly’s front leg³. Third, we screened a large collection of VNC neurons sparsely labelled with the multi-colour Flp-out (MCFO) technique to identify GAL4 driver lines that label leg MNs². We imaged GFP expression of each genetic driver line in the front leg to identify the muscle target of each MN axon. We then compared the dendritic morphology of the genetically labelled MNs to those reconstructed from FANC (Fig. 3). We focused our efforts on the left MNs because the left prothoracic leg nerve (ProLN) of the FANC dataset is more intact than the right ProLN (Extended Data Fig. 4a–g).

We illustrate the process of MN identification across the three imaging modalities with two examples (Fig. 3). First we consider the fast and slow tibia extensor MNs (Fig. 3b,c) that innervate the tibia extensor muscle²⁵. In the XNH volume, we identified two candidate MN axons that leave the leg nerve in the femur and pass through a membrane separating the tibia extensor muscles from the tibia flexor muscles (arrows in Fig. 3b). We identified four GAL4 lines from the Janelia Fly-Light collection² that label two MNs that project through the same nerve to innervate the extensor muscles. The dendritic morphology of the genetically labelled MNs closely matched the dendritic morphology of two MNs that we reconstructed in the FANC electron microscopy dataset (Fig. 3c). We therefore concluded that these are the fast extensor of the tibiae (FETi) and slow extensor of the tibiae (SETi) MNs. The SETi receives 7,090 input synapses, and the FETi receives 14,904 input synapses, the most for any front leg (T1) MN. The only other neuron in FANC with more synapses is the hind leg FETi, with 15,474 input synapses.

As a second example, we consider the large trochanter flexor muscle in the coxa (Fig. 3d). Compared with the tibia, whose extension is controlled by the SETi and FETi MNs, little is known about the neural control or biomechanics of the trochanter. Anterior muscle fibres attach to one side of the trochanter flexor tendon and posterior fibres attach to the other side (Fig. 3d). We traced eight MNs that innervate this muscle in the XNH volume: five innervate the anterior fibres via the prothoracic accessory nerve (ProAN) and three innervate the posterior fibres via the ventral prothoracic nerve (VProN; Fig. 3e). In FANC, we found six MNs with morphology that resembled previous images of trochanter flexor MNs²³: three that leave via the ProAN and three that leave via the VProN (Fig. 3f). Confocal imaging of genetic driver lines showed that at least one MN innervates the proximal fibres of the trochanter flexor and has a soma on the posterior cortex of the neuropil⁵. From this evidence, we deduced the identity of all eight MNs that innervate the trochanter flexor muscle.

We applied a similar process of deduction to map all 69 left T1 MNs to 18 target muscles (Fig. 4). A comprehensive description of leg MN identification from electron microscopy, XNH, light-level imaging of genetic driver lines and the literature is provided in the Supplementary Methods. We also used the XNH dataset to determine the number of muscle fibres in each leg muscle (Fig. 4a), along with the fibre origins and cuticle attachment points. Combining this anatomical information with joint kinematics recorded from walking flies²⁶ (Fig. 4b), leads to predictions of how MNs may contribute to distinct aspects of the step cycle, such as leg swing versus stance (Fig. 4c). We note, however, that the precise timing of MN spikes dictates how muscle contraction controls body movement²⁷. Because the dendritic morphology of leg MNs is stereotyped from fly to fly, this atlas (Fig. 4c,d) can also be used to identify MNs in other datasets¹⁹.

In the process of reconstructing and identifying leg MNs, we made other observations that were only possible due to the reconstruction of neuromuscular anatomy across imaging scales. For example, the right front leg of FANC contained 70 MNs (compared with 69 in the left T1). The extra cell appears to be a second tarsus levator MN. Supernumerary MNs have been described in the locust leg motor system²⁸, but—to our knowledge—have not been described in flies. We also found that six MNs innervate the femur reductor muscle in the trochanter. It was previously assumed that the Drosophila trochanter and femur are functionally fused²⁹; however, the fact that the trochanter segment possesses its own dedicated MNs and musculature suggests that either the two segments retain some independent biomechanical function or that the current morphology represents a palimpsest of the ancestral state.

We found that some muscle fibres are innervated by multiple MNs, confirming that adult fly muscles can exhibit polyneuronal innervation, as is the case in locusts³⁰ and larval Drosophila³¹. Mammalian muscle fibres are often poly-innervated in neonatal animals, but they are pruned during development, leading to singly innervated fibres in adults³². We found clear evidence from the XNH data for polyneuronal innervation in the long-tendon muscle (LTM), the proximal fibres of the trochanter flexor (Fig. 4d) and the femur reductor muscle. LTM fibres originate in both the femur (ltm2) and the tibia (ltm1) and insert on the long tendon (retractor unguis), which extends from the femur to the claw at the distal tip of the leg³³ (Fig. 4d, gold line). Polyneuronal innervation may endow the motor system with more flexibility and may help explain how insect limbs achieve precision despite their sparse MN innervation. Another mechanism for enhancing motor flexibility in other arthropods is the presence of GABAergic (γ-aminobutyric acid-producing) MNs that innervate many target muscles³⁴; however, we did not find such ‘common inhibitor’ MNs in the fly leg.

