Drosophila Connectome Experiments

A Brian2 leaky integrate-and-fire (LIF) model of the FlyWire Drosophila connectome — the brain half of a connectome → body pipeline whose other half is NeuroMechFly v2.

The LIF model is a reimplementation of Shiu et al., Nature 634:210–219 (2024), following the reference code at philshiu/Drosophila_brain_model (vendored under data/dbm/).

src/
  lif_model.py            the model: parameters, connectome loading, network construction
  neurons.py              named neuron sets (21 sugar GRNs, MN9, SEZ cell types)
  run_sugar_milestone.py  milestone 1 experiment
data/dbm/                 vendored connectome (FlyWire v630 + v783 parquet) + reference impl
results/                  spike parquet + firing-rate CSV + log, one set per run

The model

Per neuron, integrated only outside the refractory period:

dv/dt = (v_0 - v + g) / t_mbr     : volt
dg/dt = -g / tau                  : volt

Spike when v > v_th; reset v = v_rst; g = 0 mV; then refractory for rfc. A presynaptic spike does g += w after a fixed delay t_dly, with w = (synapse count) × (NT sign) × w_syn — synapse count from the FlyWire connectivity table, sign from the predicted neurotransmitter (ACh → +1, GABA/Glu → −1). External ("optogenetic") drive is a PoissonInput onto v.

symbol	value	meaning	source
v_0, v_rst	−52 mV	resting / reset potential	Kakaria & de Bivort 2017
v_th	−45 mV	spike threshold	Kakaria & de Bivort 2017
t_mbr	20 ms	membrane time constant (C·R)	Kakaria & de Bivort 2017
tau	5 ms	synaptic decay	Jürgensen et al. 2021
t_rfc	2.2 ms	refractory period	Lazar et al. 2021
t_dly	1.8 ms	synaptic transmission delay	Paul et al. 2015
w_syn	0.275 mV	voltage step per synapse	free parameter, fit in Shiu et al. 2024
dt	0.1 ms	integration step (method='exact')	—

Connectome sizes as loaded: v630 — 127,400 neurons, 14,687,178 edges (the materialisation the paper used); v783 — 138,639 neurons, 15,091,983 edges.

Findings

Milestone 1 — sugar activation drives the feeding motor neuron MN9

Reproduces the feeding experiment of Shiu et al. Drive the 21 right-hemisphere sugar-sensing gustatory receptor neurons (Gr64f class) with Poisson input and check that MN9 — the proboscis motor neuron for pharyngeal pumping — fires, against an otherwise identical unstimulated control.

python3 src/run_sugar_milestone.py --dataset 630 --n-run 30 --rate 150 --n-proc 5

Result (v630, 150 Hz, 30 trials × 1 s):

	firing rate
MN9, sugar-stimulated	83.47 ± 4.60 Hz
MN9, unstimulated control	0.00 Hz
neurons active, sugar condition	431 (21 stimulated + 410 downstream)
neurons active, control	0
total spikes recorded	409,883

Interpretation.

• The circuit works end to end. Stimulating a sensory population 410 synaptic partners away from MN9 reliably drives the motor neuron. The sugar → feeding pathway is present and functional in the connectome-constrained LIF network, matching the paper's optogenetically validated result.
• Every spike is causal. The network is completely silent in the control condition (0 active neurons, 0 Hz everywhere). There is no background activity, no self-sustaining loop — so all 409,883 spikes in the stimulated condition are directly attributable to the sugar drive. This is a clean stimulus–response assay with zero baseline to subtract.
• The stimulated neurons track their drive. The 21 sugar GRNs fire at 140–154 Hz under 150 Hz Poisson input (they are given zero refractory period so the imposed rate is not clipped); the highest-rate downstream neurons sit just below them (~110–140 Hz), consistent with strong monosynaptic relay before divergence dilutes the signal.
• Cost. ~112 s wall time (75 s stimulated + 37 s control) on 5 CPU cores, ~3 GB per worker, no GPU. A whole-brain spiking model is tractable on a laptop.

