Emulation of the Drosophila Fly Brain

Whole-brain leaky integrate-and-fire model of the adult fruit fly, built from the FlyWire connectome (~138k neurons, ~5M synapses). Activate and silence arbitrary neurons; observe downstream spike propagation.

Based on the paper A leaky integrate-and-fire computational model based on the connectome of the entire adult Drosophila brain reveals insights into sensorimotor processing (Shiu et al.).

Usage

With this computational model, one can manipulate the neural activity of a set of Drosophila neurons. The output of the model is the spike times and rates of all affected neurons.

Two types of manipulations are currently implemented:

• Activation: Neurons can be activated at a fixed frequency to model optogenetic activation. This triggers Poisson spiking in the target neurons. Two sets of neurons with distinct frequencies can be defined.
• Silencing: In addition to activation, a different set of neurons can be silenced to model optogenetic silencing. This sets all synaptic connections to and from those neurons to zero.

The entrypoint is main.py, which parses CLI arguments and calls code/benchmark.py -- the central orchestrator that dispatches to framework-specific runners: run_brian2_cuda.py, run_pytorch.py, and run_nestgpu.py.

# Run all 4 frameworks with default durations (0.1s–1000s) and trials (1,4,8,16,32)
python main.py

# Specific durations and trial count
python main.py --t_run 0.1 1 10 --n_run 1

# Single framework
python main.py --nestgpu --t_run 1 --n_run 1

# Combine frameworks
python main.py --brian2-cpu --pytorch --t_run 0.1 1 --n_run 1 4 8 16 32

Results are incrementally saved to data/benchmark-results.csv as each benchmark completes, with separate columns for setup time (loading, compilation) and simulation time (the always-on cost).

Ground truth comparison

Brian2 (CPU) serves as the ground truth for neural accuracy: it implements the canonical LIF model from Shiu et al. (Nature 2024), which achieved 91% prediction accuracy against experimental Drosophila data. Each backend also saves per-neuron spike trains to data/results/, and a comparison script measures how closely the other backends reproduce Brian2's output:

python code/compare_ground_truth.py                  # default: t_run=1s, n_run=1
python code/compare_ground_truth.py --t_run 10 --n_run 4   # longer / averaged

This computes active-neuron overlap (Jaccard), per-neuron firing-rate correlation, and spike-count ratios, and writes structured results to data/ground-truth-comparison.json.

Installation

Conda environment

The brain-fly conda environment provides everything needed to run the Brian2, Brian2CUDA, and PyTorch backends (including CUDA-enabled PyTorch):

conda env create -f environment.yml
conda activate brain-fly

NEST GPU

NEST GPU requires a separate build from source with a custom neuron model (user_m1). This is only needed if you want to use the --nestgpu backend.

Prerequisites:

• NVIDIA CUDA Toolkit (12.x) — follow the official installation guide.
• CMake — sudo apt install cmake (or see cmake.org).

Steps:

1. Clone NEST GPU:

git clone https://github.com/nest/nest-gpu

2. Copy the custom source files into the NEST GPU tree. You must replace /path/to/nest-gpu with your own local path:

cp scripts/nestgpu_source_files/src/user_m1.{h,cu}    /path/to/nest-gpu/src/
cp scripts/nestgpu_source_files/pythonlib/nestgpu.py   /path/to/nest-gpu/pythonlib/

The patched nestgpu.py fixes weight array initialization (lines 2225-2227).

3. Build and install (set -DCMAKE_CUDA_ARCHITECTURES to match your GPU, e.g. 89 for RTX 4070):

cmake -DCMAKE_CUDA_ARCHITECTURES=89 \
      -DCMAKE_INSTALL_PREFIX=$HOME/.nest-gpu-build \
      /path/to/nest-gpu
make -j$(nproc) && make install

For a full setup from a fresh Windows machine (WSL2 + CUDA + Miniconda), see scripts/setup_WSL_CUDA.sh.

---

Frameworks

Framework	Backend	Status
Brian2	C++ standalone (multi-core CPU)	ready
Brian2CUDA	CUDA standalone (GPU)	ready
PyTorch	CUDA (GPU)	ready
NEST GPU	CUDA (GPU, custom user_m1 neuron)	ready

All four frameworks share the same data, model parameters, and folder structure. A single conda environment (brain-fly) plus a system-level NEST GPU install runs everything.