We did not observe any chemical output synapses from leg MNs in the VNC, consistent with past work in the locust³⁵. This lack of output synapses is notable because the vast majority of fly neurons possess apposed input and output synapses, even in dendritic compartments. Larval fly MNs innervating abdominal muscles also lack output synapses^36,37. However, larval MNs do appear to form gap junctions in specific segments, enabling motor neurons to participate in rhythm generation³⁸. In other arthropods such as crustaceans, the central output synapses of MNs are essential for their role in rhythmic pattern generation³⁹.

Identification of wing motor neurons

We next focused our analysis on the wing motor system. Wing muscles (Fig. 5a) can be broadly divided into three groups on the basis of their functional roles in flight^8,40. First, 13 large muscle fibres power flight indirectly by causing the thorax to resonate at high frequencies, which changes the conformation of the mechanically complex wing hinge to drive the wings to flap back and forth⁴¹. The 13 fibres are organized into 4 muscles on each side, a single dorsal longitudinal muscle (DLM) (consisting of 6 fibres (DLM1–6)) and 3 dorsoventral muscles (DVMs) (consisting of 3 fibres (DVM1) and 2 fibres (DVM2 and DVM3)). Second, a set of 12 steering muscles attach directly to the hardened cuticle elements (sclerites) of the wing hinge to rapidly adjust wing motion. Third, the lateral thorax is equipped with four tension muscles, which, although not directly connecting to the sclerites at the base of the wing, are thought to adjust wing motion via their effects on the mechanical stiffness of the thorax⁴⁰.

We identified the MNs that innervate wing muscles by locating neurons in FANC with cell bodies in the VNC and axons that leave through one of the three nerves associated with the wing neuropil (Fig. 5b): the anterior dorsal mesothoracic nerve (ADMN), posterior dorsal mesothoracic nerve (PDMN) and the mesothoracic accessory nerve¹ (MesoAN). We found 37 neurons on each side, 29 of which we determined to be wing MNs (the other 8 are discussed below and in Extended Data Fig. 5a–c). We identified left–right homologous pairs, but here we focus our analysis on the left side.

As with the legs, we identified the muscle targets of all the wing MNs in FANC. Most identifications were based on dendritic morphology from the literature, including MNs that innervate direct and tension muscles^9,10 and MNs that innervate indirect muscles^11,12 (Fig. 5c–e).

The MNs of the DLMs were the only MNs that we found to possess presynaptic sites in the VNC (Extended Data Fig. 5d). These synapses were rare (around ten predicted synapses in total). The presence of weak synaptic connectivity between DLM MNs complements recent work demonstrating that gap junctions between DLM MNs offset their spike timing during flight⁴².

We found three leg MNs and five efferent neurons that innervate these peripheral nerves but that we did not categorize as wing MNs. One is the previously described peripherally synapsing interneuron, which has an important role in escape behaviour⁴³. Three are previously undescribed efferent neurons, one of which has a branch that ascends to the brain (Extended Data Fig. 5a–c).

Coordination of legs and wings during take-off

With knowledge of the muscle targets of the front leg and wing MNs, the connectome points to clear hypotheses about the physiological and behavioural function of premotor circuits. We illustrate the potential of combining the VNC connectome and MN projection map with identification and analysis of preMNs that coordinate the legs and wings during escape take-off (Fig. 6a).

Previous research identified a class of VNC local neurons (giant fibre-coupled (GFC)) that are electrically coupled to the giant fibre⁴⁴, a large descending neuron that drives evasive escape take-off⁴⁵. In the FANC volume, we identified and proofread two subtypes of GFC interneurons (GFC2 and GFC4). We then used FANC synapse predictions to understand how they are synaptically connected to leg and wing MNs (Fig. 6b). The morphology of the GFC2 and GFC4 interneurons indicates that they develop from stem cell hemilineages (18B and 11A, respectively) that release the excitatory neurotransmitter acetylcholine⁴⁶.

We found that the GFC2 interneurons synapse onto indirect and tension wing MNs as well as tergotrochanter MNs that innervate the middle leg (T2), which is consistent with previous literature^44,45. In addition to its role in T2 trochanter–femur extension during jumping, the enormous tergotrochanter muscle is also thought to initiate the first cycle of mechanical oscillation in the power muscles that maintain flight⁴⁷. Thus, GFC2 interneurons are positioned to excite MNs that drive two parallel motor programmes: jumping and initiation of the flight motor. The other interneuron subtype, GFC4, synapses onto front leg (T1) MNs that flex the trochanter–femur or tibia (Fig. 3d–g). This connectivity suggests that descending input from the giant fibre drives synchronized extension of the middle legs (jumping) and flexion of the front legs (lift-off). Although the role of the T1 leg retraction during take-off has not been previously described, it is visible in high-speed videos of Drosophila escape (for example, in supplementary video 1 of ref. 45).

