Neuronal wiring diagram of an adult brain

Connectome analysis

We use the term ‘connectome’ to mean a neuronal wiring diagram of an entire nervous system, or at least an entire brain. This is in keeping with the intent of the original definition, which emphasized comprehensiveness. Similarly, the term ‘genome’ refers to the entire DNA sequence of an organism, or at least the entirety of genes. Our neuronal wiring diagram of a whole fly brain arguably crosses the threshold for being called a connectome, although it would be reasonable to insist that a connectome should include the ventral nerve cord as well as the brain. Either way, the comprehensiveness of our wiring diagram has significant benefits for brain research and enables many kinds of studies that were not previously possible using wiring diagrams of portions of the fly brain. The optic lobes and the SEZ are two prominent regions that are mostly absent from the hemibrain. Both sides of the brain are now included, which enables the tracing of pathways that cross the midline. Owing to the presence of afferent and efferent neurons, pathways can be traced from sensory inputs to intrinsic neurons and brain outputs (motor, endocrine and descending neurons). This was done in a global manner to analyse the neuropil projectome, by using the information flow model, and more specifically to uncover the structure and hypothesize a circuit mechanism for behaviours supported by the ocelli. A set of companion studies provides additional global analyses of the connectome and studies of specific families of pathways.

For the first time, one can now compare entire connectomes of different species, starting with D. melanogaster and C. elegans, as touched on here and explored in more depth by Lin et al.. It also enables comparison of connectomes of the same species at different developmental stages. Although FlyWire is currently the only adult fly connectome, it can be compared with the hemibrain reconstruction in regions where they overlap to detect wiring differences between adults of the same species and to validate and extend cell-type definitions.

Finally, the connectome now enables brain simulations—partial connectomes of the early visual system of the fly had already inspired simulations of visual processing. This effort has now been extended to leverage the full connectome and to—for example—predict taste responses of neurons. These simulations assume that that physiological connection strength is proportional to anatomical synapse count, either globally, or for synapses sharing the same presynaptic and postsynaptic cell types, and have inferred connection signs (excitatory versus inhibitory) from neurotransmitter identity as predicted from electron microscopy images or from transcriptomics. Ongoing discoveries regarding the biophysics of fly neurons will guide efforts to make simulations more realistic. For example, inhibition can be shunting rather than subtractive in some fly neurons, and the conductance of an inhibitory synapse can be ten times higher than that of an excitatory synapse. Whereas the simulations mentioned above were based on point neuron models, future simulations could utilize multicompartmental neuron models constructed using the synapse locations and reconstructed neuronal morphologies provided by FlyWire, as well as emerging data about ion channels and receptors from transcriptomics and proteomics.

Electron microscopy data acquisition and reconstruction

The hemibrain was reconstructed from 8 × 8 × 8 nm³ images acquired by focused ion beam scanning electron microscopy (FIB-SEM), a form of block face electron microscopy. By contrast, FlyWire’s reconstruction is based on a full adult fly brain (FAFB) dataset of 4 × 4 × 40 nm³ images acquired by serial section transmission electron microscopy (ssTEM). Initially, the lower z resolution and higher prevalence of artefacts made alignment and reconstruction of ssTEM datasets challenging. These were cited by the hemibrain effort to justify the use of FIB-SEM despite its higher cost, slower speed and complex operation requiring many 20-µm slabs to be imaged individually and then stitched together. Computational advances have now closed this gap and FAFB images were accurately aligned with a new approach that leverages convolutional nets. The hemibrain images were automatically segmented using flood-filling convolutional nets, whereas FlyWire used the older, less computationally expensive approach of boundary-detecting convolutional nets. Overall, from acquisition to reconstruction to analysis to dissemination, the technology stack used by FlyWire is distinct from that used for the hemibrain. A notable overlap is the use of neuroglancer for browser-based 3D visualization.

FlyWire’s whole-brain automated segmentation was proofread with an estimated 33 person-years of effort (Methods), whereas hemibrain proofreading required 50 person-years for a part of the brain. Notably, the accuracy of our proofread wiring diagram is similar to that of the hemibrain (Extended Data Fig. 2 and Supplementary Fig. 3). For both FIB-SEM and ssTEM, incomplete attachment of twigs to backbones is currently the main factor that limits the accuracy of reconstructing synaptic connectivity, and in both cases synaptic connectivity is limited to chemical synapses. Higher resolution might enable the reconstruction of electrical synapses, which are included in the C. elegans connectome.

