henhaobio
Sign in

bioRxivNeurosciencePreprint

Whole-brain annotation and multi-connectome cell typing quantifies circuit stereotypy in Drosophila

Philipp SchlegeliD, Yijie YiniD, Alexander Shakeel BatesiD, Sven DorkenwaldiD, Katharina EichleriD, Paul S BrooksiD, Daniel HaniD, Marina GkantiaiD, Marcia dos SantosiD, Eva J MunnellyiD, Griffin S. BadalamenteiD, Laia Serratosa CapdevilaiD, Varun Aniruddha SaneiD, Markus William PleijzieriD, Imaan FM TamimiiD, Christopher R DunneiD, Irene SalgarellaiD, Alexandre JavieriD, Siqi FangiD, Eric PerlmaniD, Tom KazimiersiD, Sridhar R. JagannathaniD, Arie MatsliahiD, Amy Robinson SterlingiD, Szi-chieh YuiD, , Gregory S.X.E. JefferisiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

The fruit fly Drosophila melanogaster combines surprisingly sophisticated behaviour with a highly tractable nervous system. A large part of the flys success as a model organism in modern neuroscience stems from the concentration of collaboratively generated molecular genetic and digital resources. As presented in our FlyWire companion paper1, this now includes the first full brain connectome of an adult animal. Here we report the systematic and hierarchical annotation of this [~]130,000-neuron connectome including neuronal classes, cell types and developmental units (hemilineages). This enables any researcher to navigate this huge dataset and find systems and neurons of interest, linked to the literature through the Virtual Fly Brain database2. Crucially, this resource includes 4,552 cell types. 3,094 are rigorous consensus validations of cell types previously proposed in the "hemibrain" connectome3. In addition, we propose 1,458 new cell types, arising mostly from the fact that the FlyWire connectome spans the whole brain, whereas the hemibrain derives from a subvolume. Comparison of FlyWire and the hemibrain showed that cell type counts and strong connections were largely stable, but connection weights were surprisingly variable within and across animals. Further analysis defined simple heuristics for connectome interpretation: connections stronger than 10 unitary synapses or providing >1% of the input to a target cell are highly conserved. Some cell types showed increased variability across connectomes: the most common cell type in the mushroom body, required for learning and memory, is almost twice as numerous in FlyWire as the hemibrain. We find evidence for functional homeostasis through adjustments of the absolute amount of excitatory input while maintaining the excitation-inhibition ratio. Finally, and surprisingly, about one third of the cell types proposed in the hemibrain connectome could not yet be reliably identified in the FlyWire connectome. We therefore suggest that cell types should be defined to be robust to inter-individual variation, namely as groups of cells that are quantitatively more similar to cells in a different brain than to any other cell in the same brain. Joint analysis of the FlyWire and hemibrain connectomes demonstrates the viability and utility of this new definition. Our work defines a consensus cell type atlas for the fly brain and provides both an intellectual framework and open source toolchain for brain-scale comparative connectomics.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (15 Jul 2023) · CC BY 4.0

Annotating the 127,978-neuron FlyWire fly connectome yields 4,552 cell types, 3,094 cross-validated against the hemibrain; cell-type edges above 10 synapses or 1.1% of input recur over 90% across hemispheres.

With 127,978 neurons and about 16.5M weighted edges, the FlyWire connectome is hard to use without a systematic parts list, a sense of which edges are reliable, and evidence on whether one brain represents the species. About 90% of the 5,235 hemibrain types were proposed from one hemisphere of a single animal; here they are treated as predictions and tested against two more hemispheres, motivating a cell type definition based on variation between individuals.

Key results

  1. ShownAll neurons were given flow and superclass labels, and most central brain neurons hemilineage and morphology-group labels. The atlas totals 4,552 cell types: 3,094 consensus types validated against the hemibrain plus 1,458 proposed from FlyWire's two hemispheres.

    Over 730k annotations for all 127,978 neurons across 9 superclasses; 57% of central-brain-associated neurons received a terminal type; 120 lineages and 183 hemilineages cover 30,078 (86.4%) central brain neurons; hemilineage sizes differ 3% (+/-4) between hemispheres, and central brain intrinsic neurons differ by 42 (0.1%). · Abstract; Hierarchical annotation of a connectome; A full atlas of neuronal lineages in the central…; Toward multi-connectome cell typing

  2. Shown61% of hemibrain morphology types were re-identified in FlyWire. Matched types are highly stereotyped in cell number and synapse counts, both within and across brains. The other 39% of hemibrain types could not yet be found.

    3,214 of 5,235 types matched, collapsing to 3,094 consensus types (4% merged as composites, 1% split) on 43,600 neurons covering 39.7% of central brain edges; 2,021 unmatched. Mean cells-per-type difference 0.3 (+/-1.7) within vs 0.7 (+/-10) across brains; per-type synapse counts Pearson R 0.99 within, 0.95/0.8 (pre/post) across. · Validating cell types across brains; Cell types are highly stereotyped within and…

  3. ShownThe chance that an edge recurs across hemispheres rises with its weight. Cell-type edges above 10 synapses or 1.1% of the target's input recur over 90% of the time. In a simulation, left-right weight differences of up to about 30% fit technical noise alone.