Leg MNs in Drosophila and other animals are typically recruited in a hierarchical sequence, with small, low-force producing MNs firing prior to large, high force producing MNs⁷. In paired electrophysiological recordings from tibia flexor MNs, we previously observed occasional violations of the recruitment order, in which MNs at the top of the hierarchy fire before MNs at the bottom²². We hypothesized that these violations could occur during ballistic escape behaviours, such as escape take-off. Consistent with this prediction, the only MNs that receive synaptic input from GFC4 interneurons are the largest MNs that innervate trochanter–femur and tibia flexor muscles (Fig. 6c). Thus, our analysis of the connectome suggests that descending input from the giant fibre is positioned to subvert the standard leg MN recruitment order to drive rapid, ballistic escape behaviour via the GFC4 interneurons.

Discussion

By revealing the structure of neurons and synapses that underlie animal behaviour, connectomes provide a foundation for generating and falsifying theories of neural circuit function. Although comprehensive connectomes have been mapped for small crawling animals such as Caenorhabditis elegans⁴⁸ and Drosophila larva⁴⁹, we previously lacked a synapse-level wiring diagram of motor circuits for any limbed animal. Here we provide tools for collaboratively proofreading a draft connectome of a female adult Drosophila VNC, an atlas linking the connectome to the body, and analysis of a circuit that coordinates the legs and wings during escape behaviour. The existence of multiple VNC connectomes will enable comparison of neural wiring across individuals and the identification of sexually dimorphic neural circuits^14,15.

The community effort to comprehensively reconstruct neurons and measure synaptic connections in FANC is ongoing. At the time of publication, the community has proofread roughly half of the total FANC dataset, yet the automated reconstruction, proofreading environment and analysis tools already enable interactive inquiry into the neurobiology of the VNC. Our work parallels similar efforts in the adult fly brain¹⁶, the male VNC^14,15,19 and the larval nervous system⁵⁰. The distributed nature of the community-driven FANC effort is a strength, as it enables the connectome to be continuously refined as each circuit is proofread by experts in that area. Even before the connectome is complete, researchers can analyse subcircuits as a means of generating and falsifying hypotheses. Using tools to match VNC neurons reconstructed in the electron microscopy volume with GAL4 lines imaged by light microscopy (Extended Data Fig. 2), it is then possible to test these hypotheses with in vivo recordings and manipulations of neural activity. Details on how to join the FANC community and participate in reconstruction of FANC can be found in Methods. There also exist software tools that facilitate comparison of neuronal morphology and connectivity across individual animals and to bridge electron microscopy datasets¹⁷. Even with a single connectome, however, it is possible to gain insight into morphology and neural circuitry that differ between related species. For example, our reconstruction revealed independent control of the femur and trochanter leg segments, which were previously thought to be mechanically fused in Drosophila²⁹.

Many challenges remain for interpreting a synaptic wiring diagram⁵¹. First, synapses can be either excitatory or inhibitory. Fortunately, VNC neurons develop from stem cell hemilineages that share identifiable morphological features as well as neurotransmitter identity⁴⁶. Therefore, hemilineage identification of VNC cell types can identify candidate neurotransmitters, as we did for the escape circuitry (Fig. 6). As done previously for the whole fly brain electron microscopy dataset, one can also use hemilineage identities as ground truth for training CNNs to predict neurotransmitter identity directly from the electron microscopy image data^16,52.

A second challenge is the inability to see electrical synapses or sites of neuromodulation in current connectome datasets. The spatial resolution of existing electron microscopy datasets is insufficient to resolve gap junctions, which are known to have an important role in sensorimotor processing within the VNC^44,53,54. Sample-preserving imaging approaches such as serial-section transmission electron microscopy have the capacity for higher-resolution imaging at identified locations, potentially enabling visualization of gap junctions. It is also not currently possible to identify sites of neuromodulation, which are essential for flexible motor control⁵⁵. Because electron microscopy reconstruction provides a map of chemical synaptic transmission between neurons, physiological measurements will also be necessary to test for the effects of gap junctions and neuromodulators.

Many organisms use their limbs to maintain posture in the face of gravity, navigate uneven terrain and manipulate objects in the environment. The FANC connectome provides a valuable tool for investigating these fundamental motor control problems and beyond. For example, in a companion paper⁵⁶, we identify similarities and differences in leg and wing premotor networks, which reflect the unique biomechanical specializations of these appendages. The FANC connectome will also enable the identification and analysis of neural circuits for descending modulation⁵⁷, ascending communication with the brain^14,58, and sensory organs distributed across the fly limbs, thorax and abdomen⁴. By creating a bridge between the VNC connectome and the body, the MN projection map will facilitate development and analysis of neuromechanical models for flexible motor control⁵⁹. Matching cell types and comparing connectivity motifs to the larval connectome^37,49 may provide insight into the development and evolution of sensorimotor circuits^60,61. Finally, as volumetric electron microscopy datasets from other animal species are collected, comparative connectomics provides a powerful tool to study the evolution of the nervous system⁶².

Methods

Extended Data

Supplementary Material

Data availability

The electron microscopy dataset and reconstructions are freely available¹. The segmentation of the FANC dataset is available by joining the FANC community. Instructions on joining the FANC community can be found at https://github.com/htem/FANC_auto_recon/wiki#collaborative-community (ref. 80).