Limitations of our reconstruction

We showed that the attachment rates of twigs is sufficient to facilitate detection of nearly all large connections (those with more than nine synapses). Nonetheless, the observed synapse counts underrepresent the actual number of synapses and some connections with few synapses remain undetected. Substantial improvements in twig attachment are unobtainable with further proofreading, as increasing the postsynaptic attachment rate from 44.7% to 50% would require further proofreading of more than 700,000 fragments. Therefore, increases in twig attachments will rely on improvements in image acquisition, image alignment and automated reconstruction. Although proofreading was largely carried out in a neuropil-agnostic manner, attachment rates differ between neuropils (Supplementary Fig. 2) owing to differences in the number of synapses on twigs and backbones and how challenging a neuropil was to reconstruct. Although these effects are largely symmetric, the optic lobe was affected by a one-sided artefact. The left lamina was partially severed from the medulla in the left hemisphere (Fig. 1a), reducing the reconstruction accuracy for some of the lamina neurons on one side (Supplementary Fig. 2).

The automated synapse detection currently used by FlyWire was performed by Buhmann et al. in an independent effort. By combining the FlyWire brain reconstruction with these synapses, the resulting connectome inherits the limitations from both. In the case of the synapse detection, users should be aware that the ground truth was limited to a few neuropils. As a result, synapse detection performance is lower for some cell types, and we are aware that sensory neurons are particularly affected. FlyWire’s reconstruction is compatible with any synapse prediction method and improved synapse prediction will be made available in the future.

Imaging larger

Imaging a larger volume would open up other interesting opportunities. Reconstructing and proofreading a full CNS would enable the mapping of all pathways linking the brain and VNC. In the meantime, it is already possible to establish correspondences between FlyWire and FANC, a reconstruction of a VNC from another female fly. The first C. elegans connectome was obtained similarly as a mosaic drawn from multiple worms. Imaging an entire fly, both CNS and body, would enable the addition of sensory organs and muscles to the reconstruction. This also has precedent in the C. elegans connectome, which includes neuromuscular junctions, the Platynereis dumerilii larva, and the 1st instar Drosophila larva for which a whole-animal electron microscopy dataset was recently published.

Technologies developed for FlyWire have already been applied to millimetre-scale chunks of mammalian brain, which are more than 50 times larger in volume than a fly brain. The US National Institutes of Health has begun a transformative project to reconstruct a whole mouse brain from an exabyte of electron microscopy images and a report from the Wellcome trust recently examined the road to a whole mouse brain connectome.

Openness

The 1996 Bermuda Principles mandated daily release of Human Genome Project sequences into the public domain. We believe that openness is also important for large-scale connectomics projects, particularly because these projects are expensive, require coordinated effort and take several years to complete—sharing connectomes only after proofreading and annotation are completed prevents scientific discovery that can occur while the connectome is being completed. Shortly after its inception in 2019, FlyWire has been open to any Drosophila researcher, and set forth clear principles for coordination of scientific effort that prioritized attribution through keeping track of edits to and annotations of the dataset. Hundreds of scientists and proofreaders from more than 50 laboratories joined FlyWire with more than 200 of them contributing more than 100 edits (Supplementary Table 1) and 86 contributing ten or more annotations (Supplementary Table 2). As a result of openness, there are multiple studies that used completed portions of FlyWire’s brain connectome as proofreading proceeded (Supplementary Table 3). Openness has also enabled FlyWire to move faster by incorporating data sources from the community. The electron microscopy data on which FlyWire’s brain connectome is built was shared in 2018 by Bock and colleagues. The synapse data was published by Buhmann et al., neurotransmitter labels were published by Eckstein et al., numerous annotations were contributed by Schlegel et al., neck connective neuron annotations were contributed by Stürner et al., optic lobe annotations were contributed by Matsliah et al. and so far, more than 90,000 cell annotations have been shared by the community. Many cells have received multiple annotations from these sources, and discrepancies will continue to be adjudicated by the community, a process that has improved accuracy in cell-type classification. Overall, we anticipate that similar approaches based on an open ecosystem will enable connectomics to scale more efficiently, economically and equitably.