Quickstart

# Create the conda environment (includes CUDA-enabled PyTorch)
conda env create -f environment.yml
conda activate brain-fly

# Run a 1-second benchmark on all backends
python main.py --t_run 1 --n_run 1 --no_log_file

# Specific backends (combinable)
python main.py --brian2-cpu                    # Brian2 CPU only
python main.py --brian2cuda-gpu               # Brian2CUDA GPU only
python main.py --pytorch                      # PyTorch only
python main.py --nestgpu                      # NEST GPU only
python main.py --pytorch --nestgpu            # PyTorch + NEST GPU

# Full benchmark suite (all durations, n_run=1,4,8,16,32, all backends)
python main.py

main.py options

Flag	Description
(default)	Run all: Brian2 (CPU) → Brian2CUDA (GPU) → PyTorch → NEST GPU
--brian2-cpu	Brian2 C++ standalone (CPU) only
--brian2cuda-gpu	Brian2CUDA (GPU) only
--pytorch	PyTorch (GPU/CPU) only
--nestgpu	NEST GPU only
--t_run	Simulation duration(s) in seconds, e.g. --t_run 0.1 1 10
--n_run	Number of independent trials, e.g. --n_run 1 4 8 16 32
--log_file FILE	Write log to file (default: data/results/benchmarks.log)
--no_log_file	Console output only

Backend flags are combinable: --brian2-cpu --pytorch runs Brian2 CPU then PyTorch.

Project structure

fly-brain/
├── main.py                     # Entrypoint (benchmark runner CLI)
├── environment.yml             # Conda env definition (brain-fly)
├── code/
│   ├── benchmark.py            # Orchestrator: config, logging, dispatcher
│   ├── run_brian2_cuda.py      # Brian2 / Brian2CUDA benchmark runner
│   ├── run_pytorch.py          # PyTorch benchmark runner (model + utils)
│   ├── run_nestgpu.py          # NEST GPU benchmark runner (subprocess per trial)
│   ├── compare_ground_truth.py # Compare backends against Brian2 (CPU) ground truth
│   └── paper-brian2/           # Original paper code (not used by benchmarks)
│       ├── model.py            # Core LIF network model (Brian2)
│       ├── utils.py            # Analysis helpers (load_exps, get_rate)
│       ├── example.ipynb       # Tutorial: activation, silencing, rate analysis
│       └── figures.ipynb       # Reproduce paper figures (uses archive 630 data)
├── data/
│   ├── 2025_Completeness_783.csv       # Neuron list (FlyWire v783)
│   ├── 2025_Connectivity_783.parquet   # Synapse connectivity (FlyWire v783)
│   ├── benchmark-results.csv           # Accumulated benchmark timings
│   ├── ground-truth-comparison.json   # Backend accuracy vs Brian2 (CPU)
│   ├── sez_neurons.pickle              # SEZ neuron subset (for figures)
│   ├── weight_coo.pkl                  # Cached sparse weights COO (gitignored)
│   ├── weight_csr.pkl                  # Cached sparse weights CSR (gitignored)
│   ├── archive/
│   │   ├── 2023_Completeness_630.csv   # Legacy v630 data
│   │   └── 2023_Connectivity_630.parquet
└── scripts/
    └── setup_WSL_CUDA.sh       # WSL2 + CUDA + Miniconda setup

Data

The model uses FlyWire connectome data version 783 (public release). Legacy version 630 data is kept in data/archive/ for paper figure reproduction.

File	Description	Size
2025_Completeness_783.csv	Neuron IDs and metadata	3.2 MB
2025_Connectivity_783.parquet	Pre/post-synaptic indices + weights	97 MB
weight_coo.pkl	Sparse weight matrix (COO), auto-generated by PyTorch	~288 MB
weight_csr.pkl	Sparse weight matrix (CSR), auto-generated by PyTorch	~289 MB

Architecture per framework

	Brian2 / Brian2CUDA	PyTorch	NEST GPU
Build step	C++ / CUDA codegen + compile	None (eager mode)	None
Trial parallelism	Sequential (device.run)	Batched (batch_size=n_run)	Subprocess per trial (cannot reset in-process)
Weight format	Brian2 Synapses object	Sparse CSR tensor	Array-based Connect
Neuron model	Brian2 equations	Custom nn.Module classes	Custom CUDA kernel (user_m1)
Timestep	0.1 ms	0.1 ms	0.1 ms

System requirements

• Linux (tested on Ubuntu 22.04 under WSL2 on Windows 11)
• NVIDIA GPU with CUDA 12.x (tested on RTX 4070)
• Miniconda / Anaconda
• NEST GPU compiled from source (for --nestgpu backend)
• scripts/setup_WSL_CUDA.sh documents the full setup from a fresh Windows machine