Awesome Fly: Curated Fruit Fly Connectome Projects
A curated collection of fruit fly (Drosophila melanogaster) connectome projects, covering MaleCNS, FlyWire, brain simulations, embodied models, games, and research tools by Mert Cobanov.
Curator & Maintainer: Mert Cobanov
Repository: cobanov/awesome-fly
Starter Template: cobanov/fly-connectome-template
License: CC0 1.0 Universal (Linked projects retain their own licenses)
Scope: Connectome-grounded projects across adult Drosophila melanogaster datasets (MaleCNS, FlyWire / FAFB, BANC), spanning biological network simulations, embodied biomechanical agents, game controllers, interactive desktop applications, and connectomic analysis libraries.
Conceptual & Scientific Foundations
A connectome is an empirical map of anatomical neurons and their synaptic connections derived from serial-section electron microscopy (EM). Transforming a static connectomic graph into an active simulation requires defining physiological rules:
Biological Topology $\neq$ Functional Dynamics: Wiring diagrams constrain communication pathways, but simulations require explicit assumptions regarding synaptic weights, neurotransmitter sign (excitatory vs. inhibitory), leak conductances, membrane time constants, sensory transduction, and motor readouts.
Key Connectomic Datasets:
MaleCNS (Janelia / Berg et al. 2026): Complete adult male central nervous system, spanning both the brain and the ventral nerve cord (VNC), providing seamless brain-to-leg motor pathways.
FlyWire / FAFB (Princeton / Seung & Murthy Labs / Dorkenwald et al. 2024): Complete adult female brain connectome (~139,000 neurons and tens of millions of synapses).
BANC (Brain and Nerve Cord): Adult female whole-CNS connectome integrating cephalic and thoracic-abdominal ganglia.
Embodied Simulation (Brain + Body + Physics): Neural activity alone does not explain behavior. Coupling connectome-constrained controllers to physical engines (e.g., MuJoCo) with anatomical biomechanical bodies (Flybody, NeuroMechFly) enables closed-loop sensorimotor experimentation.
Directory of Highlighted Projects & Toolkits
1. Games, Controls & Reinforcement Learning Experiments
Mert Cobanov
Chromium Dino runner controlled by an 80-neuron MaleCNS circuit with leaky tanh dynamics and a 243-parameter readout trained via Cross-Entropy Method (CEM). Play
Circuit subset; open-source benchmark
nftechie
MaleCNS simulation connected to ViZDoom with modeled visual inputs and fixed button readouts; documents negative validation results for survival.
Research experiment
eganeganegan
Benchmarks MaleCNS recurrent controllers against rewired random graphs and standard ANNs on ViZDoom tasks.
Control framework
ornata
Connects MaleCNS simulation to Super Mario 64 with a local macOS dashboard.
Playful controller
blendi-remade
Fabric mod driving fly mobs via filtered MaleCNS graphs with live neural HUDs.
Interactive mod
MarkUnthank
Connects MaleCNS model to CARLA simulator to evaluate steering through simulated limbs.
Bounded steering
hrook1
Browser game featuring a 6,000-neuron MaleCNS escape circuit influencing movement.
Circuit subset
jonatasperaza
Pong paddle controller from MaleCNS subgraph testing dopamine-inspired plasticity.
Negative control
Unjuno
MaleCNS controller interfacing FightingICE with explicit sensory-motor mappings.
Research ledger
tolatolatop
FlyWire chess experiment with Rust/WASM Leaky Integrate-and-Fire (LIF) simulation.
Untrained readout
liuzihe02
Connectome simulation connected to Craftax with PPO descending-neuron training.
Research prototype
furkancak1r
Godot horror game where MaleCNS simulation drives an adaptive event policy.
Playable prototype
2. Desktop Companions, Art & Interactive Worlds
DesktopFly (DenisSergeevitch): macOS/Electron desktop companion combining FlyWire spiking networks with MaleCNS brain-to-leg motor circuits.
gnat (lubabs770): Linux/Hyprland desktop port of DesktopFly featuring a 668-neuron circuit subset and live stimulation GUI.
