Neuronal wiring diagram of an adult brain

Specimen, alignment and segmentation

Multiple brains of 7-day-old (iso) w1118 × (iso) Canton-S G1 adult female flies were screened by Zheng et al. and one was picked for electron microscopy imaging. Zheng et al. published the original electron microscopy stack (FAFB14) which we previously realigned (FAFB14.1) using a neural network trained to predict pairwise displacement fields. We produced transformations between the FAFB14 and FAFB14.1 which are accessible via natverse, navis and flybrains. We automatically segmented all cells in the dataset using a neuronal boundary-detecting neural network and mean affinity agglomeration.

Neuropils

Meshes for individual neuropils were based on work by Ito et al.. More specifically, we took meshes previously generated from a full brain segmentation of the JFRC2 template brain which are also used by the Virtual Fly Brain project (see also https://natverse.org/nat.flybrains/reference/JFRC2NP.surf.html). These meshes were moved from JFRC2 into FlyWire (FAFB14.1) space through a series of non-rigid transforms. In addition, we also generated two neuropil meshes for the laminae and for the ocellar ganglion. For these, the FlyWire synapse cloud was voxelized with 2 µm isotropic resolution, meshed using the marching cube algorithm using Python and manually post-processed in Blender 3d.

We calculated a volume for each neuropil using its mesh. In the aggregated volumes presented in the paper we assigned the lamina, medulla, accessory medulla, lobula and lobula plate to the optic lobe. The remaining neuropils but the ocellar ganglion were assigned to the central brain.

Neuropil synapse assignments

We assigned synapses to neuropils based on their presynaptic location. We used ncollpyde (https://pypi.org/project/ncollpyde/) to calculate whether the location was within a neuropil mesh and assigned the synapse accordingly. Some synapses remained unassigned after this step because the neuropils only resemble rough outlines of the underlying data. We then assigned all remaining synapses to the closest neuropil if the synapse was within 10 µm from it. The remaining synapses were left unassigned.

Correction of left–right inversion

Our reconstruction used the FAFB electron microscopy dataset. A number of consortium members (A. Bates, P. Kandimalla and S. Noselli) alerted us that the FAFB imagery seemed to be left–right inverted based on the cell types innervating the asymmetric body. Eventually a left–right inversion during FAFB imaging was confirmed. All side annotations in figures, in Codex and elsewhere are based on the true biological side. For technical reasons, we were unable to invert the underlying FAFB image data and therefore continue to show images and reconstructions in the same orientation as in Zheng et al., although we now know that in such frontal views the fly’s left is on the viewer’s left. For full details of this issue including approaches to display FAFB and other brain data with the correct chirality, please see the companion paper.

Proofreading system

FlyWire uses CAVE for hosting the proofreadable segmentation and all of its annotations. CAVE’s proofreading system is the PyChunkedGraph which has been described in detail elsewhere.

Proofreading annotations

Any user in FlyWire was able to mark a cell as complete, indicating that a cell was good for analysis. However, such annotations did not prevent future proofreading of a cell as commonly smaller branches were added later on. We created an annotation table for these completion markings. Each completion marking was defined by a point in space and the cell segment that overlapped with this point at any given time during proofreading was associated with the annotation. We created a webservice allowing users to submit completion markings for any cell. For convenience, we added an interface to this surface directly into Neuroglancer such that users can submit completion information for cells right after proofreading (Supplementary Fig. 10). When users submitted completion annotations we also recorded the current state of the cell. We encouraged users to submit new completion markings for a cell that they edited to indicate that edits were intentional. Recording the status of a cell at submission enabled us to calculate volumetric changes to a cell through further proofreading and flag cells for review if they received substantial changes without new completion markings.

Onboarding proofreaders

Proofreaders came from several distinct labour pools: community members, citizen scientists from Eyewire (Flyers), and professional proofreading teams at Princeton and Cambridge. Proofreaders at Princeton consisted of staff at Princeton University and at SixEleven. Similarly, proofreading at Cambridge was performed by staff at Cambridge University and Ariadne. All proofreaders completed the built-in interactive tutorial and were directed to self-guided proofreading training. For practice and learning purposes, the Sandbox, a complete replica of the FlyWire data, allowed new users to freely make edits and explore without affecting the actual ‘Production’ dataset. When ready, an onboarding coordinator tested the new proofreader before giving access to the Production dataset. Later onboarding called for users to send demonstration Sandbox edits that were reviewed by the onboarding coordinator. A new class of view-only users was introduced in early 2023, allowing researchers early data access for analysis purposes. All early access users attended a live onboarding session in Zoom prior to being granted edit or view access.

Training the professional proofreading team

The professional proofreading team received additional proofreading training. Correct proofreading relies on a diverse array of 2D and 3D visual cues. Proofreaders learned about 3D morphology, resulting from false merger or false split without knowing what types of cells they are. Proofreaders studied various types of ultrastructures as the ultrastructures provide valuable 2D cues and serve as reliable guides for accurate tracing. Before professional proofreaders were admitted into Production, each of them practiced on average >200 cells in a testing dataset where additional feedback was given. In this dataset, we determined the accuracy of test cells by comparing them to ground-truth reconstructions. To improve proofreading quality, peer learning was highly encouraged.

Recruitment of citizen scientists

The top 100 players from Eyewire, a gamified electron microscopy reconstruction platform that crowdsources reconstructions in mouse retina and zebrafish hindbrain, received an invitation to beta test proofreading in FlyWire. A new set of user onboarding and training materials were created for citizen scientists, including: a blog, forum and public Google docs. We created bite-sized introduction videos, a comprehensive FlyWire 101 resource, as well as an Optic Lobe Cell Guide to aid users in understanding the unique morphology of flies. A virtual Citizen Science Symposium introduced players to the project, after which the self-dubbed ‘Flyers’ began creating their own resources, such as a new comprehensive visual guide to cell types, conducting literature reviews, and even developing helpful FlyWire plugins. As of publication, FlyWire has 12 add-on apps ranging from a batch processor to cell naming helper (https://blog.flywire.ai/2022/08/11/flywire-addons/).

Proofreading strategy to complete the connectome

As previously described, proofreading of the connectome was focused on the microtubule-rich ‘backbones’ of neurons. Microtubule-free ‘twigs’ were only added if discovered incidentally or sought out specifically by members of the community. After proofreading, users marked neuronal segments as ‘complete’ indicating that neurons were ready for analysis but further changes remained possible. While Drosophila neuroscientist members of the FlyWire community generally contributed proofreading for their neurons of interest, the bulk of the segments was proofread by professional proofreaders in the following way. First, we proofread all segments with an automatically detected nucleus in the central brain by extending it as much as possible and removing all false mergers (pieces of other neurons or glia attached). Second, we proofread the remaining segments in descending order of their synapse count (pre+post) up to a predefined size threshold of 100 synapses. Third, we proofread remaining segments if they had at least one connection containing at least 15 synapses.