Neuronal wiring diagram of an adult brain

Images of an entire adult female fly brain (Fig. 1e,f) were previously acquired by serial section transmission electron microscopy and released into the public domain by Zheng et al.. We previously realigned the electron microscopy images, automatically segmented all neurons in the images, created a computational system that allows interactive proofreading of the segmentation, and assembled an online community (FlyWire). During the initial phase, much of the proofreading was done by a distributed community of Drosophila research groups in the FlyWire Consortium, and focused on neurons of interest to these groups. During the later phase, the remaining neurons were mainly proofread by centralized teams at Princeton and Cambridge, with contributions from citizen scientists worldwide. The recruitment and training of proofreaders and their workflows are described in the Methods.

Chemical synapses were automatically detected in the images as pairs of presynapse–postsynapse locations. The whole brain contains 0.0175 mm³ of neuropil volume and around 130 million synapses. This equates to 7.4 synapses per µm³, a much higher density than that of mammalian cortex (less than 1 synapse per µm³). The central brain and left and right optic lobes (including the lamina) contain 0.0103, 0.0036 and 0.0036 mm³ of neuropil volume, respectively, with synapse counts in approximately the same proportion. Synapses were combined with proofread neurons using the Connectome Annotation Versioning Engine (CAVE) to yield the connectome.

We next assessed completeness and accuracy of proofreading. We had already shown that FlyWire proofreading can yield accurate results through comparison with light microscopic reconstructions of neurons that are known to be highly stereotyped across individual flies. A second method is to subject reconstructed neurons to an additional round of proofreading, which was previously shown to yield few changes. Because proofreading workflows and personnel have changed over time, and accuracy can vary across brain regions, we repeated this evaluation by subjecting 826 neurons from the central brain to a further round of proofreading. Relative to this additional round, our proofread dataset achieved an average F1 score of 99.2% by volume (Extended Data Fig. 2a,b).

By quantifying how many of the automatically detected synapses are attached to proofread segments, as opposed to being isolated in tiny ‘orphan’ segments, we can estimate completeness of the proofreading. We found high attachment rates of presynapses (approximately 122 million presynapses (93.7%) attached), whereas attachment rates of postsynapses were lower (approximately 58.1 million postsynapses (44.7%) attached) owing to less proofreading and reattachment of twigs, which contain most of the postsynapses (Extended Data Fig. 2c,d). Attachment rates were generally in agreement between the two hemispheres of FlyWire and with the hemibrain (Extended Data Fig. 2e–g) and varied by neuropil (Supplementary Fig. 2). As with the hemibrain, false negative synapses are the dominant type of error but false positives also exist. For this reason, analyses using the connectome should consider thresholding to remove spurious connections. Thresholds should be adjusted to the individual analyses. For the analyses presented below (and connections indicated at https://codex.flywire.ai), we use a threshold of five synapses to determine a connection between two neurons. The accompanying paper by Matsliah et al. found a threshold of two synapses appropriate for analysing connections in the optic lobes. Assuming that such errors are statistically independent, accuracy is expected to be high for detection of connections involving multiple synapses.

We estimate that FlyWire’s brain reconstruction took around 33 person-years of manual proofreading. The reconstruction remains open for proofreading and annotations, and new versions of the resource will be released in future (the analysis presented here is from version 783). This enables correction of remaining errors as they are discovered and further rounds of validation to be performed.