    Of 434,790 edges, 53% of hemibrain edges recur in FlyWire (57%/59% between FlyWire hemispheres); a 1-synapse hemibrain edge is in both FlyWire hemispheres 14% of the time. Edges >10 synapses are 16% of edges but ~79% of synapses; edges >1.1% of input are ~7% of edges with 54% of synapses. On average 65% of left-right weight variability fell in the modelled noise range. · Interpreting connectomes; How do we know which edges are important?; Technical noise model

  4. SuggestedKenyon cells are about 30% more numerous per hemisphere in FlyWire than in the hemibrain, mainly because KCg-m is almost twice as common. Yet KCg-m keeps the same share of KC synapses: each KCg-m samples fewer odour channels, and its excitation-inhibition ratio is preserved.

    2,597 (FlyWire right), 2,580 (left) and 1,917 (hemibrain) KCs vs a mean cell-count variation of 5% (+/-12%). KCg-m makes and receives 32% and 45% of KC pre- and postsynapses in every hemisphere; mean ALPN types per KCg-m (K) 5.74, 5.88 vs 8.76; a Litwin-Kumar rate model predicts lower optimal K for FlyWire KCg-m. · Variability in the mushroom body; Calculating K from observed connectivity; Kenyon cell model

  5. SuggestedA cell type was defined as neurons that are each closer to a matched group in another brain than to any neuron in their own brain. Clustering the three hemispheres jointly under this definition re-derived or refined most hemibrain fan-shaped body types.

    FC1-3 (357 neurons, connectivity distance): 7 cross-brain types, 5 one-to-one with hemibrain types, 2 merges. FB1-9 (897 neurons, morphology plus connectivity): 114 types vs 146 in the hemibrain, 44% one-to-one, 11% splits, 12% merges, 33% recombinations; 67% recovery vs 61% for directed matching. · Toward multi-connectome cell typing; Defining robust cross-brain cell types

Methods, data and reproducibility

FlyWire neurons were annotated at several levels (superclass, class, hemilineage, morphology group, cell type) by semi-automated assignment and manual curation. Hemibrain neurons were registered into FlyWire space and matched via NBLAST and cross-dataset cosine connectivity clustering, so cell counts, synapse counts and edges could be compared across three hemispheres, with a technical-noise simulation.

  • FlyWire whole-brain connectome of an adult female fly (FAFB electron microscopy volume, release 630) · Drosophila melanogaster · n = 127,978 neurons; ~16.5M weighted edges
  • Janelia hemibrain connectome (most of one central brain hemisphere plus part of the optic lobe), skeletons downloaded from neuPrint · Drosophila · n = 5,235 morphology types; 98k skeletons (~23k typed)
  • Light-level lineage clones and anti-Neuroglian-labelled brains used to identify hemilineage tracts · Drosophila
  • GAL4 and split-GAL4 driver line images used to match descending neurons · Drosophila
Methods & tools
NBLAST, navis, natverse, fafbseg, coconatfly, CAVEclient, neuprint-python, skeletor, CATMAID Spaces, Codex, neuroglancer, VirtualFlyBrain database
Assay conditions & compute
Large NBLAST runs used one cluster node with 112 CPUs and 1TB RAM and took 1 to 2 days wall time; the all-by-all NBLAST matrix for 128k FlyWire neurons takes over 500GB (32-bit floats).
Reproducibility gaps
At the time of writing, Codex and CAVE access required signing up with a Google account; morphology groups were not yet visible in Codex; no software versions are given in the text; no code link is given for the noise model or the Kenyon cell model.
Limitations (6)
  • The optic lobes were newly proofread and may change with further quality control, so the finest cell typing targets the central brain; only 3,943 of an expected 12,800 retinal photoreceptors had been identified. · stated by the authors
  • Across-brain comparisons are confounded by hemibrain truncation, unequal synaptic completion rates and differences in synapse detection (for example far more KC to KC and APL to KC synapses in the hemibrain); restricting to the hemibrain volume does not fully correct for these. · stated by the authors
  • The technical noise model was built for FlyWire's two hemispheres. The Kenyon cell rate model assumes binary weights, uniform APL inhibition and a fixed K, and it is unproven that the circuit maximises dimensionality. · stated by the authors
  • The 1,458 new types rest mainly on FlyWire left-right matching and await further connectomes for validation; edges below the reliability threshold may still contribute to brain function. · stated by the authors
  • Only two animals (three hemispheres) are compared, so the KCg-m count difference and its proposed homeostatic compensation rest on a single pair of brains. · noted in this summary
  • The cross-brain typing definition was tested only on fan-shaped body neurons, with clusters manually reviewed (one adjusted by hand). The estimate that 95% of unmatched hemibrain types need retyping comes from a sampled subset, with the data not shown. · noted in this summary

Read this if you use the FlyWire or hemibrain connectomes and need cell type labels, thresholds for which edges are reliable, or a method for typing neurons jointly across several connectomes.

Citations
102

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISXu '20Matsliah '24Scheffer '20Mehta '22Lin '24Marin '23
76% semantic matchNeuroscience
C. Shan Xu, Michał Januszewski, Zhiyuan Lu+97 more·bioRxiv·Jan 21, 2020· 144 citations