Outputs (tag = 630_150Hz_30trials):

• results/spikes_sugar_<tag>.parquet — one row per spike (t, trial, flywire_id, exp_name)
• results/rates_sugar_<tag>.csv — mean ± sd firing rate per neuron per condition, with cell-type names
• results/milestone1.log — full run log

What has not been tested yet

• v783. Milestone 1 runs only on v630, because one of the 21 published sugar root ids does not resolve in the v783 materialisation (FlyWire root ids change between snapshots). A v783 run needs the sugar set re-identified by annotation rather than by hard-coded id.
• Inhibition / silencing. lif_model.silence() (zeroing a neuron's outgoing synapses) is implemented but no experiment exercises it.
• Any body coupling. Nothing in this repo touches NeuroMechFly yet — see TODO.
• Rate sensitivity. Only 150 Hz drive has been run; the paper sweeps rate.

Requirements

brian2 (tested 2.9.0), pandas, pyarrow, numpy, joblib. A working C compiler is needed for Brian2's Cython code generation; without it the model still runs, far more slowly. Each parallel worker holds a full copy of the network (~3 GB), so size --n-proc to RAM.

---

References

Simulator

• Stimberg M, Brette R, Goodman DFM (2019). Brian 2, an intuitive and efficient neural simulator. eLife 8:e47314. doi:10.7554/eLife.47314
• Brian 2 documentation — https://brian2.readthedocs.io/
• Nowotny T, Goodman DFM, Stimberg M (2020). Brian2GeNN: accelerating spiking neural network simulations with graphics hardware. Sci Rep 10:410. — GPU backend option.
• Alevi D, et al. (2022). Brian2CUDA: Flexible and Efficient Simulation of Spiking Neural Network Models on GPUs. Front Neuroinform 16:883700. — GPU backend option.

Connectome

• Dorkenwald S, et al. (2024). Neuronal wiring diagram of an adult brain. Nature 634(8032):124–138. doi:10.1038/s41586-024-07558-y
• Schlegel P, et al. (2024). Whole-brain annotation and multi-connectome cell typing of Drosophila. Nature 634(8032):139–152. doi:10.1038/s41586-024-07686-5 — cell-type / class annotations used to pick stimulation and readout populations.
• Eckstein N, et al. (2024). Neurotransmitter classification from electron microscopy images at synaptic sites in Drosophila melanogaster. Cell 187(10):2574–2594. doi:10.1016/j.cell.2024.03.016 — per-neuron neurotransmitter → synapse sign.
• Codex: FlyWire — https://codex.flywire.ai/ · bulk downloads https://codex.flywire.ai/api/download

LIF model reproduced here

• Shiu PK, et al. (2024). A Drosophila computational brain model reveals sensorimotor processing. Nature 634(8032):210–219. doi:10.1038/s41586-024-07763-9 — the model this project implements.
• Shiu PK, et al. (2023). Preprint: A leaky integrate-and-fire computational model based on the connectome of the entire adult Drosophila brain… bioRxiv 2023.05.02.539144. doi:10.1101/2023.05.02.539144 — explicit equations and parameter table.
• philshiu/Drosophila_brain_model — https://github.com/philshiu/Drosophila_brain_model — reference Brian2 implementation + bundled connectivity, vendored under data/dbm/.

LIF parameter provenance

• Kakaria KS, de Bivort BL (2017). Ring Attractor Dynamics Emerge from a Spiking Model of the Entire Protocerebral Bridge. Front Behav Neurosci 11:8. doi:10.3389/fnbeh.2017.00008 — v_0, v_rst, v_th, t_mbr.
• Jürgensen A-M, et al. (2021). A neuromorphic model of olfactory processing and sparse coding in the Drosophila larva brain. Neuromorphic Comput. Eng. 1:024007. doi:10.1088/2634-4386/ac3ba6 — tau.
• Lazar AA, et al. (2021). Accelerating with FlyBrainLab the discovery of the functional logic of the Drosophila brain… eLife 10:e62362. doi:10.7554/eLife.62362 — t_rfc.
• Paul MM, et al. (2015). Bruchpilot and Synaptotagmin collaborate to drive rapid glutamate release and active zone differentiation. Front Cell Neurosci 9:29. doi:10.3389/fncel.2015.00029 — t_dly.

