FlyWire Olfactory SNN (MaskedRecurrentLIFSNN)
A connectome-constrained recurrent spiking neural network for odor identity classification in Drosophila melanogaster, trained on the DoOR olfactory receptor response dataset.
Model description
The recurrent connectivity of this SNN is fixed to the FlyWire connectome subgraph (antennal lobe projection neurons + mushroom body Kenyon cells). Synaptic signs (excitatory/inhibitory) come from predicted neurotransmitter types in FlyWire. Only the weight magnitudes are learned; the topology is biological.
Architecture
Input: odor receptor vector (DoOR: ~52 receptors)
→ Linear(input_dim → hidden_dim, no bias)
→ 20 LIF timesteps with:
• Poisson spike encoding from rate-coded input
• Recurrent current: spk × (W_rec ⊙ mask ⊙ sign)ᵀ
• Norse LIFCell (surrogate gradient, α=100)
→ time-averaged spike rates
→ Linear(hidden_dim → num_classes)
- Neuron model: Leaky Integrate-and-Fire (Norse
LIFCell,method="super") - Recurrent mask: Binary from FlyWire adjacency (fixed, not learned)
- Synaptic signs: ACh/DA/5-HT/OA → +1 (excitatory); GABA/Glu → −1 (inhibitory)
- Training: Adam optimizer, CrossEntropyLoss, surrogate gradients through LIF
Files
| File | Description |
|---|---|
model.safetensors | Trained weights (best validation checkpoint) |
config.json | Architecture hyperparameters |
connectome_mask.npz | FlyWire olfactory subgraph (binary adjacency + signs) |
connectome_meta.json | Connectome metadata (neuron count, edge count, source) |
modeling_snn.py | Standalone MaskedRecurrentLIFSNN class |
Usage
import scipy.sparse as sp
import torch
from safetensors.torch import load_file
# Load the model
from modeling_snn import MaskedRecurrentLIFSNN
adjacency = sp.load_npz("connectome_mask.npz")
model = MaskedRecurrentLIFSNN(
input_dim=52, # from config.json
hidden_dim=800, # from config.json
num_classes=500, # from config.json
adjacency=adjacency,
steps=20,
alpha=100.0,
)
state_dict = load_file("model.safetensors")
model.load_state_dict(state_dict)
model.eval()
# Inference
x = torch.randn(1, 52) # receptor activation vector
logits, spike_sparsity = model(x)
predicted_odor = logits.argmax(dim=1).item()
Training details
- Dataset: DoOR (Database of Odorant Responses) — CC BY-SA 4.0
- Cross-validation: 5-fold over odor identities × 5 seeds = 25 runs
- Early stopping: patience 5 on validation accuracy
- Optimizer: Adam (lr=1e-3, weight_decay=1e-5)
- Batch size: 32
- Max epochs: 80
- SNN timesteps: 20
- Evaluation: 5× Monte Carlo averaging over stochastic Poisson encoding
Biological basis
The model's recurrent topology is extracted from the FlyWire whole-brain connectome of Drosophila melanogaster (FAFB dataset). The olfactory subgraph includes:
- Antennal Lobe Projection Neurons (ALPN): relay processed odor information
- Kenyon Cells (KC): mushroom body neurons for associative olfactory memory
This captures the AL → PN → KC pathway that the fly uses for odor discrimination and learning.
Citation
If you use this model, please cite:
- The DoOR database: Münch & Galizia (2016). DoOR 2.0 — Comprehensive mapping of Drosophila melanogaster odorant responses. Scientific Reports, 6, 21841. https://doi.org/10.1038/srep21841
- The FlyWire connectome: Dorkenwald et al. (2024). Neuronal wiring diagram of an adult brain. Nature, 634, 124–138. https://doi.org/10.1038/s41586-024-07558-y