NeuroTrailblazers banner

FlyWire Whole-Brain Connectome

Overview

The FlyWire project delivered the first complete synaptic-resolution connectome of an adult animal brain — that of the fruit fly Drosophila melanogaster. Published by Dorkenwald et al. in Nature (2024), this landmark dataset comprises approximately 139,255 neurons connected by roughly 54.5 million chemical synapses, organized into an estimated 8,453 cell types. The achievement demonstrates that whole-brain connectomics is feasible for small brains and establishes a reference framework for understanding how an entire nervous system is wired.

The significance of FlyWire extends beyond the dataset itself. The project pioneered a model of large-scale, community-driven proofreading that may define how future connectomics projects operate. By combining state-of-the-art automated segmentation with the collective effort of 287 contributors worldwide, FlyWire showed that neither machines nor humans alone can reconstruct a brain — but together, they can.

The Starting Point: FAFB

FlyWire did not begin from scratch. The project built upon the Full Adult Fly Brain (FAFB) electron microscopy volume, a serial-section transmission electron microscopy (ssTEM) dataset acquired by Zheng et al. (2018) and published in Cell. The FAFB volume captured an entire adult female Drosophila brain at synaptic resolution, producing roughly 21 million images at 4 nm x 4 nm in-plane resolution with 40 nm section thickness.

The FAFB volume was a monumental imaging achievement, but raw images alone are not a connectome. To go from images to a wiring diagram required two additional steps: automated segmentation (assigning each voxel to a specific neuron) and proofreading (correcting the inevitable errors in automated segmentation). FlyWire tackled both.

Technical Pipeline

Automated Segmentation with Flood-Filling Networks

The initial segmentation of the FAFB volume used flood-filling networks (FFNs), a deep learning architecture developed at Google Research. Unlike conventional segmentation approaches that classify each voxel independently, FFNs iteratively “grow” segments by predicting, at each step, which neighboring voxels belong to the same neuron. This approach naturally handles the complex, branching morphology of neurons.

The FFN segmentation produced an over-segmented representation of the brain — meaning that individual neurons were often split into multiple fragments. This is by design: over-segmentation is preferable to under-segmentation (merging two neurons) because splits are easier to correct than merges during proofreading.

The FlyWire Platform: CAVE and Neuroglancer

To enable collaborative proofreading at scale, the FlyWire team built a web-based platform on top of two key technologies:

Together, CAVE and Neuroglancer transformed connectome proofreading from a specialized, single-user desktop application task into a massively parallel, web-based collaborative effort.

Proofreading at Scale: The Social Engineering Challenge

Recruitment and Training

Perhaps the most innovative aspect of FlyWire was not its technology but its community model. The project recruited 287 proofreaders from around the world, including professional neuroscientists, postdocs, graduate students, and trained citizen scientists. Recruitment occurred through lab networks, conference presentations, social media, and word of mouth.

New proofreaders underwent a structured training protocol:

  1. Tutorial missions: Guided tasks that introduced the basic proofreading operations (merge and split) on pre-selected neurons with known ground truth.
  2. Supervised proofreading: New contributors worked on neurons that were subsequently reviewed by experienced proofreaders, with feedback provided.
  3. Independent proofreading: After demonstrating proficiency, contributors were given access to proofread neurons independently.

Gamification and Motivation

FlyWire incorporated gamification elements to sustain motivation over the multi-year proofreading campaign:

Quality Control and Inter-Annotator Agreement

Ensuring consistency across 287 proofreaders required robust quality control:

Key Scientific Findings

Brain-Wide Cell-Type Atlas

One of the primary outputs of FlyWire is a comprehensive cell-type atlas of the adult Drosophila brain. Schlegel et al. (2024), published concurrently in Nature, used the FlyWire connectome to classify neurons into approximately 8,453 cell types based on morphology, connectivity, and spatial position. This atlas provides the most complete catalog of cell types in any adult brain to date.

The cell-type classification revealed several important patterns:

Circuit Architecture of the Central Complex

The central complex is a midline neuropil structure involved in navigation, spatial orientation, and locomotor control. FlyWire provided the first complete wiring diagram of this structure, revealing:

Sensorimotor Pathways and Descending Neurons

The complete brain connectome enabled systematic tracing of pathways from sensory input to motor output. The catalog of descending neurons — neurons that project from the brain to the ventral nerve cord (the fly equivalent of the spinal cord) — was mapped in its entirety for the first time. This revealed:

Convergence and Divergence Patterns

Analysis of the complete connectome revealed systematic patterns of information flow:

Data Availability and Tools

The FlyWire connectome is publicly available through several access points:

Example Access Pattern (Python)

Researchers typically access FlyWire data through the caveclient Python package, which provides authenticated access to the CAVE backend. Queries can retrieve individual neuron morphologies, synapse lists between specified neuron pairs, or bulk connectivity matrices for entire brain regions.

Lessons for the Field

Whole-Brain Connectomics Is Achievable

FlyWire proved that it is possible to reconstruct the complete synaptic wiring diagram of an adult brain. While Drosophila is a small brain (~100,000 neurons), the principles and tools developed for FlyWire — automated segmentation, collaborative proofreading, versioned annotation infrastructure — are directly applicable to larger brains.

Crowd-Sourced Proofreading Works

The 287-person proofreading community demonstrated that connectome reconstruction does not require a small team of highly specialized annotators. With proper training, tooling, and quality control, a large distributed community can collectively achieve the accuracy required for scientific analysis.

Automation Alone Is Not Enough

Despite using state-of-the-art deep learning for segmentation, extensive human proofreading was still required. The automated segmentation contained numerous split and merge errors that would have corrupted downstream analyses if left uncorrected. This underscores the continued importance of human-in-the-loop approaches in connectomics.

The Connectome Is a Beginning, Not an End

The publication of the FlyWire connectome is not the conclusion of a project but the opening of a new era. The wiring diagram is a static snapshot that must be interpreted through functional experiments, computational modeling, and comparative analysis. The real scientific payoff will come from the community of researchers who use this resource over the coming decades.

Discussion Questions for Instructors

  1. Why was over-segmentation preferred to under-segmentation in the initial automated pipeline? What are the tradeoffs?
  2. How would you design a quality control system for a proofreading effort with 500+ contributors? What metrics would you track?
  3. The central complex circuit architecture matches computational models of ring attractor networks. Does this validate the models, or could the match be coincidental?
  4. What are the limitations of a connectome derived from a single individual brain? How might brain-to-brain variability affect the generality of findings?
  5. Compare the FlyWire community model with traditional academic lab structures. What are the advantages and disadvantages of each for large-scale data projects?

Key References