Body (target for coupling)

• Wang-Chen S, et al. (2024). NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila. Nat Methods 21(12):2353–2362. doi:10.1038/s41592-024-02497-y
• Lobato-Rios V, et al. (2022). NeuroMechFly, a neuromechanical model of adult Drosophila melanogaster. Nat Methods 19:620–627. doi:10.1038/s41592-022-01466-7 — morphology: 65 segments, 122 DoF.
• flygym — https://github.com/NeLy-EPFL/flygym/ · docs https://neuromechfly.org/ · pip install "flygym[examples]"

Related connectome-simulation work

• Lappalainen JK, et al. (2024). Connectome-constrained networks predict neural activity across the fly visual system. Nature 634:1132–1140. doi:10.1038/s41586-024-07939-3 · flyvis https://github.com/TuragaLab/flyvis — candidate visual front-end.
• Neuromorphic Simulation of Drosophila Melanogaster Brain Connectome on Loihi 2. arXiv:2508.16792.
• JNLiew/flylif_orientation_maps — https://github.com/JNLiew/flylif_orientation_maps — FlyWire + LIF, oriented stimuli → optic-lobe orientation maps.

---

TODO

Model / experiments

• Port milestone 1 to v783: select the sugar GRNs and MN9 by FlyWire annotation (cell type / class from Schlegel et al.) instead of hard-coded root ids, so the experiment is materialisation-independent.
• Rate sweep: run the sugar → MN9 assay across drive rates (e.g. 50–300 Hz) and plot the MN9 input–output curve; compare against Shiu et al.
• Silencing controls: use lif_model.silence() to knock out candidate relay interneurons between the sugar GRNs and MN9 and confirm MN9 drive drops — identifies the necessary pathway, not just a sufficient one.
• Second behavioural circuit as an independent check (e.g. antennal grooming: drive bristle mechanosensory neurons → neck/leg descending neurons).
• Cross-check total edge counts and NT-sign assignment against Codex directly, not only the vendored parquet.
• Switch to set_device('cpp_standalone') or a GPU backend (Brian2GeNN / Brian2CUDA) for the longer sweeps.

NeuroMechFly coupling (the point of the project)

• Install flygym; stand up a minimal closed loop that steps Brian2 and NeuroMechFly in lock-step (run the network in dt_phys-sized chunks inside the gym loop).
• Sensory encoding (body → brain): map NeuroMechFly ommatidia intensities and antennal odor concentration to Poisson rates on the corresponding FlyWire sensory neurons (R1–R8 / L1–L5, ORNs). Consider flyvis as a drop-in visual front-end.
• Motor decoding (brain → body): low-pass the spike trains of a chosen descending neuron population into joint targets / CPG parameters, and feed them as the action to nmf.step().
• Handle the brain-only limitation: FlyWire has no VNC, so leg motor neurons are absent. Read out at descending neurons and drive legs with a hand-built CPG (NeuroMechFly v1 style), or stitch in the MANC / male VNC connectome at the descending/ascending interface.
• First embodied milestone: odor-gated turning — odor on one antenna biases descending activity, body turns toward the source.

Repo

• git init, pin exact dependency versions (requirements.txt / environment.yml).
• Move the large parquet files out of the repo (or Git LFS) and add a fetch script from Codex.
• Restore the citations/ set (SOURCES.md, PARAMETERS.md, THEORY.md, references.bib) if provenance detail beyond this README is needed.