DesktopFly for Linux (somsom10): Python/GTK3/4 implementation for GNOME on X11/Wayland.
flyverse (tel-0s): Multi-sensory simulation with modeled color vision, olfaction, gustation, and wind mechanosensation.
Infinite Sugar (cnqso): Browser artwork where continuous sweet-sensing input drives neural firing in a terrarium. Live Demo
FLYBOARD (NullLabTests): CPU arcade board for current injection into named MaleCNS populations with live voltage field monitoring.
FLM (nftechie): Couples a frozen LLM to the MaleCNS graph through an adapter interface.
3. Brain Models, Neural Simulators & Embodied Physics
Drosophila brain model (Shiu et al. / Seung & Murthy Labs): Landmark whole-brain Leaky Integrate-and-Fire (LIF) model based on FlyWire.
flybody (TuragaLab / Google DeepMind / HHMI Janelia; Nature 2025): Anatomically accurate fruit fly body in MuJoCo with aerodynamic wings, realistic joint limits, and reinforcement learning environments for walking and flight.
FlyGym / NeuroMechFly (EPFL): Biomechanical simulation framework for embodied sensorimotor control in realistic physical environments.
flyvis (TuragaLab; Nature 2024): Connectome-constrained deep neural network predicting visual system responses and optic flow processing.
AxonWeave: Python framework exposing MaleCNS as a sparse, differentiable substrate for NumPy, PyTorch, and TensorFlow.
FastFly: High-performance CUDA/CuPy GPU simulator running the FlyWire v783 graph at real-time or faster speeds.
webgpu-fly: Browser-based WebGPU/WASM simulator joining FlyWire cephalic circuits, MANC nerve cord data, and Flybody meshes. Demo
4. Connectomic Analysis Toolkits, Registries & Explorers
Codex (Murthy Lab): Frontend and backend powering the FlyWire Connectome Data Explorer.
FlyBrainLab: Interactive 3D computing environment for constructing, analyzing, and executing Drosophila neural circuits.
connectome-interpreter: High-performance path-tracing, effective connectivity, and differentiable graph manipulation library.
natverse (Cambridge Fly Connectomics / Jefferis Lab): Premier R suite for neuroanatomy and connectomics (
malecns,coconatfly,fafbseg,neuprintr,bancr).CAVEclient: Programmatic interface for the Connectome Annotation Versioning Engine.
5. Developer Starter Kit: Fly Connectome Template
Mert Cobanov provides a ready-to-build scaffolding template:
Soma Atlas: Prepared 3D spatial coordinates from MaleCNS.
Body Mesh: Anatomically scaled Flybody 3D asset pipeline.
Interface: React + Three.js interactive workbench for visualizing neural activation and body articulation.
Decoupled Architecture: Cleanly separates neural model execution from rendering, providing a standardized I/O format for custom neural networks or spiking models.
Relevance to Computational Pathology & Spatial Biology
While connectomics operates on neuronal synapses rather than human tissue specimens, the computational methodologies directly cross-pollinate with digital pathology:
Spatial Graphs & Cell Topologies: Connectomic path-finding, motif discovery, and co-clustering (as seen in
cocoaandconnectome-interpreter) share identical mathematical foundations with spatial cell-interaction graphs in cancer immunology (e.g., [[Tumor budding T-cell graphs for pT1 colorectal cancer]]).Multi-Resolution Registration: Aligning EM volumes, confocal stacks, and light microscopy in neuroanatomy leverages the same non-rigid deformation and transform engines used in cross-stain histology registration ([[Cross-Stain Registration]], [[From Samples to Knowledge 2025: QuPath Training Course]]).
Biological Inductive Biases: Benchmarking connectome-constrained sparse architectures against unconstrained artificial neural networks (e.g., in
flyvisandFlyDoom) mirrors computational pathology efforts to incorporate tissue architecture priors rather than treating histology as unconstrained generic natural image patches.
Last updated